feat(quota): seal per-account quota overrides behind a revision-guarded owner action (#1245)

Store the override record with a state field kept separate from the admin
wire type. Every set and clear is revision-guarded and refuses a sealed
record with Blocked, and every read of a sealed record, including
admission, returns Blocked. The read-failure fallback to configured limits
is unchanged.

Add Service.SealAccountOverrides, accepted only when IAM already holds the
account as TERMINATING. It creates the override bucket if needed after that
check, creates or updates the record as sealed, and succeeds only on a
leader-accepted write, returning its revision. awsgw serves it on
admission.v1.SealAccountQuotaOverride. Nothing calls it yet.

Replace the unwired delete-based CleanupAccount.

.github/workflows/actions.lock

@@ -73,7 +73,7 @@ workflows:
     '.github/workflows/ui.yml':
         - 'actions/checkout@v7.0.1'
         - 'actions/setup-node@v7.1.0'
-        - 'pnpm/action-setup@v6.0.10'
+        - 'pnpm/action-setup@v6.1.0'
 dependencies:
     'actions/cache@v6.1.0':
         ref: 'v6.1.0'
@@ -125,9 +125,9 @@ dependencies:
         commit: 'sha1-dfe3c3f87815947d99a8997f908cb6525fc44e9e'
         owner_id: 761456
         repo_id: 258271226
-    'pnpm/action-setup@v6.0.10':
-        ref: 'v6.0.10'
-        commit: 'sha1-0977fd99725f1db4007ccb2928dbb4e90d06cc86'
+    'pnpm/action-setup@v6.1.0':
+        ref: 'v6.1.0'
+        commit: 'sha1-ea17c68df8912ef543352723c149a84f56e3d413'
         owner_id: 21320719
         repo_id: 262289747
     'softprops/action-gh-release@v3.0.3':

.github/workflows/ui.yml

@@ -35,7 +35,7 @@ jobs:
           sparse-checkout: spinifex/runtime/roles/spinifexui/frontend
 
       - name: Setup pnpm
-        uses: pnpm/action-setup@v6.0.10
+        uses: pnpm/action-setup@v6.1.0
         with:
           package_json_file: spinifex/runtime/roles/spinifexui/frontend/package.json
 

Makefile

@@ -379,9 +379,9 @@ install-system:
 		ovn-central ovn-host openvswitch-switch dhcpcd-base
 
 install-go:
-	@echo -e "\n....Installing Go 1.27.2 for $(ARCH) ($(GO_ARCH))...."
+	@echo -e "\n....Installing Go 1.27.3 for $(ARCH) ($(GO_ARCH))...."
 	@if [ ! -d "/usr/local/go" ]; then \
-		curl -L https://go.dev/dl/go1.27.2.linux-$(GO_ARCH).tar.gz | tar -C /usr/local -xz; \
+		curl -L https://go.dev/dl/go1.27.3.linux-$(GO_ARCH).tar.gz | tar -C /usr/local -xz; \
 	else \
 		echo "Go already installed in /usr/local/go"; \
 	fi
@@ -412,7 +412,7 @@ fix:
 
 govulncheck:
 	@echo "Running govulncheck..."
-	$(_Q)scripts/run-gate.sh govulncheck scripts/govulncheck-gate.sh
+	$(_Q)scripts/run-gate.sh govulncheck go tool govulncheck ./...
 	@echo "  govulncheck ok"
 
 # NilAway — advisory nil-panic analysis. Not in preflight due to false positives

build/Dockerfile.distro

@@ -1,6 +1,6 @@
 ARG VERSION=dev
 
-FROM golang:1.27.2-bookworm@sha256:5cf287a799e6b94384bad13d16b14904c531f51ba65792237e122ce42b392f61 AS builder
+FROM golang:1.27.3-bookworm@sha256:5c3e9b43cfd3bb95c7937462e29d87ada57b481a96fa23ad0eb61778d7a00110 AS builder
 
 ARG VERSION
 

contracts/admission/v1/README.md

@@ -0,0 +1,42 @@
+# Admission contract, version 1
+
+This package owns the cross-process contract for the admission domain's
+quota override. It is not an in-process service interface or the owner of
+quota state.
+
+## Seal account quota override
+
+`SealAccountQuotaOverrideSubject` (`admission.v1.SealAccountQuotaOverride`) is
+a request/reply route on which a gateway, in the `awsgw-quota-seal` queue
+group, irreversibly seals one account's quota override record.
+`domains/admission/quota` owns it: `awsgw` serves it.
+`accountteardown` is the consumer: it calls this subject to fence every
+override writer at the quota owner before the account record is removed.
+
+The request names the account in its body (`account_id`) and may carry an
+opaque `correlation_ref` that is echoed for logs.
+`awsgw` ignores `X-Account-ID` for this subject: the account sealed is the one
+named in the body.
+A seal is refused with `IncorrectState` unless IAM already holds the account
+as `TERMINATING`; a precondition read failure, or no termination checker
+wired, answers `InternalError`.
+Both refusals write nothing and never open the override bucket.
+
+A reply is sent only on success: `disposition` is always `sealed` and
+`revision` is the stream leader's accepted write for that seal.
+A failure is the standard awserrors error payload, not this reply shape.
+Sealing is idempotent: a seal on an already-sealed record re-writes the same
+value and returns a fresh, higher revision.
+Any current gateway may answer; the leader-accepted revision is the evidence
+whichever one does, and the override bucket is opened at seal time even on a
+gateway whose quotas were disabled at startup.
+
+### Compatibility boundary
+
+The subject name carries the contract major; `account_quota_seal_test.go`
+preserves the version-1 JSON shape.
+Producer and consumer are in the coordinated Spinifex release, so the window
+is N and N-1.
+A gateway that predates the subject has no subscription, so the request gets
+no responder.
+An incompatible change to the subject or fields needs a new major.

contracts/admission/v1/account_quota_seal.go

@@ -0,0 +1,31 @@
+// Package admissionv1 holds the admission domain's cross-process contract:
+// the subjects, request/reply shapes and wire vocabulary the quota owner and
+// its consumers share.
+package admissionv1
+
+// SealAccountQuotaOverrideSubject is the request/reply subject on which a
+// gateway seals one account's quota override record. Sealing is irreversible
+// and refused unless IAM already holds the account as TERMINATING.
+const SealAccountQuotaOverrideSubject = "admission.v1.SealAccountQuotaOverride"
+
+// AccountQuotaOverrideDispositionSealed is the only disposition a successful
+// reply carries: a reply is sent only on success, so there is no other value.
+const AccountQuotaOverrideDispositionSealed = "sealed"
+
+// SealAccountQuotaOverrideRequest asks for one account's quota override to be
+// sealed. CorrelationRef is an opaque caller reference echoed for logs; it
+// grants nothing and is never a fence or authority to write.
+type SealAccountQuotaOverrideRequest struct {
+	AccountID      string `json:"account_id"`
+	CorrelationRef string `json:"correlation_ref,omitempty"`
+}
+
+// SealAccountQuotaOverrideReply confirms accountID's quota override record
+// was sealed at Revision, the stream leader's accepted write for that seal.
+// A failure is the standard awserrors error payload, not this type.
+type SealAccountQuotaOverrideReply struct {
+	AccountID      string `json:"account_id"`
+	CorrelationRef string `json:"correlation_ref,omitempty"`
+	Disposition    string `json:"disposition"`
+	Revision       uint64 `json:"revision"`
+}

contracts/admission/v1/account_quota_seal_test.go

@@ -0,0 +1,43 @@
+package admissionv1
+
+import (
+	"encoding/json"
+	"testing"
+
+	"github.com/stretchr/testify/require"
+)
+
+func TestSealAccountQuotaOverrideSubject(t *testing.T) {
+	require.Equal(t, "admission.v1.SealAccountQuotaOverride", SealAccountQuotaOverrideSubject)
+}
+
+func TestSealAccountQuotaOverrideRequestJSONContract(t *testing.T) {
+	data, err := json.Marshal(SealAccountQuotaOverrideRequest{AccountID: "111122223333", CorrelationRef: "req-1"})
+	require.NoError(t, err)
+	require.JSONEq(t, `{"account_id":"111122223333","correlation_ref":"req-1"}`, string(data))
+
+	data, err = json.Marshal(SealAccountQuotaOverrideRequest{AccountID: "111122223333"})
+	require.NoError(t, err)
+	require.JSONEq(t, `{"account_id":"111122223333"}`, string(data))
+}
+
+func TestSealAccountQuotaOverrideReplyJSONContract(t *testing.T) {
+	reply := SealAccountQuotaOverrideReply{
+		AccountID:      "111122223333",
+		CorrelationRef: "req-1",
+		Disposition:    AccountQuotaOverrideDispositionSealed,
+		Revision:       42,
+	}
+	data, err := json.Marshal(reply)
+	require.NoError(t, err)
+	require.JSONEq(t, `{
+		"account_id": "111122223333",
+		"correlation_ref": "req-1",
+		"disposition": "sealed",
+		"revision": 42
+	}`, string(data))
+
+	data, err = json.Marshal(SealAccountQuotaOverrideReply{AccountID: "111122223333", Disposition: AccountQuotaOverrideDispositionSealed})
+	require.NoError(t, err)
+	require.JSONEq(t, `{"account_id":"111122223333","disposition":"sealed","revision":0}`, string(data))
+}

contracts/ec2/v1/README.md

@@ -40,7 +40,7 @@ The request (`account_id`) and reply (`vpc_id`, `error`) bodies are still declar
 ## Platform inventory
 
 `PlatformInventorySubject` (`ec2.v1.PlatformInventory`) is a request/reply route on which a daemon in the `spinifex-workers` queue group reports one account's key pairs, placement groups, launch templates and their `ec2-tags` entries from the EC2 owner stores.
-EC2 owns it: `domains/ec2/accountinventory` produces the reply and the daemon serves it; the only consumer today is `accountteardown`, which binds an adapter but does not call it.
+EC2 owns it: `domains/ec2/accountinventory` produces the reply and the daemon serves it; the only consumer today is `accountteardown`, which calls it only on an explicit operator dry run (`spx admin account delete --dry-run` or the admin API `DryRun`) and only logs a comparison with its legacy listings; no teardown calls it.
 
 The request names the account in its body (`account_id`, a 12-digit ID) and may carry an opaque `correlation_ref` that is echoed for logs.
 The daemon ignores `X-Account-ID` for this subject.
@@ -60,3 +60,42 @@ The subject name carries the contract major; `platform_inventory_test.go` preser
 Producer and consumer are in the coordinated Spinifex release, so the window is N and N-1.
 A daemon that predates the subject has no subscription, so the request gets no responder; the consumer treats no responder, a timeout, an error reply, an undecodable reply, a reply for another account, and a missing, repeated or unknown-status kind as `scan-incomplete`, never clean.
 An incompatible change to the subject, fields or vocabulary needs a new major.
+
+## Seal account settings
+
+`SealAccountSettingsSubject` (`ec2.v1.SealAccountSettings`) is a request/reply route on which a daemon in the `spinifex-workers` queue group irreversibly seals one account's EC2 settings record.
+EC2 owns it: `domains/ec2/account` serves it through the daemon.
+`accountteardown`'s StageAccount seal reaper is the consumer: it calls this subject only after the seal-capability probe below reports every configured node compatible, and completes only once it holds a successful sealed reply.
+
+The request names the account in its body (`account_id`) and may carry an opaque `correlation_ref` that is echoed for logs.
+The daemon ignores `X-Account-ID` for this subject: the account sealed is the one named in the body.
+A seal is refused with `IncorrectState` unless IAM already holds the account as `TERMINATING`; a precondition read failure answers `InternalError`.
+Both refusals write nothing.
+
+A reply is sent only on success: `disposition` is always `sealed` and `revision` is the stream leader's accepted write for that seal.
+A failure is the standard awserrors error payload, not this reply shape.
+Sealing is idempotent: a seal on an already-sealed record re-writes the same value and returns a fresh, higher revision.
+
+### Compatibility boundary
+
+The subject name carries the contract major; `account_settings_seal_test.go` preserves the version-1 JSON shape.
+Producer and consumer are in the coordinated Spinifex release, so the window is N and N-1.
+A daemon that predates the subject has no subscription, so the request gets no responder.
+An incompatible change to the subject or fields needs a new major.
+
+## Seal-account-settings capability
+
+`AccountSettingsSealCapabilitySubject(node)` (`ec2.v1.AccountSettingsSealCapability.<node>`) is a per-node, read-only request/reply route on which `accountteardown` asks one configured daemon whether it supports the account-settings seal, immediately before every seal attempt.
+EC2 owns it: the daemon serves it with no queue group, so a request addressed to one node reaches only subscribers configured as that node, and two daemons sharing a node ID both answer.
+`accountteardown` is the consumer: its StageAccount seal reaper probes every configured node and sends no seal request unless every one answers compatible.
+
+`AccountSettingsSealCapabilityVersion` is a claim about the whole binary's writers, not about the responder: version 1 means every EC2 account-settings writer in the binary is revision-guarded and refuses a sealed record, and the binary serves the seal subject.
+A daemon binary without this responder — predating the capability, or carrying only the seal subject without it — has no subscriber, so the probe for that node reads as incomplete (`no-reply`), the same as a stopped or partitioned node.
+The reply proves nothing beyond the moment it was sent: a daemon downgraded or replaced after a successful probe is outside the guarantee.
+
+### Compatibility boundary
+
+The subject name carries the contract major; `account_settings_seal_capability_test.go` preserves the version-1 JSON shape and the subject builder.
+Producer and consumer are in the coordinated Spinifex release, so the window is N and N-1.
+A daemon that predates the capability has no subscription, so the probe gets no responder for that node.
+An incompatible change to the subject or fields needs a new major.

contracts/ec2/v1/account_settings_seal.go

@@ -0,0 +1,28 @@
+package ec2v1
+
+// SealAccountSettingsSubject is the request/reply subject on which a daemon
+// seals one account's EC2 settings record. Sealing is irreversible and refused
+// unless IAM already holds the account as TERMINATING.
+const SealAccountSettingsSubject = "ec2.v1.SealAccountSettings"
+
+// AccountSettingsDispositionSealed is the only disposition a successful reply
+// carries: a reply is sent only on success, so there is no other value.
+const AccountSettingsDispositionSealed = "sealed"
+
+// SealAccountSettingsRequest asks for one account's EC2 settings to be
+// sealed. CorrelationRef is an opaque caller reference echoed for logs; it
+// grants nothing and is never a fence or authority to write.
+type SealAccountSettingsRequest struct {
+	AccountID      string `json:"account_id"`
+	CorrelationRef string `json:"correlation_ref,omitempty"`
+}
+
+// SealAccountSettingsReply confirms accountID's settings record was sealed at
+// Revision, the stream leader's accepted write for that seal. A failure is
+// the standard awserrors error payload, not this type.
+type SealAccountSettingsReply struct {
+	AccountID      string `json:"account_id"`
+	CorrelationRef string `json:"correlation_ref,omitempty"`
+	Disposition    string `json:"disposition"`
+	Revision       uint64 `json:"revision"`
+}

contracts/ec2/v1/account_settings_seal_capability.go

@@ -0,0 +1,39 @@
+package ec2v1
+
+import "strings"
+
+// AccountSettingsSealCapabilityVersion is the version this binary reports.
+// Version 1: every EC2 account-settings writer in the binary is
+// revision-guarded and refuses a sealed record, and the seal subject is served.
+const AccountSettingsSealCapabilityVersion = 1
+
+// AccountSettingsSealCapabilitySubject is one daemon's read-only
+// seal-capability route, addressed per node so a probe reaches exactly the
+// node it asks about.
+func AccountSettingsSealCapabilitySubject(node string) string {
+	return "ec2.v1.AccountSettingsSealCapability." + node
+}
+
+// ValidSealCapabilityNode reports whether node is exactly one NATS subject
+// token. Responder and probe both check it before building a subject, so
+// neither can address a wildcard or another node's route.
+func ValidSealCapabilityNode(node string) bool {
+	return node != "" && !strings.ContainsAny(node, ".*>\t\n\r ")
+}
+
+// AccountSettingsSealCapabilityRequest asks node whether it supports the
+// account-settings seal. CorrelationRef is an opaque caller reference echoed
+// for logs; it grants nothing and is never a fence or authority to write.
+type AccountSettingsSealCapabilityRequest struct {
+	Node           string `json:"node"`
+	CorrelationRef string `json:"correlation_ref,omitempty"`
+}
+
+// AccountSettingsSealCapabilityReply reports the responding node's capability
+// version. It carries no account data and proves nothing beyond the moment it
+// was sent.
+type AccountSettingsSealCapabilityReply struct {
+	Node              string `json:"node"`
+	CapabilityVersion int    `json:"capability_version"`
+	CorrelationRef    string `json:"correlation_ref,omitempty"`
+}

contracts/ec2/v1/account_settings_seal_capability_test.go

@@ -0,0 +1,43 @@
+package ec2v1
+
+import (
+	"encoding/json"
+	"testing"
+
+	"github.com/stretchr/testify/require"
+)
+
+func TestAccountSettingsSealCapabilitySubject(t *testing.T) {
+	require.Equal(t, "ec2.v1.AccountSettingsSealCapability.node-1", AccountSettingsSealCapabilitySubject("node-1"))
+	require.Equal(t, "ec2.v1.AccountSettingsSealCapability.node-2", AccountSettingsSealCapabilitySubject("node-2"))
+}
+
+func TestAccountSettingsSealCapabilityRequestJSONContract(t *testing.T) {
+	data, err := json.Marshal(AccountSettingsSealCapabilityRequest{Node: "node-1", CorrelationRef: "req-1"})
+	require.NoError(t, err)
+	require.JSONEq(t, `{"node":"node-1","correlation_ref":"req-1"}`, string(data))
+
+	data, err = json.Marshal(AccountSettingsSealCapabilityRequest{Node: "node-1"})
+	require.NoError(t, err)
+	require.JSONEq(t, `{"node":"node-1"}`, string(data))
+}
+
+func TestAccountSettingsSealCapabilityReplyJSONContract(t *testing.T) {
+	reply := AccountSettingsSealCapabilityReply{Node: "node-1", CapabilityVersion: 1, CorrelationRef: "req-1"}
+	data, err := json.Marshal(reply)
+	require.NoError(t, err)
+	require.JSONEq(t, `{"node":"node-1","capability_version":1,"correlation_ref":"req-1"}`, string(data))
+
+	data, err = json.Marshal(AccountSettingsSealCapabilityReply{Node: "node-1", CapabilityVersion: 1})
+	require.NoError(t, err)
+	require.JSONEq(t, `{"node":"node-1","capability_version":1}`, string(data))
+}
+
+func TestValidSealCapabilityNode(t *testing.T) {
+	for _, node := range []string{"node-1", "node_2", "n3"} {
+		require.True(t, ValidSealCapabilityNode(node), node)
+	}
+	for _, node := range []string{"", "a.b", "*", ">", "node-*", "node >", "node\t1", "node\n1"} {
+		require.False(t, ValidSealCapabilityNode(node), "%q", node)
+	}
+}

contracts/ec2/v1/account_settings_seal_test.go

@@ -0,0 +1,43 @@
+package ec2v1
+
+import (
+	"encoding/json"
+	"testing"
+
+	"github.com/stretchr/testify/require"
+)
+
+func TestSealAccountSettingsSubject(t *testing.T) {
+	require.Equal(t, "ec2.v1.SealAccountSettings", SealAccountSettingsSubject)
+}
+
+func TestSealAccountSettingsRequestJSONContract(t *testing.T) {
+	data, err := json.Marshal(SealAccountSettingsRequest{AccountID: "111122223333", CorrelationRef: "req-1"})
+	require.NoError(t, err)
+	require.JSONEq(t, `{"account_id":"111122223333","correlation_ref":"req-1"}`, string(data))
+
+	data, err = json.Marshal(SealAccountSettingsRequest{AccountID: "111122223333"})
+	require.NoError(t, err)
+	require.JSONEq(t, `{"account_id":"111122223333"}`, string(data))
+}
+
+func TestSealAccountSettingsReplyJSONContract(t *testing.T) {
+	reply := SealAccountSettingsReply{
+		AccountID:      "111122223333",
+		CorrelationRef: "req-1",
+		Disposition:    AccountSettingsDispositionSealed,
+		Revision:       42,
+	}
+	data, err := json.Marshal(reply)
+	require.NoError(t, err)
+	require.JSONEq(t, `{
+		"account_id": "111122223333",
+		"correlation_ref": "req-1",
+		"disposition": "sealed",
+		"revision": 42
+	}`, string(data))
+
+	data, err = json.Marshal(SealAccountSettingsReply{AccountID: "111122223333", Disposition: AccountSettingsDispositionSealed})
+	require.NoError(t, err)
+	require.JSONEq(t, `{"account_id":"111122223333","disposition":"sealed","revision":0}`, string(data))
+}

docs/COMMANDS.md

@@ -109,7 +109,7 @@ Operational commands for inspecting cluster state. These fan out NATS requests t
 | Command | Flags | Description |
 |---------|-------|-------------|
 | `spx admin kv digest` | `--store-dir` (default `<spinifex-dir>/nats/jetstream`), `--work-dir` (holds the temporary copy, default the system temp dir), `--stream` (repeatable filter), `--seqs` (also print a hash per sequence), `--json` (print one JSON document, the input `spx admin kv compare` reads) | Read-only per-replica consistency check. Copies this node's `$G` streams tree (without consumer state) to a temporary directory, opens the copy in a private in-process NATS server with no listener, and prints one line per stream with first/last sequence, message count and a SHA-256 over every stored message (sequence, subject, timestamp, headers, data). Run it on every node with `sudo` and diff the output with the `#` lines removed: replicas of one stream must match. A direct get over NATS is not used because any replica may answer it. Writes landing during the copy can change a replica's tail, so gate API writes or re-run before calling a tail difference divergence. Prefer `--json --seqs` on every node followed by `spx admin kv compare`, which tells a diverged replica from one that is merely behind. |
-| `spx admin kv compare <digest.json>...` | — (`-` reads any number of digests from stdin) | Reads the `spx admin kv digest --json --seqs` output of every node and checks that each stream's replicas hold the same content. A replica that is merely behind is not divergence; these are: the same sequence holding different content on two replicas, a message missing from one replica that nothing on it explains (no later write to the same subject, no expiry under `max_age` or a message TTL, no stream limit dropping its head), a stream held by fewer or more nodes than its replica count, replicas disagreeing on that count, or no sequence range common to every replica. Sequences past the slowest replica's last are not compared. Exits 0 when every stream is consistent, 1 on divergence, 2 on bad input. Digests are taken from live stores seconds apart, so re-run before trusting a failure that a write landing between two copies could explain. `install-node.sh` gates cluster formation on it. |
+| `spx admin kv compare <digest.json>...` | — (`-` reads any number of digests from stdin) | Reads the `spx admin kv digest --json --seqs` output of every node and checks that each stream's replicas hold the same content. A replica that is merely behind is not divergence; these are: the same sequence holding different content on two replicas, a message missing from one replica that nothing on it explains (no later write to the same subject, no expiry under `max_age` or a message TTL, no stream limit dropping its head), a stream held by fewer or more nodes than its replica count, replicas disagreeing on that count, or no sequence range common to every replica. A stream with a `max_age` whose replicas were copied at different times and share no sequence range is reported as `skipped` (volatile) rather than diverged, since a short-lived bucket such as a leader lease turns over between copies; any sequence two of its replicas share must still match, and each message a replica lacks must be superseded, expired, or written after that replica was copied. Sequences past the slowest replica's last are not compared. Exits 0 when every stream is consistent, 1 on divergence, 2 on bad input. Digests are taken from live stores seconds apart, so re-run before trusting a failure that a write landing between two copies could explain. `install-node.sh` gates cluster formation on it. |
 | `spx admin cluster shutdown` | `--force` (shutdown even if nodes don't respond), `--timeout` (max wait per phase, default 120s), `--dry-run` (print phase plan without executing) | Performs coordinated, phased shutdown of entire cluster. Phases execute in order: GATE (stop API/UI) → DRAIN (stop VMs) → STORAGE (stop viperblock) → PERSIST (stop predastore) → INFRA (stop NATS/daemon). Each phase waits for all nodes to ACK before proceeding. Uses JetStream state tracking. |
 | `spx admin cluster drain-dhcp` | `--timeout` (reply-collection window, default 30s) | Asks each vpcd to DHCPRELEASE every external-pool DHCP lease it currently holds, returning them to the upstream DHCP server. Run on teardown before stopping services — an env reset otherwise strands held leases upstream until their TTL expires, eventually exhausting the upstream scope. Best-effort: warns and exits 0 if the cluster is already down. |
 | `spx admin node drain --local` | `--local` (drain the local node only, required), `--timeout` (max wait per phase, default 120s), `--unless-restarting` (drain on a real host shutdown or a plain target stop, and skip only when a restart of the stack is already queued; unset, it always drains) | Runs the GATE and DRAIN phases against the local node only: powers down its guests via QMP and unmounts their volumes (flushing the viperblock WAL) while every service is still up. STORAGE/PERSIST/INFRA are left to systemd's ordered teardown. Wired as the `ExecStop` of `spinifex-shutdown.service`, so `systemctl stop spinifex.target` and host reboot/poweroff drain guests before any storage service stops. |

docs/PACKAGE_BOUNDARY_MIGRATION.md

@@ -446,7 +446,7 @@ These residuals were recorded by individual slices and must be resolved once the
 - `providers/ebs/viperblock` → `runtime/compute/nbd` and `runtime/host/process` (dependency, moved unchanged from `services/viperblockd`): the adapter formats NBD endpoints and manages nbdkit processes through runtime packages. ADR-0001 does not list providers → runtime as forbidden and the layering test does not enforce it (same category as `providers/ebs/conformance` → `runtime/compute/nbd`), but a provider should not depend on node runtime. Clears when NBD endpoint naming has a provider-facing home (a candidate for `contracts/ebs/v1`) and process supervision is injected by the role. The adapter also imports `providers/objectstore`, a provider-to-provider edge that is allowed.
 - `gateway/ec2.go` (registration): the EC2 dispatch map stays a `GatewayConfig` method in `gateway` and imports every `domains/ec2/awsapi/*` package; `gateway/ec2_idempotency.go` also imports `domains/ec2/awsapi/idem` (`KVBucket`, `TTL`, `TokenAndParams`) for the dispatch-time idempotency wrapper. ADR-0001 makes the domain own its action registration; moving the map waits on the generic ingress registration seam and must not be done as a package move.
 - `gateway/requestaudit.go`: the outermost request-audit middleware records AWS fields and is updated from four places in the gateway's authentication flow through the gateway's own context key. Moving it would export about seven names used only by that flow, so it stays in `gateway` until the generic ingress registration seam defines how ingress middleware receives authentication results.
-- `accountteardown` (dependency): its EC2 reapers call `domains/ec2/awsapi/*` action adapters directly, as they did under `gateway/ec2`. The ADR's `accountteardown` current-to-target map row replaces this with domain-owned cleanup capabilities before the package moves. It also holds a read-only consumer of the EC2 platform-inventory contract (`contracts/ec2/v1`, `ec2.v1.PlatformInventory`), served by `domains/ec2/accountinventory`; no engine path calls it yet, and the key-pair, placement-group and launch-template reapers still use the `awsapi` adapters.
+- `accountteardown` (dependency): its EC2 reapers call `domains/ec2/awsapi/*` action adapters directly, as they did under `gateway/ec2`. The ADR's `accountteardown` current-to-target map row replaces this with domain-owned cleanup capabilities before the package moves. It also holds a read-only consumer of the EC2 platform-inventory contract (`contracts/ec2/v1`, `ec2.v1.PlatformInventory`), served by `domains/ec2/accountinventory`; only the explicit dry run (`Engine.DryRun`) calls it, to log a shadow comparison, and the key-pair, placement-group and launch-template reapers still use the `awsapi` adapters. A domain-owned cleanup capability for the proven-safe subset of those three kinds (key pairs and empty placement groups no instance, spot request or live launch-template version names, and launch-template residue; never live launch templates) exists as `domains/ec2/accountcleanup` (a Go interface with injected owners and reference readers, no NATS subject); nothing calls it yet. It also holds the StageAccount settings-seal reaper, a consumer of `contracts/ec2/v1`'s `ec2.v1.SealAccountSettings` and per-node `ec2.v1.AccountSettingsSealCapability.<node>` subjects, served by `domains/ec2/account` through the daemon; the reaper reads no KV and imports no domain package, reaching both subjects only through NATS and the configured node set `NewClusterEngine` already receives.
 - `domains/admission/quota` (dependency, moved from `handlers/quota` unchanged): `live.go` imports `gateway/elbv2` and `handlers/rds` for its live ELBv2 and RDS counters; `records.go` imports `runtime/compute/vm` for the instance-record vCPU sweep; `live.go`, `vcpu.go` and `storage.go` call `domains/ec2/awsapi/{vpc,eip,volume}` describe adapters to count VPCs, subnets, addresses and volumes, as they did under `gateway/ec2`. The ADR's `handlers/quota` current-to-target map row replaces these with reader capabilities once their final locations exist.
 - `domains/ec2/awsapi/instance` (dependency): imports `handlers/iam` (`IAMService` in `DescribeInstances.go`, `RunInstances.go` and `IamInstanceProfileAssociation.go`; `InstanceProfile` in the latter two), `runtime/compute/cache` (`NodeState`, `NodeUnknown`, `NodeLive`) and `runtime/compute/vm` (`VM`), both in `status_synthesis.go`. The IAM edge needs an IAM capability; the runtime edges are domain-to-runtime and need the instance record and node liveness exposed through an EC2-owned projection.
 - `domains/ec2/awsapi/spotinstance` (dependency): imports `handlers/iam.IAMService` and `domains/admission/quota.Service` in `RequestSpotInstances.go`.

docs/terraform-workbooks/demo-app/Dockerfile

@@ -4,7 +4,7 @@
 # distroless base (no shell, no package manager) for a small, production-shaped
 # image suitable for pushing to ECR and pulling onto EKS workers.
 
-FROM golang:1.27.2-alpine@sha256:f92b6ef800e499660581efdabdf25d9d817a9d124eaf900924f0504e7e27e12d AS build
+FROM golang:1.27.3-alpine@sha256:f2c872bcee949b7be9c490efd02b48a8a2fbf947cdf533f1c58525ec9d8170fd AS build
 WORKDIR /src
 COPY go.mod ./
 COPY main.go ./

docs/terraform-workbooks/demo-app/go.mod

@@ -1,3 +1,3 @@
 module github.com/mulgadc/spinifex-demo
 
-go 1.27.2
+go 1.27.3

go.mod

@@ -1,6 +1,6 @@
 module github.com/mulgadc/spinifex
 
-go 1.27.2
+go 1.27.3
 
 require (
 	github.com/aws/aws-sdk-go v1.55.8

govulncheck-accepted.txt

@@ -1,3 +0,0 @@
-# Reachable vulnerabilities accepted for a bounded time, read by scripts/govulncheck-gate.sh.
-# Format: <OSV-ID> <expiry YYYY-MM-DD> <reason>
-# The gate fails once the expiry passes: fix the dependency or renew deliberately.

scripts/build-microvm-image.sh

@@ -17,7 +17,7 @@ BUILD_DIR="$REPO_ROOT/build/microvm"
 INIT_SH="$BUILD_DIR/init.sh"
 INITTAB="$BUILD_DIR/inittab"
 
-ALPINE_VERSION="${ALPINE_VERSION:-3.20}"
+ALPINE_VERSION="${ALPINE_VERSION:-3.24}"
 ALPINE_MIRROR="${ALPINE_MIRROR:-https://dl-cdn.alpinelinux.org/alpine}"
 # haproxy: ALB (L7) data plane. nginx + nginx-mod-stream: NLB (L4) data plane —
 # the `stream` module (a separate Alpine package) load-balances TCP/UDP/TLS,

scripts/build-system-image.sh

@@ -284,6 +284,12 @@ if [[ "$DISTRO" == "alpine" ]] && [[ -n "${APK_PACKAGES:-}" ]]; then
     CUST+=(--run-command 'sed -i "s|^#\(.*community\)|\1|" /etc/apk/repositories 2>/dev/null || true')
     CUST+=(--run-command 'grep -q community /etc/apk/repositories || grep main /etc/apk/repositories | head -1 | sed "s|/main|/community|" >> /etc/apk/repositories')
     CUST+=(--run-command 'apk update')
+    # The base cloud image's kernel is only as fresh as its release; upgrade so
+    # the AMI ships the branch's current linux-virt and security fixes.
+    CUST+=(--run-command 'apk upgrade')
+    # apk's mkinitfs trigger does not always fire in the appliance; a new kernel
+    # with the old initramfs cannot load virtio_blk and never mounts root.
+    CUST+=(--run-command 'k=$(ls /lib/modules | sort -V | tail -1) && mkinitfs "$k" && zcat /boot/initramfs-virt | cpio -t | grep -qE "^(usr/)?lib/modules/$k/.*/virtio_blk\.ko"')
     CUST+=(--run-command "apk add ${APK_PACKAGES}")
 elif [[ "$DISTRO" == "ubuntu" ]] && [[ -n "${APT_PACKAGES:-}" ]]; then
     echo "Will install packages: ${APT_PACKAGES}"

scripts/govulncheck-gate.sh

@@ -1,50 +0,0 @@
-#!/usr/bin/env bash
-# Fails on any reachable vulnerability that is not listed in
-# govulncheck-accepted.txt, and on any acceptance whose expiry has passed.
-# Each line is "<OSV-ID> <YYYY-MM-DD expiry> <reason>"; '#' starts a comment.
-# GOVULNCHECK_JSON substitutes a saved -format json report, for tests.
-set -euo pipefail
-
-accepted_file=${GOVULNCHECK_ACCEPTED:-govulncheck-accepted.txt}
-today=${GOVULNCHECK_TODAY:-$(date -u +%F)}
-
-if [ -n "${GOVULNCHECK_JSON:-}" ]; then
-	report=$(cat "$GOVULNCHECK_JSON")
-else
-	report=$(go tool govulncheck -format json ./...)
-fi
-
-# A finding whose first trace frame names a function is reachable from our
-# code; text mode reports only these, so module- and package-level ones pass.
-mapfile -t called < <(printf '%s\n' "$report" | jq -rs \
-	'[.[] | select(.finding) | .finding | select(.trace[0].function) | .osv] | unique | .[]')
-
-declare -A expiry reason
-if [ -f "$accepted_file" ]; then
-	while read -r id until why; do
-		[[ -z "$id" || "$id" == \#* ]] && continue
-		expiry[$id]=$until
-		reason[$id]=$why
-	done <"$accepted_file"
-fi
-
-fail=0
-for id in "${called[@]}"; do
-	if [ -z "${expiry[$id]:-}" ]; then
-		echo "  $id is reachable and not accepted; run 'go tool govulncheck ./...' for the trace"
-		fail=1
-	elif [[ "$today" > "${expiry[$id]}" ]]; then
-		echo "  $id acceptance expired on ${expiry[$id]}: ${reason[$id]}"
-		fail=1
-	else
-		echo "  $id accepted until ${expiry[$id]}: ${reason[$id]}"
-	fi
-done
-
-for id in "${!expiry[@]}"; do
-	if ! printf '%s\n' "${called[@]}" | grep -qx "$id"; then
-		echo "  $id is accepted but no longer reported; remove it from $accepted_file"
-	fi
-done
-
-exit "$fail"

scripts/govulncheck-gate_test.sh

@@ -1,53 +0,0 @@
-#!/bin/sh
-# Tests the accepted-vulnerability gate against saved reports, so the logic is
-# checked without running the scanner.
-set -u
-
-here=$(cd "$(dirname "$0")" && pwd)
-gate="$here/govulncheck-gate.sh"
-tmp=$(mktemp -d)
-trap 'rm -rf "$tmp"' EXIT
-
-cat >"$tmp/accepted.txt" <<'EOF'
-# comment
-GO-2026-6597 2026-11-07 accepted for the test
-EOF
-cat >"$tmp/reachable.json" <<'EOF'
-{"config":{}}
-{"finding":{"osv":"GO-2026-6597","trace":[{"module":"m","package":"p","function":"Pull"}]}}
-{"finding":{"osv":"GO-2026-0001","trace":[{"module":"m"}]}}
-EOF
-cat >"$tmp/new.json" <<'EOF'
-{"finding":{"osv":"GO-2026-6597","trace":[{"module":"m","package":"p","function":"Pull"}]}}
-{"finding":{"osv":"GO-2026-7777","trace":[{"module":"m","package":"p","function":"F"}]}}
-EOF
-echo '{"config":{}}' >"$tmp/clean.json"
-
-fails=0
-check() {
-	want=$1 name=$2 report=$3 today=$4 accepted=$5 expect=$6
-	out=$(GOVULNCHECK_JSON="$tmp/$report" GOVULNCHECK_TODAY=$today GOVULNCHECK_ACCEPTED=$accepted bash "$gate" 2>&1)
-	rc=$?
-	if [ "$rc" != "$want" ] || ! printf '%s' "$out" | grep -q "$expect"; then
-		echo "FAIL [$name] rc=$rc want=$want output: $out"
-		fails=$((fails + 1))
-	else
-		echo "ok   [$name]"
-	fi
-}
-
-check 0 "accepted reachable finding passes" reachable.json 2026-10-07 "$tmp/accepted.txt" "GO-2026-6597 accepted until 2026-11-07"
-check 0 "expiry day itself still passes" reachable.json 2026-11-07 "$tmp/accepted.txt" "accepted until"
-check 1 "expired acceptance fails" reachable.json 2026-11-08 "$tmp/accepted.txt" "acceptance expired on 2026-11-07"
-check 1 "new reachable finding fails" new.json 2026-10-07 "$tmp/accepted.txt" "GO-2026-7777 is reachable and not accepted"
-out=$(GOVULNCHECK_JSON="$tmp/reachable.json" GOVULNCHECK_TODAY=2026-10-07 GOVULNCHECK_ACCEPTED="$tmp/accepted.txt" bash "$gate" 2>&1)
-if printf '%s' "$out" | grep -q "GO-2026-0001"; then
-	echo "FAIL [module-level finding is not reachable] output: $out"
-	fails=$((fails + 1))
-else
-	echo "ok   [module-level finding is not reachable]"
-fi
-check 0 "fixed finding warns that acceptance is stale" clean.json 2026-10-07 "$tmp/accepted.txt" "no longer reported"
-check 1 "missing acceptance file fails reachable finding" reachable.json 2026-10-07 "$tmp/none" "not accepted"
-
-[ "$fails" -eq 0 ] || exit 1

scripts/setup-ovn_test.sh

@@ -43,7 +43,13 @@ esac
 # Teardown, dry run. Every mutation must go through run(), so a dry run can
 # report the whole plan without performing any of it.
 # --------------------------------------------------------------------------
-out=$("$SETUP" --teardown --dry-run 2>&1) && rc=0 || rc=$?
+# The teardown plans the veth steps only when ovs-vsctl exists and the link is
+# present, so stub both or the order check depends on the machine running it.
+mkdir "$TMP/bin"
+printf '#!/bin/sh\nexit 0\n' > "$TMP/bin/ip"
+cp "$TMP/bin/ip" "$TMP/bin/ovs-vsctl"
+chmod +x "$TMP/bin/ip" "$TMP/bin/ovs-vsctl"
+out=$(PATH="$TMP/bin:$PATH" "$SETUP" --teardown --dry-run 2>&1) && rc=0 || rc=$?
 [ "$rc" = 0 ] && pass "--teardown --dry-run exits 0" || fail "--teardown --dry-run exited $rc, want 0"
 
 case "$out" in

scripts/uninstall-spx_test.sh

@@ -81,9 +81,15 @@ mkcfg() { # mkcfg <dir> <node>...
 mkcfg "$TMP/three" node1 node2 node3
 mkcfg "$TMP/one" node1
 
+# The node count is read through sudo, so a pass-through stub keeps the refusal
+# from depending on whether the machine running this has sudo.
+mkdir "$TMP/bin"
+printf '#!/bin/sh\nexec "$@"\n' > "$TMP/bin/sudo"
+chmod +x "$TMP/bin/sudo"
+
 for args in "" "--yes" "--yes --dry-run"; do
     # shellcheck disable=SC2086
-    SPX_ETC_DIR="$TMP/three" "$UNINSTALL" $args >"$TMP/out" 2>"$TMP/err" && rc=0 || rc=$?
+    PATH="$TMP/bin:$PATH" SPX_ETC_DIR="$TMP/three" "$UNINSTALL" $args >"$TMP/out" 2>"$TMP/err" && rc=0 || rc=$?
     label=${args:-"(no flags)"}
     [ "$rc" = 1 ] \
         && pass "three-node config refuses with $label" \

spinifex/accountteardown/cluster.go

@@ -70,6 +70,11 @@ func NewClusterEngine(
 	reapers = append(reapers, S3Reapers(buckets)...)
 	reapers = append(reapers, ACMReapers(nc)...)
 	reapers = append(reapers, IAMReapers(svc)...)
+
+	// Before the account counters and reservations: the seal must still be
+	// able to find the account IAM holds as TERMINATING, which the account
+	// record reaper does not remove.
+	reapers = append(reapers, EC2AccountSettingsReapers(nc, nodeIDs)...)
 	reapers = append(reapers, AccountReapers(svc, names, usage)...)
 
 	engine := NewEngine(NewIAMAccounts(svc), reapers...)

spinifex/accountteardown/cluster_test.go

@@ -41,6 +41,27 @@ func TestNewClusterEngineCoversEveryStage(t *testing.T) {
 	}
 }
 
+// The EC2 settings seal reaper must be registered in StageAccount on every
+// surface that builds a cluster engine, or teardown would finalise accounts
+// without ever sealing their settings.
+func TestNewClusterEngineRegistersTheEC2SettingsSealReaper(t *testing.T) {
+	_, nc, _ := testutil.StartTestJetStream(t)
+
+	masterKey, err := handlers_iam.GenerateMasterKey()
+	require.NoError(t, err)
+	svc, err := handlers_iam.NewIAMServiceImpl(t.Context(), nc, masterKey)
+	require.NoError(t, err)
+
+	engine, err := NewClusterEngine(t.Context(), nc, 1, []string{"node-1"}, svc, newFakeBuckets())
+	require.NoError(t, err)
+
+	var kinds []string
+	for _, reaper := range engine.reapersFor(StageAccount) {
+		kinds = append(kinds, reaper.Kind())
+	}
+	assert.Contains(t, kinds, "ec2-account-settings")
+}
+
 // A teardown must refuse the accounts no credential may delete before it reads
 // anything, and refuse a name confirmation that does not match the record.
 func TestClusterEnginePrechecks(t *testing.T) {

spinifex/accountteardown/ec2_account_settings_seal.go

@@ -0,0 +1,125 @@
+package accountteardown
+
+import (
+	"context"
+	"errors"
+	"fmt"
+	"log/slog"
+	"sync"
+	"time"
+
+	ec2v1 "github.com/mulgadc/spinifex/contracts/ec2/v1"
+	natsmsg "github.com/mulgadc/spinifex/spinifex/foundation/messaging/nats"
+	"github.com/nats-io/nats.go"
+)
+
+// ec2AccountSettingsSealTimeout bounds one seal request; an earlier caller
+// deadline wins.
+const ec2AccountSettingsSealTimeout = 30 * time.Second
+
+// ec2AccountSettingsSealer is EC2's seal of one account's settings record.
+type ec2AccountSettingsSealer interface {
+	SealAccountSettings(ctx context.Context, accountID string) (revision uint64, err error)
+}
+
+// ec2AccountSettingsSealClient reaches the EC2 settings-seal contract over
+// NATS. Only a reply for this account, disposition sealed and a non-zero
+// revision is a seal; anything else, including no responder, is an error.
+type ec2AccountSettingsSealClient struct {
+	nc      *nats.Conn
+	timeout time.Duration
+}
+
+var _ ec2AccountSettingsSealer = (*ec2AccountSettingsSealClient)(nil)
+
+func newEC2AccountSettingsSealClient(nc *nats.Conn) *ec2AccountSettingsSealClient {
+	return &ec2AccountSettingsSealClient{nc: nc, timeout: ec2AccountSettingsSealTimeout}
+}
+
+func (c *ec2AccountSettingsSealClient) SealAccountSettings(ctx context.Context, accountID string) (uint64, error) {
+	timeout := c.timeout
+	if deadline, ok := ctx.Deadline(); ok {
+		timeout = min(timeout, time.Until(deadline))
+	}
+	if ctx.Err() != nil || timeout <= 0 {
+		return 0, fmt.Errorf("seal account settings: %w", context.DeadlineExceeded)
+	}
+
+	reply, err := natsmsg.NATSRequest[ec2v1.SealAccountSettingsReply](ctx, c.nc, ec2v1.SealAccountSettingsSubject,
+		&ec2v1.SealAccountSettingsRequest{AccountID: accountID}, timeout, accountID)
+	if err != nil {
+		return 0, fmt.Errorf("seal account settings: %w", err)
+	}
+	if reply.AccountID != accountID {
+		return 0, fmt.Errorf("seal account settings: reply named account %q, not %q", reply.AccountID, accountID)
+	}
+	if reply.Disposition != ec2v1.AccountSettingsDispositionSealed {
+		return 0, fmt.Errorf("seal account settings: unexpected disposition %q", reply.Disposition)
+	}
+	if reply.Revision == 0 {
+		return 0, errors.New("seal account settings: reply carried no revision")
+	}
+	return reply.Revision, nil
+}
+
+// accountSettingsSealReaper seals EC2's settings record in StageAccount. List
+// does no I/O and stays pending until this instance has sealed, so DryRun never
+// seals; state is per engine, so every fresh teardown run seals again.
+type accountSettingsSealReaper struct {
+	sealer     ec2AccountSettingsSealer
+	capability sealCapabilityChecker
+
+	mu     sync.Mutex
+	sealed map[string]uint64 // account ID -> leader revision, this run only
+}
+
+func (r *accountSettingsSealReaper) Kind() string { return "ec2-account-settings" }
+func (r *accountSettingsSealReaper) Stage() Stage { return StageAccount }
+
+func (r *accountSettingsSealReaper) alreadySealed(accountID string) bool {
+	r.mu.Lock()
+	defer r.mu.Unlock()
+	_, ok := r.sealed[accountID]
+	return ok
+}
+
+func (r *accountSettingsSealReaper) List(_ context.Context, accountID string) ([]Resource, error) {
+	if r.alreadySealed(accountID) {
+		return nil, nil
+	}
+	return []Resource{{Kind: r.Kind(), ID: accountID, Detail: "seal"}}, nil
+}
+
+// Delete probes every configured daemon immediately before each seal attempt,
+// never reusing an earlier probe. Incomplete capability sends no seal and is
+// retryable; force is ignored, since nothing may force past the probe.
+func (r *accountSettingsSealReaper) Delete(ctx context.Context, accountID string, _ Resource, _ bool) error {
+	report := r.capability.CheckSealCapability(ctx, accountID)
+	if !report.Complete {
+		return errSealCapabilityIncomplete(report)
+	}
+
+	revision, err := r.sealer.SealAccountSettings(ctx, accountID)
+	if err != nil {
+		return err
+	}
+
+	r.mu.Lock()
+	if r.sealed == nil {
+		r.sealed = map[string]uint64{}
+	}
+	r.sealed[accountID] = revision
+	r.mu.Unlock()
+	slog.InfoContext(ctx, "Account teardown sealed EC2 account settings", "accountID", accountID, "revision", revision)
+	return nil
+}
+
+// EC2AccountSettingsReapers returns the StageAccount reaper that seals EC2's
+// settings record. nodeIDs is the configured node set the capability probe
+// treats as "every relevant writer"; an empty set is unknown, never complete.
+func EC2AccountSettingsReapers(nc *nats.Conn, nodeIDs []string) []Reaper {
+	return []Reaper{&accountSettingsSealReaper{
+		sealer:     newEC2AccountSettingsSealClient(nc),
+		capability: newSealCapabilityClient(nc, nodeIDs),
+	}}
+}

spinifex/accountteardown/ec2_account_settings_seal_capability.go

@@ -0,0 +1,190 @@
+package accountteardown
+
+import (
+	"context"
+	"encoding/json"
+	"sort"
+	"strings"
+	"sync"
+	"time"
+
+	ec2v1 "github.com/mulgadc/spinifex/contracts/ec2/v1"
+	awserrors "github.com/mulgadc/spinifex/spinifex/foundation/aws/errors"
+	natsmsg "github.com/mulgadc/spinifex/spinifex/foundation/messaging/nats"
+	"github.com/nats-io/nats.go"
+)
+
+// sealCapabilityWait bounds the whole probe; shortened by the caller's
+// deadline. Every configured node is probed in parallel against this one
+// wait, so N nodes cost one wait, not N.
+const sealCapabilityWait = 3 * time.Second
+
+// sealCapabilityGrace is how long a node's probe keeps listening after its
+// one expected responder has answered, long enough for a genuine second
+// subscriber on the same node to be caught as a duplicate.
+const sealCapabilityGrace = 150 * time.Millisecond
+
+// Seal capability reasons, one per failing node. Each names a category, never
+// wire error text.
+const (
+	sealCapabilityReasonNoReply       = "no-reply"
+	sealCapabilityReasonDuplicate     = "duplicate"
+	sealCapabilityReasonWrongIdentity = "wrong-identity"
+	sealCapabilityReasonErrorReply    = "error-reply"
+	sealCapabilityReasonMalformed     = "malformed"
+	sealCapabilityReasonOldVersion    = "old-version"
+	sealCapabilityReasonTransport     = "transport"
+	sealCapabilityReasonInvalidNodeID = "invalid-node-id"
+	sealCapabilityReasonSetUnknown    = "configured-node-set-unknown"
+)
+
+// sealCapabilityChecker reports whether every configured daemon supports the
+// seal.
+type sealCapabilityChecker interface {
+	CheckSealCapability(ctx context.Context, correlationRef string) sealCapabilityReport
+}
+
+// sealCapabilityReport is one probe's result, one entry per configured node.
+type sealCapabilityReport struct {
+	Complete bool
+	Nodes    []nodeSealCapability // one per configured node, sorted by node
+}
+
+// nodeSealCapability is one configured node's compatibility. Reason is ""
+// when the node is compatible.
+type nodeSealCapability struct {
+	Node    string
+	Version int    // 0 unless a valid reply arrived
+	Reason  string // "" when compatible
+}
+
+// sealCapabilityClient probes every configured node's seal-capability subject
+// in parallel. "Configured" is the same node set NewClusterEngine already
+// receives; there is no separate liveness or membership source.
+type sealCapabilityClient struct {
+	nc      *nats.Conn
+	nodeIDs []string
+	wait    time.Duration // 3 s, shortened by the caller's deadline
+	grace   time.Duration // 150 ms ResponderGrace
+}
+
+var _ sealCapabilityChecker = (*sealCapabilityClient)(nil)
+
+func newSealCapabilityClient(nc *nats.Conn, nodeIDs []string) *sealCapabilityClient {
+	return &sealCapabilityClient{nc: nc, nodeIDs: nodeIDs, wait: sealCapabilityWait, grace: sealCapabilityGrace}
+}
+
+// CheckSealCapability probes every configured node in parallel, bounded by
+// one shared deadline, run immediately before each seal attempt. A cached or
+// earlier result is never reused.
+func (c *sealCapabilityClient) CheckSealCapability(ctx context.Context, correlationRef string) sealCapabilityReport {
+	if len(c.nodeIDs) == 0 {
+		return sealCapabilityReport{Complete: false}
+	}
+
+	wait := c.wait
+	if deadline, ok := ctx.Deadline(); ok {
+		wait = min(wait, time.Until(deadline))
+	}
+
+	nodes := append([]string(nil), c.nodeIDs...)
+	sort.Strings(nodes)
+
+	results := make([]nodeSealCapability, len(nodes))
+	if ctx.Err() != nil || wait <= 0 {
+		for i, node := range nodes {
+			results[i] = nodeSealCapability{Node: node, Reason: sealCapabilityReasonNoReply}
+		}
+		return sealCapabilityReport{Complete: false, Nodes: results}
+	}
+
+	var wg sync.WaitGroup
+	for i, node := range nodes {
+		wg.Add(1)
+		go func(i int, node string) {
+			defer wg.Done()
+			results[i] = c.checkNode(ctx, node, correlationRef, wait)
+		}(i, node)
+	}
+	wg.Wait()
+
+	complete := true
+	for _, r := range results {
+		if r.Reason != "" {
+			complete = false
+			break
+		}
+	}
+	return sealCapabilityReport{Complete: complete, Nodes: results}
+}
+
+// checkNode probes one node and classifies the result. A node is compatible
+// only when exactly one identified, decodable, matching, version >= 1 frame
+// arrived, with no duplicate, conflict or cap hit.
+func (c *sealCapabilityClient) checkNode(ctx context.Context, node, correlationRef string, wait time.Duration) nodeSealCapability {
+	if !ec2v1.ValidSealCapabilityNode(node) {
+		return nodeSealCapability{Node: node, Reason: sealCapabilityReasonInvalidNodeID}
+	}
+
+	data, err := json.Marshal(ec2v1.AccountSettingsSealCapabilityRequest{Node: node, CorrelationRef: correlationRef})
+	if err != nil {
+		return nodeSealCapability{Node: node, Reason: sealCapabilityReasonMalformed}
+	}
+
+	frames, sum, err := natsmsg.Gather(ctx, c.nc, ec2v1.AccountSettingsSealCapabilitySubject(node), data, natsmsg.GatherOpts{
+		Timeout:            wait,
+		ExpectedResponders: 1,
+		Mode:               natsmsg.CollectUntilDeadline,
+		ResponderGrace:     c.grace,
+	})
+	if err != nil {
+		return nodeSealCapability{Node: node, Reason: sealCapabilityReasonTransport}
+	}
+
+	switch {
+	case sum.CapHit, len(sum.ConflictNodes) > 0, sum.DuplicateFrames > 0:
+		return nodeSealCapability{Node: node, Reason: sealCapabilityReasonDuplicate}
+	case sum.Unidentified > 0:
+		return nodeSealCapability{Node: node, Reason: sealCapabilityReasonWrongIdentity}
+	case len(sum.ErrorResponders) > 0:
+		return nodeSealCapability{Node: node, Reason: sealCapabilityReasonErrorReply}
+	case sum.Received == 0, len(frames) == 0:
+		return nodeSealCapability{Node: node, Reason: sealCapabilityReasonNoReply}
+	}
+
+	frame := frames[0]
+	if frame.NodeID != node {
+		return nodeSealCapability{Node: node, Reason: sealCapabilityReasonWrongIdentity}
+	}
+
+	var reply ec2v1.AccountSettingsSealCapabilityReply
+	if json.Unmarshal(frame.Data, &reply) != nil {
+		return nodeSealCapability{Node: node, Reason: sealCapabilityReasonMalformed}
+	}
+	if reply.Node != node {
+		return nodeSealCapability{Node: node, Reason: sealCapabilityReasonWrongIdentity}
+	}
+	if reply.CapabilityVersion < 1 {
+		return nodeSealCapability{Node: node, Version: reply.CapabilityVersion, Reason: sealCapabilityReasonOldVersion}
+	}
+	return nodeSealCapability{Node: node, Version: reply.CapabilityVersion}
+}
+
+// errSealCapabilityIncomplete builds the retryable refusal from report. It
+// names node IDs and reason categories only, never wire error text.
+func errSealCapabilityIncomplete(report sealCapabilityReport) error {
+	detail := sealCapabilityReasonSetUnknown
+	var parts []string
+	for _, n := range report.Nodes {
+		if n.Reason != "" {
+			parts = append(parts, n.Node+"="+n.Reason)
+		}
+	}
+	if len(parts) > 0 {
+		detail = strings.Join(parts, ", ")
+	}
+	return awserrors.WithRetryAfter(
+		awserrors.Errorf(awserrors.ErrorServiceUnavailable, "EC2 account-settings seal capability incomplete: %s", detail),
+		5*time.Second,
+	)
+}

spinifex/accountteardown/ec2_account_settings_seal_capability_test.go

@@ -0,0 +1,311 @@
+package accountteardown
+
+//test:in-package — the probe client, its report types and reason constants
+// are unexported, and what they guarantee (incomplete never seals, complete
+// does) is only observable from inside the package.
+
+import (
+	"encoding/json"
+	"sync/atomic"
+	"testing"
+	"time"
+
+	ec2v1 "github.com/mulgadc/spinifex/contracts/ec2/v1"
+	"github.com/mulgadc/spinifex/internal/testkit"
+	awserrors "github.com/mulgadc/spinifex/spinifex/foundation/aws/errors"
+	natsmsg "github.com/mulgadc/spinifex/spinifex/foundation/messaging/nats"
+	"github.com/nats-io/nats.go"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+// testCapabilityClient builds a probe client with test-sized timeouts so the
+// incomplete cases below do not each pay the production 3 s wait.
+func testCapabilityClient(nc *nats.Conn, nodeIDs []string) *sealCapabilityClient {
+	c := newSealCapabilityClient(nc, nodeIDs)
+	c.wait = 200 * time.Millisecond
+	c.grace = 30 * time.Millisecond
+	return c
+}
+
+// serveCapability subscribes reply on subject with no queue group, which is
+// how the real responder is addressed: per node, so two subscribers on the
+// same node's subject both answer.
+func serveCapability(t *testing.T, nc *nats.Conn, subject string, reply nats.MsgHandler) *nats.Subscription {
+	t.Helper()
+	sub, err := nc.Subscribe(subject, reply)
+	require.NoError(t, err)
+	t.Cleanup(func() { _ = sub.Unsubscribe() })
+	require.NoError(t, nc.Flush())
+	return sub
+}
+
+// capabilityReply answers with a well-formed reply naming replyNode (which
+// may deliberately differ from the subject's node) and headerNode as the
+// X-Node-ID ("" omits the header).
+func capabilityReply(replyNode, headerNode string, version int) nats.MsgHandler {
+	return func(msg *nats.Msg) {
+		data, _ := json.Marshal(ec2v1.AccountSettingsSealCapabilityReply{Node: replyNode, CapabilityVersion: version})
+		reply := nats.NewMsg(msg.Reply)
+		reply.Data = data
+		if headerNode != "" {
+			reply.Header.Set(natsmsg.NodeIDHeader, headerNode)
+		}
+		_ = msg.RespondMsg(reply)
+	}
+}
+
+func nodeResult(t *testing.T, report sealCapabilityReport, node string) nodeSealCapability {
+	t.Helper()
+	for _, n := range report.Nodes {
+		if n.Node == node {
+			return n
+		}
+	}
+	t.Fatalf("no result for node %q in %+v", node, report)
+	return nodeSealCapability{}
+}
+
+// Each sub-case names exactly one way a node fails to prove capability
+// version 1, and that failure alone makes the whole probe incomplete.
+func TestSealCapabilityIncomplete(t *testing.T) {
+	t.Run("missing", func(t *testing.T) {
+		_, nc, _ := testutil.StartTestJetStream(t)
+		report := testCapabilityClient(nc, []string{"node-1"}).CheckSealCapability(testCtx(t), "req")
+		assert.False(t, report.Complete)
+		assert.Equal(t, sealCapabilityReasonNoReply, nodeResult(t, report, "node-1").Reason)
+	})
+
+	t.Run("duplicate", func(t *testing.T) {
+		_, nc, _ := testutil.StartTestJetStream(t)
+		serveCapability(t, nc, ec2v1.AccountSettingsSealCapabilitySubject("node-1"), capabilityReply("node-1", "node-1", 1))
+		serveCapability(t, nc, ec2v1.AccountSettingsSealCapabilitySubject("node-1"), capabilityReply("node-1", "node-1", 1))
+		report := testCapabilityClient(nc, []string{"node-1"}).CheckSealCapability(testCtx(t), "req")
+		assert.False(t, report.Complete)
+		assert.Equal(t, sealCapabilityReasonDuplicate, nodeResult(t, report, "node-1").Reason)
+	})
+
+	t.Run("unreachable timeout", func(t *testing.T) {
+		_, nc, _ := testutil.StartTestJetStream(t)
+		serveCapability(t, nc, ec2v1.AccountSettingsSealCapabilitySubject("node-1"), func(msg *nats.Msg) {
+			time.Sleep(time.Second)
+			_ = msg.Respond(nil)
+		})
+		client := testCapabilityClient(nc, []string{"node-1"})
+		report := client.CheckSealCapability(testCtx(t), "req")
+		assert.False(t, report.Complete)
+		assert.Equal(t, sealCapabilityReasonNoReply, nodeResult(t, report, "node-1").Reason)
+	})
+
+	t.Run("transport closed connection", func(t *testing.T) {
+		_, nc, _ := testutil.StartTestJetStream(t)
+		nc.Close()
+		report := testCapabilityClient(nc, []string{"node-1"}).CheckSealCapability(testCtx(t), "req")
+		assert.False(t, report.Complete)
+		assert.Equal(t, sealCapabilityReasonTransport, nodeResult(t, report, "node-1").Reason)
+	})
+
+	t.Run("malformed body", func(t *testing.T) {
+		_, nc, _ := testutil.StartTestJetStream(t)
+		serveCapability(t, nc, ec2v1.AccountSettingsSealCapabilitySubject("node-1"), func(msg *nats.Msg) {
+			reply := nats.NewMsg(msg.Reply)
+			reply.Data = []byte("not json")
+			reply.Header.Set(natsmsg.NodeIDHeader, "node-1")
+			_ = msg.RespondMsg(reply)
+		})
+		report := testCapabilityClient(nc, []string{"node-1"}).CheckSealCapability(testCtx(t), "req")
+		assert.False(t, report.Complete)
+		assert.Equal(t, sealCapabilityReasonMalformed, nodeResult(t, report, "node-1").Reason)
+	})
+
+	t.Run("error reply", func(t *testing.T) {
+		_, nc, _ := testutil.StartTestJetStream(t)
+		serveCapability(t, nc, ec2v1.AccountSettingsSealCapabilitySubject("node-1"), func(msg *nats.Msg) {
+			reply := nats.NewMsg(msg.Reply)
+			reply.Data = awserrors.GenerateErrorPayload(awserrors.ErrorServerInternal)
+			reply.Header.Set(natsmsg.NodeIDHeader, "node-1")
+			_ = msg.RespondMsg(reply)
+		})
+		report := testCapabilityClient(nc, []string{"node-1"}).CheckSealCapability(testCtx(t), "req")
+		assert.False(t, report.Complete)
+		assert.Equal(t, sealCapabilityReasonErrorReply, nodeResult(t, report, "node-1").Reason)
+	})
+
+	t.Run("wrong identity no header", func(t *testing.T) {
+		_, nc, _ := testutil.StartTestJetStream(t)
+		serveCapability(t, nc, ec2v1.AccountSettingsSealCapabilitySubject("node-1"), capabilityReply("node-1", "", 1))
+		report := testCapabilityClient(nc, []string{"node-1"}).CheckSealCapability(testCtx(t), "req")
+		assert.False(t, report.Complete)
+		assert.Equal(t, sealCapabilityReasonWrongIdentity, nodeResult(t, report, "node-1").Reason)
+	})
+
+	t.Run("wrong identity another node header", func(t *testing.T) {
+		_, nc, _ := testutil.StartTestJetStream(t)
+		serveCapability(t, nc, ec2v1.AccountSettingsSealCapabilitySubject("node-1"), capabilityReply("node-1", "node-9", 1))
+		report := testCapabilityClient(nc, []string{"node-1"}).CheckSealCapability(testCtx(t), "req")
+		assert.False(t, report.Complete)
+		assert.Equal(t, sealCapabilityReasonWrongIdentity, nodeResult(t, report, "node-1").Reason)
+	})
+
+	t.Run("wrong identity body node differs", func(t *testing.T) {
+		_, nc, _ := testutil.StartTestJetStream(t)
+		serveCapability(t, nc, ec2v1.AccountSettingsSealCapabilitySubject("node-1"), capabilityReply("node-9", "node-1", 1))
+		report := testCapabilityClient(nc, []string{"node-1"}).CheckSealCapability(testCtx(t), "req")
+		assert.False(t, report.Complete)
+		assert.Equal(t, sealCapabilityReasonWrongIdentity, nodeResult(t, report, "node-1").Reason)
+	})
+
+	t.Run("lower version", func(t *testing.T) {
+		_, nc, _ := testutil.StartTestJetStream(t)
+		serveCapability(t, nc, ec2v1.AccountSettingsSealCapabilitySubject("node-1"), capabilityReply("node-1", "node-1", 0))
+		report := testCapabilityClient(nc, []string{"node-1"}).CheckSealCapability(testCtx(t), "req")
+		assert.False(t, report.Complete)
+		assert.Equal(t, sealCapabilityReasonOldVersion, nodeResult(t, report, "node-1").Reason)
+	})
+
+	t.Run("configured set empty", func(t *testing.T) {
+		_, nc, _ := testutil.StartTestJetStream(t)
+		report := testCapabilityClient(nc, nil).CheckSealCapability(testCtx(t), "req")
+		assert.False(t, report.Complete)
+		assert.Empty(t, report.Nodes)
+	})
+
+	t.Run("invalid node id", func(t *testing.T) {
+		_, nc, _ := testutil.StartTestJetStream(t)
+		report := testCapabilityClient(nc, []string{"bad.node"}).CheckSealCapability(testCtx(t), "req")
+		assert.False(t, report.Complete)
+		assert.Equal(t, sealCapabilityReasonInvalidNodeID, nodeResult(t, report, "bad.node").Reason)
+	})
+}
+
+// Every configured node answering version 1 or later makes the probe
+// complete, and the reaper sends exactly one seal request.
+func TestSealCapabilityCompleteAllowsSeal(t *testing.T) {
+	_, nc, _ := testutil.StartTestJetStream(t)
+	serveCapability(t, nc, ec2v1.AccountSettingsSealCapabilitySubject("node-1"), capabilityReply("node-1", "node-1", 1))
+	serveCapability(t, nc, ec2v1.AccountSettingsSealCapabilitySubject("node-2"), capabilityReply("node-2", "node-2", 2))
+
+	client := testCapabilityClient(nc, []string{"node-1", "node-2"})
+	report := client.CheckSealCapability(testCtx(t), "req")
+	require.True(t, report.Complete)
+	assert.Equal(t, 1, nodeResult(t, report, "node-1").Version)
+	assert.Equal(t, 2, nodeResult(t, report, "node-2").Version)
+
+	sealer := &fakeSealer{}
+	reaper := &accountSettingsSealReaper{sealer: sealer, capability: client}
+	require.NoError(t, reaper.Delete(testCtx(t), testAccountID, Resource{Kind: reaper.Kind(), ID: testAccountID}, false))
+	assert.Equal(t, []string{testAccountID}, sealer.calls)
+}
+
+// An incomplete probe keeps the account TERMINATING through the full engine,
+// names the silent node and its reason, sends no seal request, and leaves the
+// settings record untouched (the fake sealer records no call).
+func TestSealCapabilityIncompleteKeepsAccountTerminating(t *testing.T) {
+	_, nc, _ := testutil.StartTestJetStream(t)
+	serveCapability(t, nc, ec2v1.AccountSettingsSealCapabilitySubject("node-1"), capabilityReply("node-1", "node-1", 1))
+	// node-2 has no subscriber.
+
+	sealer := &fakeSealer{}
+	reaper := &accountSettingsSealReaper{sealer: sealer, capability: testCapabilityClient(nc, []string{"node-1", "node-2"})}
+	accounts := newFakeAccounts()
+	engine := quietEngine(accounts, reaper)
+
+	result, err := engine.Teardown(testCtx(t), request())
+	require.ErrorIs(t, err, ErrResourcesStuck)
+	assert.False(t, result.AccountDeleted)
+	assert.Equal(t, AccountStatusTerminating, accounts.account.Status)
+	assert.Empty(t, sealer.calls)
+
+	var reason string
+	for _, stage := range result.Stages {
+		for _, stuck := range stage.Stuck {
+			if stuck.Resource.Kind == "ec2-account-settings" {
+				reason = stuck.Reason
+			}
+		}
+	}
+	assert.Contains(t, reason, "node-2="+sealCapabilityReasonNoReply)
+}
+
+// The silent node starting to answer during the drain makes the very next
+// Delete re-probe and succeed, which proves the result is retryable and the
+// probe is never cached across attempts.
+func TestSealCapabilityRetriesThenSeals(t *testing.T) {
+	_, nc, _ := testutil.StartTestJetStream(t)
+	serveCapability(t, nc, ec2v1.AccountSettingsSealCapabilitySubject("node-1"), capabilityReply("node-1", "node-1", 1))
+
+	var node2Up atomic.Bool
+	sub, err := nc.Subscribe(ec2v1.AccountSettingsSealCapabilitySubject("node-2"), func(msg *nats.Msg) {
+		if !node2Up.Load() {
+			return
+		}
+		capabilityReply("node-2", "node-2", 1)(msg)
+	})
+	require.NoError(t, err)
+	t.Cleanup(func() { _ = sub.Unsubscribe() })
+	require.NoError(t, nc.Flush())
+
+	sealer := &fakeSealer{}
+	reaper := &accountSettingsSealReaper{sealer: sealer, capability: testCapabilityClient(nc, []string{"node-1", "node-2"})}
+	accounts := newFakeAccounts()
+	engine := quietEngine(accounts, reaper)
+	engine.Timeouts = Timeouts{StageDrain: time.Second, DrainPoll: 5 * time.Millisecond}
+
+	go func() {
+		time.Sleep(80 * time.Millisecond)
+		node2Up.Store(true)
+	}()
+
+	result, err := engine.Teardown(testCtx(t), request())
+	require.NoError(t, err)
+	assert.True(t, result.AccountDeleted)
+	assert.Equal(t, []string{testAccountID}, sealer.calls)
+}
+
+// Two Delete calls on one reaper probe twice: the probe runs in Delete
+// itself, never cached from a previous attempt.
+func TestSealCapabilityProbedBeforeEverySeal(t *testing.T) {
+	_, nc, _ := testutil.StartTestJetStream(t)
+	var probes int
+	serveCapability(t, nc, ec2v1.AccountSettingsSealCapabilitySubject("node-1"), func(msg *nats.Msg) {
+		probes++
+		capabilityReply("node-1", "node-1", 1)(msg)
+	})
+
+	sealer := &fakeSealer{}
+	client := testCapabilityClient(nc, []string{"node-1"})
+	reaper := &accountSettingsSealReaper{sealer: sealer, capability: client}
+
+	require.NoError(t, reaper.Delete(testCtx(t), testAccountID, Resource{Kind: reaper.Kind(), ID: testAccountID}, false))
+	// alreadySealed now makes List empty, but Delete itself does not consult
+	// it, so a second direct call still probes and seals again.
+	require.NoError(t, reaper.Delete(testCtx(t), testAccountID, Resource{Kind: reaper.Kind(), ID: testAccountID}, false))
+	assert.Equal(t, 2, probes)
+	assert.Len(t, sealer.calls, 2)
+}
+
+// Five silent nodes still return within one deadline, not five: every node
+// is probed in parallel against the same wait.
+func TestSealCapabilityBoundedByOneDeadline(t *testing.T) {
+	_, nc, _ := testutil.StartTestJetStream(t)
+	client := testCapabilityClient(nc, []string{"node-1", "node-2", "node-3", "node-4", "node-5"})
+
+	started := time.Now()
+	report := client.CheckSealCapability(testCtx(t), "req")
+	elapsed := time.Since(started)
+
+	assert.False(t, report.Complete)
+	assert.Less(t, elapsed, client.wait+time.Second, "five silent nodes must not cost five deadlines")
+}
+
+// The stuck reason is logged and stored as one line naming each failing node;
+// the retry hint and the ServiceUnavailable code travel with it.
+func TestSealCapabilityIncompleteErrorIsOneLine(t *testing.T) {
+	err := errSealCapabilityIncomplete(sealCapabilityReport{Nodes: []nodeSealCapability{{Node: "node-1", Reason: sealCapabilityReasonDuplicate}}})
+
+	assert.Equal(t, "EC2 account-settings seal capability incomplete: node-1=duplicate: ServiceUnavailable", err.Error())
+	d, ok := awserrors.ResolveRetryAfter(err)
+	assert.True(t, ok)
+	assert.Equal(t, 5*time.Second, d)
+	assert.True(t, awserrors.IsErrorCode(err, awserrors.ErrorServiceUnavailable))
+}

spinifex/accountteardown/ec2_account_settings_seal_test.go

@@ -0,0 +1,273 @@
+package accountteardown
+
+//test:in-package — the reaper, sealer and capability checker are unexported;
+// no seal without a complete probe, completion only after a sealed reply and
+// idempotent fresh runs are only observable from inside the package.
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"sync"
+	"testing"
+	"time"
+
+	"github.com/aws/aws-sdk-go/service/ec2"
+	ec2v1 "github.com/mulgadc/spinifex/contracts/ec2/v1"
+	"github.com/mulgadc/spinifex/internal/testkit"
+	ec2account "github.com/mulgadc/spinifex/spinifex/domains/ec2/account"
+	awserrors "github.com/mulgadc/spinifex/spinifex/foundation/aws/errors"
+	natsmsg "github.com/mulgadc/spinifex/spinifex/foundation/messaging/nats"
+	"github.com/nats-io/nats.go"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+// fakeSealer is a recording ec2AccountSettingsSealer. Each call beyond
+// failFor succeeds with a fresh, strictly increasing revision, unless err is
+// set, in which case every call after failFor fails with it.
+type fakeSealer struct {
+	mu       sync.Mutex
+	calls    []string
+	failFor  int
+	err      error
+	revision uint64
+}
+
+func (f *fakeSealer) SealAccountSettings(_ context.Context, accountID string) (uint64, error) {
+	f.mu.Lock()
+	defer f.mu.Unlock()
+	f.calls = append(f.calls, accountID)
+	if len(f.calls) <= f.failFor {
+		return 0, errors.New("account settings owner temporarily unavailable")
+	}
+	if f.err != nil {
+		return 0, f.err
+	}
+	f.revision++
+	return f.revision, nil
+}
+
+// fakeSealCapability is a recording sealCapabilityChecker returning a fixed report.
+type fakeSealCapability struct {
+	calls  int
+	report sealCapabilityReport
+}
+
+func (f *fakeSealCapability) CheckSealCapability(context.Context, string) sealCapabilityReport {
+	f.calls++
+	return f.report
+}
+
+// completeReport builds a sealCapabilityReport where every named node answers
+// compatible at version 1.
+func completeReport(nodeIDs ...string) sealCapabilityReport {
+	var nodes []nodeSealCapability
+	for _, node := range nodeIDs {
+		nodes = append(nodes, nodeSealCapability{Node: node, Version: 1})
+	}
+	return sealCapabilityReport{Complete: true, Nodes: nodes}
+}
+
+// alwaysTerminatingChecker stands in for IAM in the end-to-end test: the
+// account under test is always TERMINATING.
+type alwaysTerminatingChecker struct{}
+
+func (alwaysTerminatingChecker) AccountTerminating(context.Context, string) (bool, error) {
+	return true, nil
+}
+
+// No subscriber on either subject makes the capability probe incomplete, so
+// no seal request is ever sent and the account stays TERMINATING with one
+// stuck ec2-account-settings item.
+func TestSealReaperNoResponderKeepsAccountTerminating(t *testing.T) {
+	_, nc, _ := testutil.StartTestJetStream(t)
+
+	capability := newSealCapabilityClient(nc, []string{"node-1"})
+	capability.wait = 50 * time.Millisecond
+	capability.grace = 10 * time.Millisecond
+	sealer := newEC2AccountSettingsSealClient(nc)
+	sealer.timeout = 50 * time.Millisecond
+
+	reaper := &accountSettingsSealReaper{sealer: sealer, capability: capability}
+	accounts := newFakeAccounts()
+	engine := quietEngine(accounts, reaper)
+
+	result, err := engine.Teardown(testCtx(t), request())
+	require.ErrorIs(t, err, ErrResourcesStuck)
+	assert.False(t, result.AccountDeleted)
+	assert.Equal(t, AccountStatusTerminating, accounts.account.Status)
+
+	var stuck []Stuck
+	for _, stage := range result.Stages {
+		stuck = append(stuck, stage.Stuck...)
+	}
+	require.Len(t, stuck, 1)
+	assert.Equal(t, "ec2-account-settings", stuck[0].Resource.Kind)
+}
+
+// Whatever the owner refuses with, the account stays TERMINATING: an owner
+// refusal, a timeout, and each reply shape the client rejects (wrong account,
+// wrong disposition, no revision).
+func TestSealReaperOwnerFailureKeepsAccountTerminating(t *testing.T) {
+	cases := []struct {
+		name string
+		err  error
+	}{
+		{"incorrect-state", errors.New(awserrors.ErrorIncorrectState)},
+		{"internal-error", errors.New(awserrors.ErrorInternalError)},
+		{"timeout", context.DeadlineExceeded},
+		{"wrong-account", errors.New(`seal account settings: reply named account "999999999999", not "000000000042"`)},
+		{"wrong-disposition", errors.New(`seal account settings: unexpected disposition "pending"`)},
+		{"zero-revision", errors.New("seal account settings: reply carried no revision")},
+	}
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			sealer := &fakeSealer{err: tc.err}
+			reaper := &accountSettingsSealReaper{sealer: sealer, capability: &fakeSealCapability{report: completeReport("node-1")}}
+			accounts := newFakeAccounts()
+			engine := quietEngine(accounts, reaper)
+
+			result, err := engine.Teardown(testCtx(t), request())
+			require.ErrorIs(t, err, ErrResourcesStuck)
+			assert.False(t, result.AccountDeleted)
+			assert.Equal(t, AccountStatusTerminating, accounts.account.Status)
+			assert.NotEmpty(t, sealer.calls)
+		})
+	}
+}
+
+// The reaper retries through the engine's own drain mechanics: it fails
+// twice, then succeeds, and only the success lets the account be deleted.
+// List is non-empty before the success and empty only after it.
+func TestSealReaperCompletesOnlyAfterSealed(t *testing.T) {
+	sealer := &fakeSealer{failFor: 2}
+	reaper := &accountSettingsSealReaper{sealer: sealer, capability: &fakeSealCapability{report: completeReport("node-1")}}
+	accounts := newFakeAccounts()
+	engine := quietEngine(accounts, reaper)
+
+	before, err := reaper.List(testCtx(t), testAccountID)
+	require.NoError(t, err)
+	require.Len(t, before, 1)
+
+	result, err := engine.Teardown(testCtx(t), request())
+	require.NoError(t, err)
+	assert.True(t, result.AccountDeleted)
+	assert.GreaterOrEqual(t, len(sealer.calls), 3)
+
+	after, err := reaper.List(testCtx(t), testAccountID)
+	require.NoError(t, err)
+	assert.Empty(t, after)
+}
+
+// Each teardown builds a fresh engine and a fresh reaper, so a retried or
+// taken-over run seals again: two engines against the same owner both
+// succeed, each with a higher revision than the last.
+func TestSealReaperFreshRunSealsAgain(t *testing.T) {
+	sealer := &fakeSealer{}
+	capability := &fakeSealCapability{report: completeReport("node-1")}
+
+	firstEngine := quietEngine(newFakeAccounts(), &accountSettingsSealReaper{sealer: sealer, capability: capability})
+	firstResult, err := firstEngine.Teardown(testCtx(t), request())
+	require.NoError(t, err)
+	assert.True(t, firstResult.AccountDeleted)
+
+	secondEngine := quietEngine(newFakeAccounts(), &accountSettingsSealReaper{sealer: sealer, capability: capability})
+	secondResult, err := secondEngine.Teardown(testCtx(t), request())
+	require.NoError(t, err)
+	assert.True(t, secondResult.AccountDeleted)
+
+	require.Len(t, sealer.calls, 2)
+	assert.Equal(t, uint64(2), sealer.revision)
+}
+
+// Inventory and the explicit dry run, and a Teardown request marked dry run,
+// all list the pending seal and send no seal request: only a real teardown
+// may seal.
+func TestSealReaperInventoryDoesNotSeal(t *testing.T) {
+	sealer := &fakeSealer{}
+	capability := &fakeSealCapability{report: completeReport("node-1")}
+	reaper := &accountSettingsSealReaper{sealer: sealer, capability: capability}
+	accounts := newFakeAccounts()
+	engine := quietEngine(accounts, reaper)
+
+	inventory, err := engine.Inventory(testCtx(t), testAccountID)
+	require.NoError(t, err)
+	assert.Empty(t, sealer.calls)
+	assert.Zero(t, capability.calls)
+	found := false
+	for _, stage := range inventory.Stages {
+		for _, resource := range stage.Deleted {
+			if resource.Kind == "ec2-account-settings" {
+				found = true
+			}
+		}
+	}
+	assert.True(t, found, "inventory must still list the pending seal")
+
+	dryRun := request()
+	dryRun.DryRun = true
+	_, err = engine.Teardown(testCtx(t), dryRun)
+	require.NoError(t, err)
+	assert.Empty(t, sealer.calls)
+	assert.Zero(t, capability.calls)
+}
+
+// A real settings service subscribed on the real seal subject, behind a real
+// engine: teardown seals the record, replacing the characterisation test that
+// showed it used to survive.
+func TestTeardownSealsAccountSettings(t *testing.T) {
+	_, nc, js := testutil.StartTestJetStream(t)
+	ctx := testCtx(t)
+
+	settings, err := ec2account.NewAccountSettingsServiceImplWithNATS(ctx, nil, nc)
+	require.NoError(t, err)
+	settings.SetAccountTerminationChecker(alwaysTerminatingChecker{})
+	_, err = settings.EnableEbsEncryptionByDefault(ctx, &ec2.EnableEbsEncryptionByDefaultInput{}, testAccountID)
+	require.NoError(t, err)
+
+	sealSub, err := nc.QueueSubscribe(ec2v1.SealAccountSettingsSubject, "spinifex-workers", func(msg *nats.Msg) {
+		natsmsg.ServeNATSRequestCtx(msg, func(ctx context.Context, req *ec2v1.SealAccountSettingsRequest) (*ec2v1.SealAccountSettingsReply, error) {
+			seal, err := settings.SealAccount(ctx, req.AccountID)
+			if err != nil {
+				return nil, err
+			}
+			return &ec2v1.SealAccountSettingsReply{
+				AccountID: req.AccountID, Disposition: ec2v1.AccountSettingsDispositionSealed, Revision: seal.Revision,
+			}, nil
+		})
+	})
+	require.NoError(t, err)
+	t.Cleanup(func() { _ = sealSub.Unsubscribe() })
+
+	capSub, err := nc.Subscribe(ec2v1.AccountSettingsSealCapabilitySubject("node-1"), func(msg *nats.Msg) {
+		reply, err := json.Marshal(ec2v1.AccountSettingsSealCapabilityReply{Node: "node-1", CapabilityVersion: 1})
+		require.NoError(t, err)
+		out := nats.NewMsg("")
+		out.Data = reply
+		out.Header.Set(natsmsg.NodeIDHeader, "node-1")
+		require.NoError(t, msg.RespondMsg(out))
+	})
+	require.NoError(t, err)
+	t.Cleanup(func() { _ = capSub.Unsubscribe() })
+	require.NoError(t, nc.Flush())
+
+	accounts := newFakeAccounts()
+	engine := quietEngine(accounts, EC2AccountSettingsReapers(nc, []string{"node-1"})...)
+	// The real capability probe holds CollectUntilDeadline's full ResponderGrace
+	// (150 ms) even when the one expected node answers instantly, so the quiet
+	// engine's default 200 ms StageDrain leaves no room for the sweep after it.
+	engine.Timeouts = Timeouts{StageDrain: 2 * time.Second, DrainPoll: 10 * time.Millisecond}
+
+	result, err := engine.Teardown(ctx, request())
+	require.NoError(t, err)
+	assert.True(t, result.AccountDeleted)
+
+	bucket, err := js.KeyValue(ctx, ec2account.KVBucketAccountSettings)
+	require.NoError(t, err)
+	entry, err := bucket.Get(ctx, testAccountID)
+	require.NoError(t, err)
+	var rec ec2account.AccountSettingsRecord
+	require.NoError(t, json.Unmarshal(entry.Value(), &rec))
+	assert.Equal(t, ec2account.AccountSettingsStateSealed, rec.State)
+}

spinifex/accountteardown/ec2_platform_inventory.go

@@ -12,8 +12,8 @@ import (
 	"github.com/nats-io/nats.go"
 )
 
-// ec2PlatformInventoryTimeout bounds one inventory request. The scan reads
-// four stores for one account; a reply later than this is scan-incomplete.
+// ec2PlatformInventoryTimeout bounds one inventory request; an earlier caller
+// deadline wins. A reply later than either is scan-incomplete.
 const ec2PlatformInventoryTimeout = 30 * time.Second
 
 // ec2PlatformInventory is EC2's read-only report of an account's key pairs,
@@ -39,8 +39,17 @@ func newEC2PlatformInventoryClient(nc *nats.Conn) *ec2PlatformInventoryClient {
 // responder, which is also how a daemon predating the subject answers, a
 // timeout, an error reply or a malformed reply each make every kind incomplete.
 func (c *ec2PlatformInventoryClient) PlatformInventory(ctx context.Context, accountID, correlationRef string) ec2v1.PlatformInventoryReply {
+	// The NATS request takes a timeout, not a context, so the caller's deadline
+	// shortens it here; a context already done is a timeout without a request.
+	timeout := c.timeout
+	if deadline, ok := ctx.Deadline(); ok {
+		timeout = min(timeout, time.Until(deadline))
+	}
+	if ctx.Err() != nil || timeout <= 0 {
+		return ec2v1.IncompletePlatformInventory(accountID, correlationRef, ec2v1.PlatformReasonTimeout)
+	}
 	reply, err := natsmsg.NATSRequest[ec2v1.PlatformInventoryReply](ctx, c.nc, ec2v1.PlatformInventorySubject,
-		&ec2v1.PlatformInventoryRequest{AccountID: accountID, CorrelationRef: correlationRef}, c.timeout, accountID)
+		&ec2v1.PlatformInventoryRequest{AccountID: accountID, CorrelationRef: correlationRef}, timeout, accountID)
 	if err != nil {
 		return ec2v1.IncompletePlatformInventory(accountID, correlationRef, inventoryFailureReason(err))
 	}

spinifex/accountteardown/ec2_platform_inventory_test.go

@@ -6,11 +6,14 @@ package accountteardown
 import (
 	"context"
 	"encoding/json"
+	"strings"
+	"sync/atomic"
 	"testing"
 	"time"
 
 	"github.com/aws/aws-sdk-go/aws"
 	"github.com/aws/aws-sdk-go/service/ec2"
+	"github.com/aws/aws-sdk-go/service/s3"
 	ec2v1 "github.com/mulgadc/spinifex/contracts/ec2/v1"
 	"github.com/mulgadc/spinifex/internal/testkit"
 	"github.com/mulgadc/spinifex/spinifex/domains/ec2/accountinventory"
@@ -67,9 +70,13 @@ func TestEC2PlatformInventoryClientReturnsEC2sReport(t *testing.T) {
 		KeyName: aws.String("deploy"), PublicKeyMaterial: []byte("ssh-ed25519 AAAAC3NzaC1lZDI1NTE5AAAAIOMqqnkVzrm0SdG6UOoqKLsabgH5C9okWi0dh2l9GKJl"),
 	}, testAccountID)
 	require.NoError(t, err)
-	o.store.failDelete = func(key string) bool { return key == "keys/"+testAccountID+"/deploy" }
 	_, err = o.keys.DeleteKeyPair(ctx, &ec2.DeleteKeyPairInput{KeyName: aws.String("deploy")}, testAccountID)
 	require.NoError(t, err)
+	// A public key without metadata, as an older release's failed delete left.
+	_, err = o.store.PutObject(ctx, &s3.PutObjectInput{
+		Bucket: aws.String(platformBucket), Key: aws.String("keys/" + testAccountID + "/deploy"), Body: strings.NewReader("ssh-ed25519 AAAA"),
+	})
+	require.NoError(t, err)
 
 	reply := newEC2PlatformInventoryClient(o.nc).PlatformInventory(ctx, testAccountID, "req-1")
 
@@ -153,19 +160,10 @@ func TestEC2PlatformInventoryClientDistrustsAMalformedReply(t *testing.T) {
 	}, reply.Kinds)
 }
 
-// failingInventory fails the test if anything consults it.
-type failingInventory struct {
-	t *testing.T
-}
-
-func (f failingInventory) PlatformInventory(context.Context, string, string) ec2v1.PlatformInventoryReply {
-	f.t.Error("an engine path called the EC2 platform inventory")
-	return ec2v1.PlatformInventoryReply{}
-}
-
-// The cluster wiring binds the NATS adapter, and no engine path consults it:
-// precheck, inventory, dry run and a full teardown all leave it uncalled.
-func TestNoEnginePathCallsTheEC2PlatformInventory(t *testing.T) {
+// The cluster wiring binds the NATS adapter, and only the explicit dry run
+// consults it: precheck, the inventory the CLI prints before a real teardown,
+// a Teardown request marked dry run and a full teardown all leave it uncalled.
+func TestOnlyTheExplicitDryRunCallsTheEC2PlatformInventory(t *testing.T) {
 	_, nc, _ := testutil.StartTestJetStream(t)
 	masterKey, err := handlers_iam.GenerateMasterKey()
 	require.NoError(t, err)
@@ -175,9 +173,10 @@ func TestNoEnginePathCallsTheEC2PlatformInventory(t *testing.T) {
 	require.NoError(t, err)
 	assert.IsType(t, &ec2PlatformInventoryClient{}, cluster.ec2Inventory)
 
+	inventory := &fakeInventory{reply: ec2v1.IncompletePlatformInventory(testAccountID, "req-1", ec2v1.PlatformReasonNoResponder)}
 	reaper := &fakeReaper{kind: "key-pair", stage: StagePlatform, resources: []Resource{{Kind: "key-pair", ID: "deploy"}}}
 	engine := quietEngine(newFakeAccounts(), reaper)
-	engine.ec2Inventory = failingInventory{t: t}
+	engine.ec2Inventory = inventory
 
 	require.NoError(t, engine.Precheck(testAccountID, testAccountName))
 	_, err = engine.Inventory(testCtx(t), testAccountID)
@@ -186,7 +185,49 @@ func TestNoEnginePathCallsTheEC2PlatformInventory(t *testing.T) {
 	dryRun.DryRun = true
 	_, err = engine.Teardown(testCtx(t), dryRun)
 	require.NoError(t, err)
+	assert.Empty(t, inventory.calls())
+
+	_, err = engine.DryRun(testCtx(t), testAccountID, "req-1")
+	require.NoError(t, err)
+	assert.Equal(t, []string{"req-1"}, inventory.calls())
+
 	result, err := engine.Teardown(testCtx(t), request())
 	require.NoError(t, err)
 	assert.True(t, result.AccountDeleted)
+	assert.Equal(t, []string{"req-1"}, inventory.calls(), "a real teardown never consults the inventory")
+}
+
+// The shadow matches legacy listings to inventory kinds by the reapers' kind
+// names, so the registered platform reapers must use the contract's names.
+func TestPlatformReaperKindsAreTheInventoryKinds(t *testing.T) {
+	var kinds []string
+	for _, reaper := range EC2Reapers(nil, 1, nil) {
+		if reaper.Stage() == StagePlatform {
+			kinds = append(kinds, reaper.Kind())
+		}
+	}
+	for _, kind := range ec2ShadowKinds {
+		assert.Contains(t, kinds, kind.kind)
+	}
+}
+
+// A caller deadline shorter than the adapter's own timeout ends the request
+// at that deadline, and a context already done sends nothing.
+func TestEC2PlatformInventoryClientHonoursTheCallersDeadline(t *testing.T) {
+	_, nc, _ := testutil.StartTestJetStream(t)
+	var requests atomic.Int32
+	serveInventory(t, nc, func() []byte { requests.Add(1); return nil })
+	client := newEC2PlatformInventoryClient(nc)
+
+	ctx, cancel := context.WithTimeout(t.Context(), 100*time.Millisecond)
+	defer cancel()
+	started := time.Now()
+	assertAllIncomplete(t, client.PlatformInventory(ctx, testAccountID, ""), ec2v1.PlatformReasonTimeout)
+	assert.Less(t, time.Since(started), 5*time.Second)
+	assert.Equal(t, int32(1), requests.Load())
+
+	done, cancelDone := context.WithCancel(t.Context())
+	cancelDone()
+	assertAllIncomplete(t, client.PlatformInventory(done, testAccountID, ""), ec2v1.PlatformReasonTimeout)
+	assert.Equal(t, int32(1), requests.Load())
 }

spinifex/accountteardown/ec2_platform_shadow.go

@@ -0,0 +1,286 @@
+package accountteardown
+
+import (
+	"context"
+	"log/slog"
+	"regexp"
+	"slices"
+	"time"
+
+	ec2v1 "github.com/mulgadc/spinifex/contracts/ec2/v1"
+	otelsetup "github.com/mulgadc/spinifex/spinifex/foundation/telemetry"
+)
+
+// ec2PlatformShadowBudget caps how long a dry run waits for EC2's platform
+// inventory after its own listings, never past the caller's deadline. A later
+// answer is scan-incomplete and the dry run returns exactly as it would without.
+const ec2PlatformShadowBudget = 10 * time.Second
+
+// ec2PlatformShadowMessage is the one operator audit record a dry run writes.
+const ec2PlatformShadowMessage = "Account teardown dry-run EC2 platform shadow comparison"
+
+// Shadow outcomes, per kind and for the whole record. Only agree means the
+// legacy listing and the inventory matched and nothing hidden was reported.
+const (
+	shadowAgree          = "agree"
+	shadowDiffer         = "differ"
+	shadowScanIncomplete = "scan-incomplete"
+)
+
+// Bases a kind's visible resources are compared on.
+const (
+	shadowBasisResourceID = "resource-id"
+	shadowBasisCount      = "count"
+)
+
+// shadowKind is one compared kind. The key-pair and placement-group reapers
+// list by name, which the inventory never carries, so those compare by count.
+type shadowKind struct {
+	kind   string
+	field  string
+	byName bool
+}
+
+var ec2ShadowKinds = []shadowKind{
+	{kind: ec2v1.PlatformKindKeyPair, field: "keyPair", byName: true},
+	{kind: ec2v1.PlatformKindPlacementGroup, field: "placementGroup", byName: true},
+	{kind: ec2v1.PlatformKindLaunchTemplate, field: "launchTemplate"},
+}
+
+// loggableResourceID admits only the AWS ID shape of the three kinds, so a
+// reply that put stored content where an ID belongs never reaches the log.
+var loggableResourceID = regexp.MustCompile(`^(key|pg|lt)-[0-9a-f]{1,32}$`)
+
+var knownShadowReasons = []string{
+	ec2v1.PlatformReasonStoreListFailed, ec2v1.PlatformReasonRecordReadFailed,
+	ec2v1.PlatformReasonOwnerRecordsUnresolved, ec2v1.PlatformReasonNoResponder,
+	ec2v1.PlatformReasonTimeout, ec2v1.PlatformReasonTransportFailed,
+	ec2v1.PlatformReasonErrorReply, ec2v1.PlatformReasonInvalidReply,
+	ec2v1.PlatformReasonKindNotReported,
+}
+
+var knownShadowClasses = []string{
+	ec2v1.PlatformClassPublicKeyWithoutMetadata, ec2v1.PlatformClassPublicKeyUnresolved,
+	ec2v1.PlatformClassUnreadableRecord, ec2v1.PlatformClassVersionBodyWithoutHeader,
+	ec2v1.PlatformClassNameIndexWithoutHeader, ec2v1.PlatformClassUnrecognisedTemplateEntry,
+}
+
+// shadowUnknownClass stands in for a count class this consumer does not know.
+const shadowUnknownClass = "unrecognised-class"
+
+func (e *Engine) ec2ShadowBudget() time.Duration {
+	if e.ec2PlatformShadowBudget > 0 {
+		return e.ec2PlatformShadowBudget
+	}
+	return ec2PlatformShadowBudget
+}
+
+// shadowEC2Platform compares EC2's platform inventory with the legacy listings
+// in result and logs the comparison. It reads result and never changes it.
+func (e *Engine) shadowEC2Platform(ctx context.Context, result *Result, correlationRef string) {
+	start := time.Now()
+	reply := e.awaitEC2PlatformInventory(ctx, result.AccountID, correlationRef)
+
+	reports := make(map[string]ec2v1.PlatformKindReport, len(reply.Kinds))
+	for _, report := range reply.Kinds {
+		reports[report.Kind] = report
+	}
+
+	outcome := shadowAgree
+	attrs := []slog.Attr{
+		slog.String("accountID", result.AccountID),
+		slog.String("correlationRef", correlationRef),
+	}
+	for _, kind := range ec2ShadowKinds {
+		attr, kindOutcome := compareShadowKind(kind, legacyIDs(result, kind.kind), reports[kind.kind], reply)
+		attrs = append(attrs, attr)
+		outcome = worseShadowOutcome(outcome, kindOutcome)
+	}
+	tagAttr, tagOutcome := shadowTagIndex(reports[ec2v1.PlatformKindTagIndex], reply)
+	attrs = append(attrs, tagAttr)
+	outcome = worseShadowOutcome(outcome, tagOutcome)
+
+	attrs = append(attrs, slog.String("outcome", outcome), slog.Int64("duration_ms", otelsetup.Millis(time.Since(start))))
+	e.logger().LogAttrs(ctx, slog.LevelInfo, ec2PlatformShadowMessage, attrs...)
+}
+
+// awaitEC2PlatformInventory bounds the inventory call by the shadow budget even
+// if the implementation ignores its context, and distrusts the reply's shape.
+func (e *Engine) awaitEC2PlatformInventory(ctx context.Context, accountID, correlationRef string) ec2v1.PlatformInventoryReply {
+	ctx, cancel := context.WithTimeout(ctx, e.ec2ShadowBudget())
+	defer cancel()
+
+	replies := make(chan ec2v1.PlatformInventoryReply, 1)
+	go func() { replies <- e.ec2Inventory.PlatformInventory(ctx, accountID, correlationRef) }()
+	select {
+	case reply := <-replies:
+		return normalisePlatformInventory(reply, accountID, correlationRef)
+	case <-ctx.Done():
+		return ec2v1.IncompletePlatformInventory(accountID, correlationRef, ec2v1.PlatformReasonTimeout)
+	}
+}
+
+func legacyIDs(result *Result, kind string) []string {
+	var ids []string
+	for _, stage := range result.Stages {
+		for _, resource := range stage.Deleted {
+			if resource.Kind == kind {
+				ids = append(ids, resource.ID)
+			}
+		}
+	}
+	return ids
+}
+
+// compareShadowKind compares one kind's legacy listing with the inventory's
+// live and dependency-held items, and reports its hidden records as counts.
+func compareShadowKind(kind shadowKind, legacy []string, report ec2v1.PlatformKindReport, reply ec2v1.PlatformInventoryReply) (slog.Attr, string) {
+	var visible, held []string
+	hidden := map[string]int{}
+	hiddenTotal := 0
+	for _, item := range reply.Items {
+		if item.Kind != kind.kind {
+			continue
+		}
+		switch item.Disposition {
+		case ec2v1.PlatformDispositionLive:
+			visible = append(visible, item.ResourceID)
+		case ec2v1.PlatformDispositionDependencyHeld:
+			visible = append(visible, item.ResourceID)
+			held = append(held, item.ResourceID)
+		default:
+			hidden[shadowUnknownClass]++
+			hiddenTotal++
+		}
+	}
+	for _, count := range reply.Counts {
+		if count.Kind != kind.kind || count.Count <= 0 {
+			continue
+		}
+		class := count.Class
+		if !slices.Contains(knownShadowClasses, class) {
+			class = shadowUnknownClass
+		}
+		hidden[class] += count.Count
+		hiddenTotal += count.Count
+	}
+
+	basis := shadowBasisResourceID
+	var onlyLegacy, onlyInventory []string
+	agree := len(legacy) == len(visible)
+	if kind.byName {
+		basis = shadowBasisCount
+	} else {
+		onlyLegacy, onlyInventory = setDifference(legacy, visible), setDifference(visible, legacy)
+		agree = len(onlyLegacy) == 0 && len(onlyInventory) == 0
+	}
+
+	comparison := shadowDiffer
+	if agree {
+		comparison = shadowAgree
+	}
+	outcome := comparison
+	if hiddenTotal > 0 {
+		outcome = shadowDiffer
+	}
+	if report.Status != ec2v1.PlatformStatusObservedClean && report.Status != ec2v1.PlatformStatusRecordsFound {
+		comparison, outcome = shadowScanIncomplete, shadowScanIncomplete
+	}
+
+	loggableOnlyLegacy, droppedLegacy := loggableIDs(onlyLegacy)
+	loggableOnlyInventory, droppedInventory := loggableIDs(onlyInventory)
+	loggableHeld, droppedHeld := loggableIDs(held)
+	return slog.Group(kind.field,
+		slog.String("status", report.Status),
+		slog.Any("incompleteReasons", fixedReasons(report.IncompleteReasons)),
+		slog.Time("scanStartedAt", report.ScanStartedAt),
+		slog.Time("scanEndedAt", report.ScanEndedAt),
+		slog.String("comparison", comparison),
+		slog.String("matchBasis", basis),
+		slog.Int("legacyCount", len(legacy)),
+		slog.Int("inventoryVisibleCount", len(visible)),
+		slog.Any("onlyInLegacy", loggableOnlyLegacy),
+		slog.Any("onlyInInventory", loggableOnlyInventory),
+		slog.Any("dependencyHeld", loggableHeld),
+		slog.Int("hiddenRecordCount", hiddenTotal),
+		slog.Any("hiddenRecords", hidden),
+		slog.Int("unloggableIDCount", droppedLegacy+droppedInventory+droppedHeld),
+	), outcome
+}
+
+// shadowTagIndex reports tag-index entries whose resource has no owner record,
+// which no legacy reaper lists.
+func shadowTagIndex(report ec2v1.PlatformKindReport, reply ec2v1.PlatformInventoryReply) (slog.Attr, string) {
+	var residue []string
+	for _, item := range reply.Items {
+		if item.Kind == ec2v1.PlatformKindTagIndex {
+			residue = append(residue, item.ResourceID)
+		}
+	}
+	outcome := shadowAgree
+	if len(residue) > 0 {
+		outcome = shadowDiffer
+	}
+	if report.Status != ec2v1.PlatformStatusObservedClean && report.Status != ec2v1.PlatformStatusRecordsFound {
+		outcome = shadowScanIncomplete
+	}
+	loggable, dropped := loggableIDs(residue)
+	return slog.Group("tagIndex",
+		slog.String("status", report.Status),
+		slog.Any("incompleteReasons", fixedReasons(report.IncompleteReasons)),
+		slog.Time("scanStartedAt", report.ScanStartedAt),
+		slog.Time("scanEndedAt", report.ScanEndedAt),
+		slog.Int("residueCount", len(residue)),
+		slog.Any("residue", loggable),
+		slog.Int("unloggableIDCount", dropped),
+	), outcome
+}
+
+func worseShadowOutcome(a, b string) string {
+	rank := map[string]int{shadowAgree: 0, shadowDiffer: 1, shadowScanIncomplete: 2}
+	if rank[b] > rank[a] {
+		return b
+	}
+	return a
+}
+
+// setDifference returns the sorted members of a absent from b.
+func setDifference(a, b []string) []string {
+	out := []string{}
+	for _, id := range a {
+		if !slices.Contains(b, id) && !slices.Contains(out, id) {
+			out = append(out, id)
+		}
+	}
+	slices.Sort(out)
+	return out
+}
+
+func loggableIDs(ids []string) ([]string, int) {
+	out := []string{}
+	dropped := 0
+	for _, id := range ids {
+		if loggableResourceID.MatchString(id) {
+			out = append(out, id)
+			continue
+		}
+		dropped++
+	}
+	slices.Sort(out)
+	return out, dropped
+}
+
+// fixedReasons keeps only the contract's reason categories; anything else a
+// reply carried is reported as an invalid reply rather than logged.
+func fixedReasons(reasons []string) []string {
+	out := []string{}
+	for _, reason := range reasons {
+		if !slices.Contains(knownShadowReasons, reason) {
+			reason = ec2v1.PlatformReasonInvalidReply
+		}
+		if !slices.Contains(out, reason) {
+			out = append(out, reason)
+		}
+	}
+	return out
+}

spinifex/accountteardown/ec2_platform_shadow_test.go

@@ -0,0 +1,513 @@
+package accountteardown
+
+//test:in-package — the shadow comparison, the inventory binding it reads and
+// its budget are unexported.
+
+import (
+	"bytes"
+	"context"
+	"encoding/base64"
+	"encoding/json"
+	"errors"
+	"log/slog"
+	"strings"
+	"sync"
+	"testing"
+	"time"
+
+	"github.com/aws/aws-sdk-go/aws"
+	"github.com/aws/aws-sdk-go/service/ec2"
+	"github.com/aws/aws-sdk-go/service/s3"
+	ec2v1 "github.com/mulgadc/spinifex/contracts/ec2/v1"
+	"github.com/mulgadc/spinifex/internal/testkit"
+	"github.com/mulgadc/spinifex/spinifex/domains/ec2/accountinventory"
+	natsmsg "github.com/mulgadc/spinifex/spinifex/foundation/messaging/nats"
+	"github.com/nats-io/nats.go"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+var shadowScanTime = time.Date(2026, 10, 16, 9, 0, 0, 0, time.UTC)
+
+// fakeInventory answers with reply, or never when blocked, and records the
+// correlation reference of every call.
+type fakeInventory struct {
+	reply   ec2v1.PlatformInventoryReply
+	blocked chan struct{}
+
+	mu   sync.Mutex
+	refs []string
+}
+
+func (f *fakeInventory) PlatformInventory(_ context.Context, _, correlationRef string) ec2v1.PlatformInventoryReply {
+	f.mu.Lock()
+	f.refs = append(f.refs, correlationRef)
+	f.mu.Unlock()
+	if f.blocked != nil {
+		<-f.blocked
+	}
+	return f.reply
+}
+
+func (f *fakeInventory) calls() []string {
+	f.mu.Lock()
+	defer f.mu.Unlock()
+	return append([]string(nil), f.refs...)
+}
+
+// countingReaper counts listings, so a dry run can be shown to list once.
+type countingReaper struct {
+	Reaper
+
+	lists int
+}
+
+func (r *countingReaper) List(ctx context.Context, accountID string) ([]Resource, error) {
+	r.lists++
+	return r.Reaper.List(ctx, accountID)
+}
+
+func platformKinds(status string) []ec2v1.PlatformKindReport {
+	var kinds []ec2v1.PlatformKindReport
+	for _, kind := range ec2v1.PlatformKinds() {
+		kinds = append(kinds, ec2v1.PlatformKindReport{
+			Kind: kind, Status: status, ScanStartedAt: shadowScanTime, ScanEndedAt: shadowScanTime.Add(time.Second),
+		})
+	}
+	return kinds
+}
+
+func setKind(kinds []ec2v1.PlatformKindReport, kind, status string, reasons ...string) {
+	for i := range kinds {
+		if kinds[i].Kind == kind {
+			kinds[i].Status, kinds[i].IncompleteReasons = status, reasons
+		}
+	}
+}
+
+func item(kind, id, disposition string) ec2v1.PlatformItem {
+	return ec2v1.PlatformItem{ResourceID: id, Kind: kind, Disposition: disposition}
+}
+
+// legacyPlatform is a key pair, a placement group and a launch template as the
+// legacy reapers list them: the first two by name, the third by ID.
+func legacyPlatform() []Reaper {
+	return []Reaper{
+		&fakeReaper{kind: "instance", stage: StageCompute, resources: []Resource{{Kind: "instance", ID: "i-0123456789abcdef0"}}},
+		&fakeReaper{kind: "key-pair", stage: StagePlatform, resources: []Resource{{Kind: "key-pair", ID: "deploy"}}},
+		&fakeReaper{kind: "placement-group", stage: StagePlatform, resources: []Resource{{Kind: "placement-group", ID: "spread"}}},
+		&fakeReaper{kind: "launch-template", stage: StagePlatform, resources: []Resource{{Kind: "launch-template", ID: "lt-0aaaaaaaaaaaaaaaa"}}},
+	}
+}
+
+// matchingInventory reports exactly what legacyPlatform lists.
+func matchingInventory() ec2v1.PlatformInventoryReply {
+	return ec2v1.PlatformInventoryReply{
+		AccountID: testAccountID,
+		Kinds:     platformKinds(ec2v1.PlatformStatusRecordsFound),
+		Items: []ec2v1.PlatformItem{
+			item(ec2v1.PlatformKindKeyPair, "key-0aaaaaaaaaaaaaaaa", ec2v1.PlatformDispositionLive),
+			item(ec2v1.PlatformKindPlacementGroup, "pg-0aaaaaaaaaaaaaaaa", ec2v1.PlatformDispositionLive),
+			item(ec2v1.PlatformKindLaunchTemplate, "lt-0aaaaaaaaaaaaaaaa", ec2v1.PlatformDispositionLive),
+		},
+	}
+}
+
+// shadowEngine is a quiet engine whose log is captured as JSON in logs.
+func shadowEngine(inventory ec2PlatformInventory, reapers ...Reaper) (*Engine, *bytes.Buffer) {
+	engine := quietEngine(newFakeAccounts(), reapers...)
+	logs := &bytes.Buffer{}
+	engine.Logger = slog.New(slog.NewJSONHandler(logs, nil))
+	engine.ec2Inventory = inventory
+	return engine, logs
+}
+
+// shadowRecords returns every shadow comparison record in logs.
+func shadowRecords(t *testing.T, logs *bytes.Buffer) []map[string]any {
+	t.Helper()
+	var records []map[string]any
+	for line := range strings.SplitSeq(strings.TrimSpace(logs.String()), "\n") {
+		if line == "" {
+			continue
+		}
+		var record map[string]any
+		require.NoError(t, json.Unmarshal([]byte(line), &record))
+		if record["msg"] == ec2PlatformShadowMessage {
+			records = append(records, record)
+		}
+	}
+	return records
+}
+
+func shadowRecord(t *testing.T, logs *bytes.Buffer) map[string]any {
+	t.Helper()
+	records := shadowRecords(t, logs)
+	require.Len(t, records, 1, "one shadow record per dry run")
+	return records[0]
+}
+
+func group(t *testing.T, record map[string]any, name string) map[string]any {
+	t.Helper()
+	g, ok := record[name].(map[string]any)
+	require.True(t, ok, "record has a %s group", name)
+	return g
+}
+
+func strs(values ...string) []any {
+	out := []any{}
+	for _, v := range values {
+		out = append(out, v)
+	}
+	return out
+}
+
+func resultJSON(t *testing.T, result *Result) string {
+	t.Helper()
+	data, err := json.Marshal(result)
+	require.NoError(t, err)
+	return string(data)
+}
+
+// noShadowResult is what the same dry run returns with no inventory bound.
+func noShadowResult(t *testing.T, reapers ...Reaper) string {
+	t.Helper()
+	result, err := quietEngine(newFakeAccounts(), reapers...).Inventory(testCtx(t), testAccountID)
+	require.NoError(t, err)
+	return resultJSON(t, result)
+}
+
+// Visible resources that match on both sides agree. The legacy key-pair and
+// placement-group reapers list names, so those kinds compare by count.
+func TestDryRunShadowAgreesWhenBothSidesListTheSameResources(t *testing.T) {
+	reapers := legacyPlatform()
+	counted := make([]Reaper, 0, len(reapers))
+	for _, reaper := range reapers {
+		counted = append(counted, &countingReaper{Reaper: reaper})
+	}
+	inventory := &fakeInventory{reply: matchingInventory()}
+	engine, logs := shadowEngine(inventory, counted...)
+
+	result, err := engine.DryRun(testCtx(t), testAccountID, "req-agree")
+	require.NoError(t, err)
+
+	assert.Equal(t, noShadowResult(t, legacyPlatform()...), resultJSON(t, result))
+	assert.Equal(t, []string{"req-agree"}, inventory.calls())
+	for _, reaper := range counted {
+		assert.Equal(t, 1, reaper.(*countingReaper).lists, "%s listed once: the shadow reuses the dry run's listing", reaper.Kind())
+	}
+
+	record := shadowRecord(t, logs)
+	assert.Equal(t, "INFO", record["level"])
+	assert.Equal(t, testAccountID, record["accountID"])
+	assert.Equal(t, "req-agree", record["correlationRef"])
+	assert.Equal(t, shadowAgree, record["outcome"])
+	assert.Contains(t, record, "duration_ms")
+
+	keyPair := group(t, record, "keyPair")
+	assert.Equal(t, ec2v1.PlatformStatusRecordsFound, keyPair["status"])
+	assert.Equal(t, shadowAgree, keyPair["comparison"])
+	assert.Equal(t, shadowBasisCount, keyPair["matchBasis"])
+	assert.InDelta(t, 1, keyPair["legacyCount"], 0)
+	assert.InDelta(t, 1, keyPair["inventoryVisibleCount"], 0)
+	assert.Equal(t, shadowScanTime.Format(time.RFC3339), keyPair["scanStartedAt"])
+
+	launchTemplate := group(t, record, "launchTemplate")
+	assert.Equal(t, shadowAgree, launchTemplate["comparison"])
+	assert.Equal(t, shadowBasisResourceID, launchTemplate["matchBasis"])
+	assert.Equal(t, strs(), launchTemplate["onlyInLegacy"])
+	assert.Equal(t, strs(), launchTemplate["onlyInInventory"])
+
+	tagIndex := group(t, record, "tagIndex")
+	assert.Equal(t, ec2v1.PlatformStatusRecordsFound, tagIndex["status"])
+	assert.InDelta(t, 0, tagIndex["residueCount"], 0)
+}
+
+// Visible resources one side has and the other does not are named by ID where
+// the legacy reaper lists IDs, and counted where it lists names.
+func TestDryRunShadowReportsVisibleResourcesOnlyOneSideSees(t *testing.T) {
+	reapers := legacyPlatform()
+	reapers[1].(*fakeReaper).resources = append(reapers[1].(*fakeReaper).resources, Resource{Kind: "key-pair", ID: "ci"})
+	reapers[3].(*fakeReaper).resources = []Resource{
+		{Kind: "launch-template", ID: "lt-0aaaaaaaaaaaaaaaa"}, {Kind: "launch-template", ID: "lt-0bbbbbbbbbbbbbbbb"},
+	}
+	reply := matchingInventory()
+	reply.Items = append(reply.Items, item(ec2v1.PlatformKindLaunchTemplate, "lt-0cccccccccccccccc", ec2v1.PlatformDispositionLive))
+	engine, logs := shadowEngine(&fakeInventory{reply: reply}, reapers...)
+
+	_, err := engine.DryRun(testCtx(t), testAccountID, "req-differ")
+	require.NoError(t, err)
+
+	record := shadowRecord(t, logs)
+	assert.Equal(t, shadowDiffer, record["outcome"])
+	keyPair := group(t, record, "keyPair")
+	assert.Equal(t, shadowDiffer, keyPair["comparison"])
+	assert.InDelta(t, 2, keyPair["legacyCount"], 0)
+	assert.InDelta(t, 1, keyPair["inventoryVisibleCount"], 0)
+	assert.Equal(t, strs(), keyPair["onlyInLegacy"], "legacy key-pair names are never logged")
+
+	launchTemplate := group(t, record, "launchTemplate")
+	assert.Equal(t, shadowDiffer, launchTemplate["comparison"])
+	assert.Equal(t, strs("lt-0bbbbbbbbbbbbbbbb"), launchTemplate["onlyInLegacy"])
+	assert.Equal(t, strs("lt-0cccccccccccccccc"), launchTemplate["onlyInInventory"])
+	assert.Equal(t, shadowAgree, group(t, record, "placementGroup")["comparison"])
+}
+
+// Residue the legacy reapers cannot see makes the record differ even when the
+// visible resources agree: as counts per class, and tag-index entries by ID.
+func TestDryRunShadowReportsHiddenResidueTheLegacyReapersMiss(t *testing.T) {
+	reply := matchingInventory()
+	reply.Counts = []ec2v1.PlatformRecordCount{
+		{Kind: ec2v1.PlatformKindKeyPair, Disposition: ec2v1.PlatformDispositionResidue, Class: ec2v1.PlatformClassPublicKeyWithoutMetadata, Count: 1},
+		{Kind: ec2v1.PlatformKindLaunchTemplate, Disposition: ec2v1.PlatformDispositionResidue, Class: ec2v1.PlatformClassVersionBodyWithoutHeader, Count: 2},
+		{Kind: ec2v1.PlatformKindLaunchTemplate, Disposition: ec2v1.PlatformDispositionUnreadable, Class: ec2v1.PlatformClassUnreadableRecord, Count: 1},
+	}
+	reply.Items = append(reply.Items, item(ec2v1.PlatformKindTagIndex, "pg-0dddddddddddddddd", ec2v1.PlatformDispositionResidue))
+	engine, logs := shadowEngine(&fakeInventory{reply: reply}, legacyPlatform()...)
+
+	_, err := engine.DryRun(testCtx(t), testAccountID, "req-residue")
+	require.NoError(t, err)
+
+	record := shadowRecord(t, logs)
+	assert.Equal(t, shadowDiffer, record["outcome"])
+	keyPair := group(t, record, "keyPair")
+	assert.Equal(t, shadowAgree, keyPair["comparison"], "the visible key pairs still agree")
+	assert.InDelta(t, 1, keyPair["hiddenRecordCount"], 0)
+	assert.Equal(t, map[string]any{ec2v1.PlatformClassPublicKeyWithoutMetadata: float64(1)}, keyPair["hiddenRecords"])
+	launchTemplate := group(t, record, "launchTemplate")
+	assert.InDelta(t, 3, launchTemplate["hiddenRecordCount"], 0)
+	assert.Equal(t, map[string]any{
+		ec2v1.PlatformClassVersionBodyWithoutHeader: float64(2),
+		ec2v1.PlatformClassUnreadableRecord:         float64(1),
+	}, launchTemplate["hiddenRecords"])
+	tagIndex := group(t, record, "tagIndex")
+	assert.InDelta(t, 1, tagIndex["residueCount"], 0)
+	assert.Equal(t, strs("pg-0dddddddddddddddd"), tagIndex["residue"])
+}
+
+// A placement group still holding instances is a plain listed resource to the
+// legacy dry run, which learns otherwise only when its delete fails.
+func TestDryRunShadowReportsADependencyHeldPlacementGroup(t *testing.T) {
+	reply := matchingInventory()
+	reply.Items[1].Disposition = ec2v1.PlatformDispositionDependencyHeld
+	engine, logs := shadowEngine(&fakeInventory{reply: reply}, legacyPlatform()...)
+
+	_, err := engine.DryRun(testCtx(t), testAccountID, "req-held")
+	require.NoError(t, err)
+
+	placementGroup := group(t, shadowRecord(t, logs), "placementGroup")
+	assert.Equal(t, shadowAgree, placementGroup["comparison"])
+	assert.InDelta(t, 1, placementGroup["inventoryVisibleCount"], 0)
+	assert.Equal(t, strs("pg-0aaaaaaaaaaaaaaaa"), placementGroup["dependencyHeld"])
+}
+
+// An owner store that could not be read is scan-incomplete, never agreement,
+// even where the IDs it did report match.
+func TestDryRunShadowNeverReportsAnIncompleteScanAsAgreement(t *testing.T) {
+	reply := matchingInventory()
+	setKind(reply.Kinds, ec2v1.PlatformKindLaunchTemplate, ec2v1.PlatformStatusScanIncomplete, ec2v1.PlatformReasonRecordReadFailed)
+	engine, logs := shadowEngine(&fakeInventory{reply: reply}, legacyPlatform()...)
+
+	result, err := engine.DryRun(testCtx(t), testAccountID, "req-store")
+	require.NoError(t, err)
+	assert.Equal(t, noShadowResult(t, legacyPlatform()...), resultJSON(t, result))
+
+	record := shadowRecord(t, logs)
+	assert.Equal(t, shadowScanIncomplete, record["outcome"])
+	launchTemplate := group(t, record, "launchTemplate")
+	assert.Equal(t, shadowScanIncomplete, launchTemplate["comparison"])
+	assert.Equal(t, strs(ec2v1.PlatformReasonRecordReadFailed), launchTemplate["incompleteReasons"])
+	assert.Equal(t, shadowAgree, group(t, record, "keyPair")["comparison"])
+}
+
+// However the inventory fails, the dry run returns what it returns without it
+// and the record says scan-incomplete with the fixed reason.
+func TestDryRunShadowFailureLeavesTheDryRunUnchanged(t *testing.T) {
+	want := noShadowResult(t, legacyPlatform()...)
+	for _, reason := range []string{
+		ec2v1.PlatformReasonNoResponder, ec2v1.PlatformReasonTimeout,
+		ec2v1.PlatformReasonTransportFailed, ec2v1.PlatformReasonErrorReply,
+	} {
+		t.Run(reason, func(t *testing.T) {
+			engine, logs := shadowEngine(&fakeInventory{reply: ec2v1.IncompletePlatformInventory(testAccountID, "req-fail", reason)}, legacyPlatform()...)
+			result, err := engine.DryRun(testCtx(t), testAccountID, "req-fail")
+			require.NoError(t, err)
+			assert.Equal(t, want, resultJSON(t, result))
+
+			record := shadowRecord(t, logs)
+			assert.Equal(t, shadowScanIncomplete, record["outcome"])
+			for _, name := range []string{"keyPair", "placementGroup", "launchTemplate", "tagIndex"} {
+				g := group(t, record, name)
+				assert.Equal(t, ec2v1.PlatformStatusScanIncomplete, g["status"], name)
+				assert.Equal(t, strs(reason), g["incompleteReasons"], name)
+			}
+		})
+	}
+
+	t.Run("reply for another account", func(t *testing.T) {
+		reply := matchingInventory()
+		reply.AccountID = "000000000099"
+		engine, logs := shadowEngine(&fakeInventory{reply: reply}, legacyPlatform()...)
+		result, err := engine.DryRun(testCtx(t), testAccountID, "req-fail")
+		require.NoError(t, err)
+		assert.Equal(t, want, resultJSON(t, result))
+		record := shadowRecord(t, logs)
+		assert.Equal(t, shadowScanIncomplete, record["outcome"])
+		assert.Equal(t, strs(ec2v1.PlatformReasonInvalidReply), group(t, record, "keyPair")["incompleteReasons"])
+	})
+
+	t.Run("no daemon serves the subject", func(t *testing.T) {
+		_, nc, _ := testutil.StartTestJetStream(t)
+		engine, logs := shadowEngine(newEC2PlatformInventoryClient(nc), legacyPlatform()...)
+		result, err := engine.DryRun(testCtx(t), testAccountID, "req-fail")
+		require.NoError(t, err)
+		assert.Equal(t, want, resultJSON(t, result))
+		assert.Equal(t, strs(ec2v1.PlatformReasonNoResponder), group(t, shadowRecord(t, logs), "tagIndex")["incompleteReasons"])
+	})
+}
+
+// An inventory that never answers, even one ignoring its context, holds the
+// dry run for the shadow budget at most and is logged as a timeout.
+func TestDryRunShadowIsBoundedByItsBudget(t *testing.T) {
+	inventory := &fakeInventory{reply: matchingInventory(), blocked: make(chan struct{})}
+	t.Cleanup(func() { close(inventory.blocked) })
+	engine, logs := shadowEngine(inventory, legacyPlatform()...)
+	engine.ec2PlatformShadowBudget = 20 * time.Millisecond
+
+	started := time.Now()
+	result, err := engine.DryRun(testCtx(t), testAccountID, "req-slow")
+	require.NoError(t, err)
+	assert.Less(t, time.Since(started), 5*time.Second)
+	assert.Equal(t, noShadowResult(t, legacyPlatform()...), resultJSON(t, result))
+	assert.Equal(t, strs(ec2v1.PlatformReasonTimeout), group(t, shadowRecord(t, logs), "launchTemplate")["incompleteReasons"])
+}
+
+// The caller's deadline wins over the budget, so a dry run never outlasts the
+// bound its caller already set.
+func TestDryRunShadowStopsAtTheCallersDeadline(t *testing.T) {
+	_, nc, _ := testutil.StartTestJetStream(t)
+	serveInventory(t, nc, func() []byte { return nil })
+	engine, logs := shadowEngine(newEC2PlatformInventoryClient(nc), legacyPlatform()...)
+	require.Greater(t, engine.ec2ShadowBudget(), 5*time.Second)
+
+	ctx, cancel := context.WithTimeout(t.Context(), 200*time.Millisecond)
+	defer cancel()
+	started := time.Now()
+	result, err := engine.DryRun(ctx, testAccountID, "req-deadline")
+	require.NoError(t, err)
+	assert.Less(t, time.Since(started), 5*time.Second)
+	assert.Equal(t, noShadowResult(t, legacyPlatform()...), resultJSON(t, result))
+	assert.Equal(t, strs(ec2v1.PlatformReasonTimeout), group(t, shadowRecord(t, logs), "keyPair")["incompleteReasons"])
+}
+
+// A dry run whose own listing fails returns that error as before, and nothing
+// is compared against a listing that does not exist.
+func TestDryRunWithAFailedListingDoesNotCallTheInventory(t *testing.T) {
+	reapers := legacyPlatform()
+	reapers[2].(*fakeReaper).listErr = errors.New("describe failed")
+	inventory := &fakeInventory{reply: matchingInventory()}
+	engine, logs := shadowEngine(inventory, reapers...)
+
+	_, err := engine.DryRun(testCtx(t), testAccountID, "req-list")
+	require.ErrorContains(t, err, "inventory placement-group")
+	assert.Empty(t, inventory.calls())
+	assert.Empty(t, shadowRecords(t, logs))
+}
+
+// A reply that puts anything but the contract's fixed vocabulary where a
+// reason, class or ID belongs never reaches the log; the IDs are counted.
+func TestDryRunShadowLogsOnlyTheFixedVocabulary(t *testing.T) {
+	reply := matchingInventory()
+	setKind(reply.Kinds, ec2v1.PlatformKindKeyPair, ec2v1.PlatformStatusScanIncomplete, "read keys/000000000042/secret-key-name: denied")
+	reply.Counts = []ec2v1.PlatformRecordCount{{Kind: ec2v1.PlatformKindLaunchTemplate, Disposition: ec2v1.PlatformDispositionResidue, Class: "secret-template-name", Count: 1}}
+	reply.Items = append(reply.Items,
+		item(ec2v1.PlatformKindLaunchTemplate, "lt-secret-template-name", ec2v1.PlatformDispositionLive),
+		item(ec2v1.PlatformKindTagIndex, "pg-team=secret-tag-value", ec2v1.PlatformDispositionResidue),
+	)
+	reapers := legacyPlatform()
+	reapers[1].(*fakeReaper).resources = []Resource{{Kind: "key-pair", ID: "secret-key-name"}}
+	engine, logs := shadowEngine(&fakeInventory{reply: reply}, reapers...)
+
+	_, err := engine.DryRun(testCtx(t), testAccountID, "req-disclosure")
+	require.NoError(t, err)
+
+	assert.NotContains(t, logs.String(), "secret")
+	record := shadowRecord(t, logs)
+	assert.Equal(t, strs(ec2v1.PlatformReasonInvalidReply), group(t, record, "keyPair")["incompleteReasons"])
+	launchTemplate := group(t, record, "launchTemplate")
+	assert.Equal(t, map[string]any{shadowUnknownClass: float64(1)}, launchTemplate["hiddenRecords"])
+	assert.InDelta(t, 1, launchTemplate["unloggableIDCount"], 0)
+	assert.InDelta(t, 1, group(t, record, "tagIndex")["unloggableIDCount"], 0)
+}
+
+// End to end over the real owner stores and the NATS adapter: the record names
+// the residue each reaper misses and nothing a customer named or stored.
+func TestDryRunShadowOverTheOwnerStores(t *testing.T) {
+	o := newPlatformOwners(t)
+	ctx := testCtx(t)
+	inv := accountinventory.New(o.keys, o.groups, o.templates, o.tags)
+	sub, err := o.nc.QueueSubscribe(ec2v1.PlatformInventorySubject, "spinifex-workers", func(msg *nats.Msg) {
+		natsmsg.ServeNATSRequestCtx(msg, inv.PlatformInventory)
+	})
+	require.NoError(t, err)
+	t.Cleanup(func() { _ = sub.Unsubscribe() })
+
+	material := []byte("ssh-ed25519 AAAAC3NzaC1lZDI1NTE5AAAAIOMqqnkVzrm0SdG6UOoqKLsabgH5C9okWi0dh2l9GKJl")
+	for _, name := range []string{"secret-key-name", "secret-orphan-key"} {
+		_, err = o.keys.ImportKeyPair(ctx, &ec2.ImportKeyPairInput{KeyName: aws.String(name), PublicKeyMaterial: material}, testAccountID)
+		require.NoError(t, err)
+	}
+	_, err = o.keys.DeleteKeyPair(ctx, &ec2.DeleteKeyPairInput{KeyName: aws.String("secret-orphan-key")}, testAccountID)
+	require.NoError(t, err)
+	// Residue an older release's failed deletes left behind.
+	_, err = o.store.PutObject(ctx, &s3.PutObjectInput{
+		Bucket: aws.String(platformBucket), Key: aws.String("keys/" + testAccountID + "/secret-orphan-key"), Body: bytes.NewReader(material),
+	})
+	require.NoError(t, err)
+
+	_, err = o.groups.CreatePlacementGroup(ctx, &ec2.CreatePlacementGroupInput{
+		GroupName: aws.String("secret-group-name"), Strategy: aws.String(ec2.PlacementStrategySpread),
+	}, testAccountID)
+	require.NoError(t, err)
+	doomed, err := o.groups.CreatePlacementGroup(ctx, &ec2.CreatePlacementGroupInput{
+		GroupName: aws.String("secret-doomed-group"), Strategy: aws.String(ec2.PlacementStrategySpread),
+	}, testAccountID)
+	require.NoError(t, err)
+	lt, err := o.templates.CreateLaunchTemplate(ctx, &ec2.CreateLaunchTemplateInput{
+		LaunchTemplateName: aws.String("secret-template-name"),
+		LaunchTemplateData: &ec2.RequestLaunchTemplateData{UserData: aws.String(base64.StdEncoding.EncodeToString([]byte("secret-user-data")))},
+	}, testAccountID)
+	require.NoError(t, err)
+	doomedID, ltID := aws.StringValue(doomed.PlacementGroup.GroupId), aws.StringValue(lt.LaunchTemplate.LaunchTemplateId)
+	_, err = o.tags.CreateTags(ctx, &ec2.CreateTagsInput{
+		Resources: []*string{aws.String(doomedID), aws.String(ltID)},
+		Tags:      []*ec2.Tag{{Key: aws.String("secret-tag-key"), Value: aws.String("secret-tag-value")}},
+	}, testAccountID)
+	require.NoError(t, err)
+	_, err = o.groups.DeletePlacementGroup(ctx, &ec2.DeletePlacementGroupInput{GroupName: aws.String("secret-doomed-group")}, testAccountID)
+	require.NoError(t, err)
+	_, err = o.tags.CreateTags(ctx, &ec2.CreateTagsInput{
+		Resources: []*string{aws.String(doomedID)},
+		Tags:      []*ec2.Tag{{Key: aws.String("secret-tag-key"), Value: aws.String("secret-tag-value")}},
+	}, testAccountID)
+	require.NoError(t, err)
+
+	engine, logs := shadowEngine(newEC2PlatformInventoryClient(o.nc), platformReapers(o.nc)...)
+	result, err := engine.DryRun(ctx, testAccountID, "req-owners")
+	require.NoError(t, err)
+	plain, err := quietEngine(newFakeAccounts(), platformReapers(o.nc)...).Inventory(ctx, testAccountID)
+	require.NoError(t, err)
+	assert.Equal(t, resultJSON(t, plain), resultJSON(t, result))
+
+	assert.NotContains(t, logs.String(), "secret")
+	assert.NotContains(t, logs.String(), base64.StdEncoding.EncodeToString([]byte("secret-user-data")))
+	record := shadowRecord(t, logs)
+	assert.Equal(t, shadowDiffer, record["outcome"])
+	keyPair := group(t, record, "keyPair")
+	assert.Equal(t, shadowAgree, keyPair["comparison"])
+	assert.Equal(t, map[string]any{ec2v1.PlatformClassPublicKeyWithoutMetadata: float64(1)}, keyPair["hiddenRecords"])
+	assert.Equal(t, shadowAgree, group(t, record, "placementGroup")["comparison"])
+	launchTemplate := group(t, record, "launchTemplate")
+	assert.Equal(t, shadowAgree, launchTemplate["comparison"])
+	assert.Equal(t, shadowBasisResourceID, launchTemplate["matchBasis"])
+	assert.Equal(t, strs(doomedID), group(t, record, "tagIndex")["residue"])
+}

spinifex/accountteardown/engine.go

@@ -29,9 +29,12 @@ type Engine struct {
 	// and this is the only view into it while it runs.
 	OnStage func(ctx context.Context, stage StageResult)
 
-	// ec2Inventory is EC2's read-only platform inventory. Bound by the cluster
-	// wiring but consulted by no teardown or inventory path yet.
+	// ec2Inventory is EC2's read-only platform inventory. Only DryRun consults
+	// it, and only to log a comparison; no teardown or Inventory call does.
 	ec2Inventory ec2PlatformInventory
+
+	// ec2PlatformShadowBudget overrides ec2PlatformShadowBudget when positive.
+	ec2PlatformShadowBudget time.Duration
 }
 
 // NewEngine builds an engine with the default timeouts.
@@ -87,6 +90,20 @@ func (e *Engine) Inventory(ctx context.Context, accountID string) (*Result, erro
 	return result, nil
 }
 
+// DryRun is the explicit operator dry run. It returns exactly what Inventory
+// returns; beside it, EC2's platform inventory is compared with the legacy
+// listings and only logged, under correlationRef, which authorises nothing.
+func (e *Engine) DryRun(ctx context.Context, accountID, correlationRef string) (*Result, error) {
+	result, err := e.Inventory(ctx, accountID)
+	if err != nil {
+		return nil, err
+	}
+	if e.ec2Inventory != nil {
+		e.shadowEC2Platform(ctx, result, correlationRef)
+	}
+	return result, nil
+}
+
 // Teardown empties the account and then removes it.
 //
 // The account is marked TERMINATING before anything is deleted. Without that

spinifex/accountteardown/reapers_ec2_platform_current_behaviour_test.go

@@ -34,27 +34,11 @@ import (
 
 const platformBucket = "test-bucket"
 
-// deleteFailingStore fails object deletes its predicate selects.
-type deleteFailingStore struct {
-	*objectstore.MemoryObjectStore
-
-	failDelete func(key string) bool
-}
-
-var _ objectstore.ObjectStore = (*deleteFailingStore)(nil)
-
-func (s *deleteFailingStore) DeleteObject(ctx context.Context, in *s3.DeleteObjectInput) (*s3.DeleteObjectOutput, error) {
-	if s.failDelete != nil && s.failDelete(aws.StringValue(in.Key)) {
-		return nil, errors.New("injected store failure")
-	}
-	return s.MemoryObjectStore.DeleteObject(ctx, in)
-}
-
 // platformOwners runs the real key, placement-group, launch-template and tag
 // services behind the subjects the platform reapers call.
 type platformOwners struct {
 	nc        *nats.Conn
-	store     *deleteFailingStore
+	store     *objectstore.MemoryObjectStore
 	keys      *ec2key.KeyServiceImpl
 	groups    *ec2placementgroup.PlacementGroupServiceImpl
 	templates *ec2launchtemplate.LaunchTemplateServiceImpl
@@ -81,7 +65,7 @@ func newPlatformOwners(t *testing.T) *platformOwners {
 	t.Helper()
 	_, nc, js := testutil.StartTestJetStream(t)
 	ctx := t.Context()
-	o := &platformOwners{nc: nc, store: &deleteFailingStore{MemoryObjectStore: objectstore.NewMemoryObjectStore()}}
+	o := &platformOwners{nc: nc, store: objectstore.NewMemoryObjectStore()}
 
 	var err error
 	o.tagsKV, err = ec2tags.GetOrCreateTagsBucket(ctx, js)
@@ -92,8 +76,10 @@ func newPlatformOwners(t *testing.T) *platformOwners {
 	o.keys.SetCentralTagStore(o.tags)
 	o.groups, err = ec2placementgroup.NewPlacementGroupServiceImplWithNATS(ctx, nil, nc)
 	require.NoError(t, err)
+	o.groups.SetCentralTagStore(o.tags)
 	o.templates, err = ec2launchtemplate.NewLaunchTemplateServiceImplWithNATS(ctx, nil, nc)
 	require.NoError(t, err)
+	o.templates.SetCentralTagStore(o.tags)
 	o.pgKV, err = js.KeyValue(ctx, ec2placementgroup.KVBucketPlacementGroups)
 	require.NoError(t, err)
 	o.ltKV, err = js.KeyValue(ctx, ec2launchtemplate.KVBucketLaunchTemplates)
@@ -172,9 +158,18 @@ func TestCurrentBehaviour_AccountTermination_PlatformReapersReportDrainedOverHid
 	}, testAccountID)
 	require.NoError(t, err)
 
-	// Residue each owner's Describe skips.
-	publicKey := "keys/" + testAccountID + "/deploy"
-	o.store.failDelete = func(key string) bool { return key == publicKey }
+	// Residue each owner's Describe skips, as left by older releases or a
+	// crash between writes.
+	publicKey := "keys/" + testAccountID + "/old-deploy"
+	_, err = o.store.PutObject(ctx, &s3.PutObjectInput{
+		Bucket: aws.String(platformBucket), Key: aws.String(publicKey), Body: strings.NewReader("ssh-ed25519 AAAA"),
+	})
+	require.NoError(t, err)
+	staleTagged := "pg-0stale000000000"
+	_, err = o.tags.CreateTags(ctx, &ec2.CreateTagsInput{
+		Resources: []*string{aws.String(staleTagged)}, Tags: []*ec2.Tag{{Key: aws.String("team"), Value: aws.String("blue")}},
+	}, testAccountID)
+	require.NoError(t, err)
 	_, err = o.store.PutObject(ctx, &s3.PutObjectInput{
 		Bucket: aws.String(platformBucket), Key: aws.String("keys/" + testAccountID + "/key-0badbadbadbadbad0.json"),
 		Body: strings.NewReader("{not json"),
@@ -210,10 +205,6 @@ func TestCurrentBehaviour_AccountTermination_PlatformReapersReportDrainedOverHid
 	assert.ElementsMatch(t, []string{publicKey, "keys/" + testAccountID + "/key-0badbadbadbadbad0.json"}, o.objectKeys(t))
 	assert.Equal(t, []string{testAccountID + ".broken"}, accountKVKeys(t, o.pgKV))
 	assert.Equal(t, []string{orphanBody}, accountKVKeys(t, o.ltKV))
-	assert.Equal(t, sortedStrings(testAccountID+"."+pgID, testAccountID+"."+ltID), accountKVKeys(t, o.tagsKV))
-}
-
-func sortedStrings(s ...string) []string {
-	slices.Sort(s)
-	return s
+	// The owner deletes clear their own tags; only the stale entry remains.
+	assert.Equal(t, []string{testAccountID + "." + staleTagged}, accountKVKeys(t, o.tagsKV))
 }

spinifex/accountteardown/reapers_quota_override_current_behaviour_test.go

@@ -0,0 +1,61 @@
+package accountteardown
+
+//test:in-package — drives the real final-stage reapers against the real quota service to pin that teardown leaves the override bucket alone.
+
+import (
+	"testing"
+
+	"github.com/mulgadc/spinifex/internal/testkit"
+	"github.com/mulgadc/spinifex/spinifex/domains/admission/quota"
+	"github.com/mulgadc/spinifex/spinifex/foundation/state/kvutil"
+	handlers_iam "github.com/mulgadc/spinifex/spinifex/handlers/iam"
+	"github.com/nats-io/nats.go/jetstream"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+// No reaper in this package references the quota override bucket or calls an
+// override delete. Running the real engine and the real StageAccount reapers
+// against an account with a raised vCPU override, with usage wired, shows the
+// consequence: the override outlives the account record it belongs to, while
+// the usage counter the reapers do own is removed.
+func TestCurrentBehaviour_AccountTermination_QuotaOverridesSurviveTeardown(t *testing.T) {
+	_, _, js := testutil.StartTestJetStream(t)
+	ctx := testCtx(t)
+
+	usage, err := kvutil.GetOrCreateBucket(ctx, js, quota.KVBucketAccountUsage, 1)
+	require.NoError(t, err)
+	_, err = usage.Put(ctx, testAccountID, []byte(`{"vcpus":4}`))
+	require.NoError(t, err)
+
+	overridesKV, err := kvutil.GetOrCreateBucket(ctx, js, quota.KVBucketAccountQuota, 1)
+	require.NoError(t, err)
+	svc := quota.New(quota.Limits{Enabled: true, VCPUs: 16}, usage)
+	svc.SetOverrides(overridesKV)
+	unlimited := quota.Unlimited
+	require.NoError(t, svc.PutAccountQuota(ctx, testAccountID, quota.Overrides{VCPUs: &unlimited}, "operator"))
+
+	accounts := newFakeAccounts()
+	iam := &accountIAM{account: &handlers_iam.Account{AccountID: testAccountID, AccountName: testAccountName}}
+	engine := quietEngine(accounts, AccountReapers(iam, nil, usage)...)
+
+	result, err := engine.Teardown(ctx, request())
+	require.NoError(t, err)
+	assert.True(t, result.AccountDeleted)
+	assert.True(t, accounts.deleted)
+	assert.Zero(t, result.StuckCount())
+
+	// The reaper this package does own removed the usage counter, which shows
+	// which store teardown actually clears.
+	_, err = usage.Get(ctx, testAccountID)
+	assert.ErrorIs(t, err, jetstream.ErrKeyNotFound, "teardown's quota-usage reaper must remove the vCPU counter")
+
+	_, err = overridesKV.Get(ctx, testAccountID)
+	require.NoError(t, err, "the override key must still be present after the account is torn down")
+
+	over, effective, err := svc.GetAccountQuota(ctx, testAccountID)
+	require.NoError(t, err)
+	require.NotNil(t, over.VCPUs)
+	assert.Equal(t, quota.Unlimited, *over.VCPUs)
+	assert.Equal(t, quota.Unlimited, effective.VCPUs)
+}

spinifex/accountteardown/settings_kv_boundary_test.go

@@ -0,0 +1,89 @@
+package accountteardown
+
+//test:in-package — pins ADR-0003:S4 and ADR-0001 Context item 2 for the EC2
+// settings seal: the legacy engine must reach the owner only through the
+// versioned contract and NATS, never by reading the settings bucket itself.
+
+import (
+	"go/ast"
+	"go/parser"
+	"go/token"
+	"io/fs"
+	"path/filepath"
+	"runtime"
+	"strconv"
+	"strings"
+	"testing"
+)
+
+// forbiddenSettingsImports are imports that would let this package reach the
+// EC2 settings owner or the configured node set by a path other than the
+// versioned contract and NewClusterEngine's parameters.
+var forbiddenSettingsImports = []string{
+	"github.com/mulgadc/spinifex/spinifex/domains/ec2/account",
+	"github.com/mulgadc/spinifex/spinifex/bootstrap/config",
+}
+
+// forbiddenSettingsIdentifiers would let this package name the settings
+// bucket or its key directly instead of reaching it only through
+// contracts/ec2/v1 and NATS.
+var forbiddenSettingsIdentifiers = []string{"spinifex-ec2-account-settings", "KVBucketAccountSettings"}
+
+// Teardown reaches EC2 settings only over contracts/ec2/v1: no import of the
+// settings owner or bootstrap/config (nodes arrive via NewClusterEngine), and
+// no literal naming the settings bucket or its constant.
+func TestAccountTeardownHasNoSettingsKVAccess(t *testing.T) {
+	_, file, _, ok := runtime.Caller(0)
+	if !ok {
+		t.Fatal("cannot locate test source")
+	}
+	root := filepath.Dir(file)
+
+	fset := token.NewFileSet()
+	scanned := 0
+	err := filepath.WalkDir(root, func(path string, d fs.DirEntry, err error) error {
+		if err != nil || d.IsDir() || !strings.HasSuffix(path, ".go") || strings.HasSuffix(path, "_test.go") {
+			return err
+		}
+		f, perr := parser.ParseFile(fset, path, nil, parser.ParseComments)
+		if perr != nil {
+			return perr
+		}
+		scanned++
+
+		for _, imp := range f.Imports {
+			p, _ := strconv.Unquote(imp.Path.Value)
+			for _, bad := range forbiddenSettingsImports {
+				if p == bad {
+					t.Errorf("%s: imports %s; the legacy engine must reach the EC2 settings owner only through contracts/ec2/v1 and the configured node set NewClusterEngine already receives",
+						fset.Position(imp.Pos()), p)
+				}
+			}
+		}
+
+		ast.Inspect(f, func(n ast.Node) bool {
+			lit, ok := n.(*ast.BasicLit)
+			if !ok || lit.Kind != token.STRING {
+				return true
+			}
+			value, uerr := strconv.Unquote(lit.Value)
+			if uerr != nil {
+				return true
+			}
+			for _, bad := range forbiddenSettingsIdentifiers {
+				if value == bad {
+					t.Errorf("%s: names %q; the settings bucket and its key are the owner's own, never this package's",
+						fset.Position(lit.Pos()), value)
+				}
+			}
+			return true
+		})
+		return nil
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if scanned == 0 {
+		t.Fatalf("scanned no Go files under %s; the invariant would pass vacuously", root)
+	}
+}

spinifex/daemon/daemon.go

@@ -1026,6 +1026,7 @@ func (d *Daemon) subscribeAll() error {
 		{"ec2.GetSerialConsoleAccessStatus", handleNATSRequest(d.node, d.accountService.GetSerialConsoleAccessStatus), "spinifex-workers"},
 		{"ec2.EnableSerialConsoleAccess", handleNATSRequest(d.node, d.accountService.EnableSerialConsoleAccess), "spinifex-workers"},
 		{"ec2.DisableSerialConsoleAccess", handleNATSRequest(d.node, d.accountService.DisableSerialConsoleAccess), "spinifex-workers"},
+		{ec2v1.SealAccountSettingsSubject, handleNATSRequest(d.node, d.sealAccountSettings), "spinifex-workers"},
 		{clusterv1.NodeHealthSubject(d.node), d.handleHealthCheck, ""},
 		{clusterv1.NodesDiscoverSubject, d.handleNodeDiscover, ""},
 		{clusterv1.NodeStatusSubject, d.handleNodeStatus, ""},
@@ -1288,6 +1289,15 @@ func (d *Daemon) subscribeAll() error {
 		natsSub{viperblocklegacyv1.VolumeFencedSubject(d.node), d.handleVolumeFenced, ""},
 	)
 
+	// Per-node, no queue group: a second daemon claiming this node ID also
+	// answers, so the teardown probe sees a duplicate and refuses to seal. An
+	// invalid node ID serves nothing, so this daemon reads as incompatible.
+	if ec2v1.ValidSealCapabilityNode(d.node) {
+		subs = append(subs, natsSub{ec2v1.AccountSettingsSealCapabilitySubject(d.node), handleNATSRequest(d.node, d.accountSettingsSealCapability), ""})
+	} else {
+		slog.Error("Not serving account-settings seal capability: node ID is not a single NATS subject token", "node", d.node)
+	}
+
 	return d.registerNatsSubs(subs)
 }
 
@@ -1825,9 +1835,6 @@ func (d *Daemon) startCluster() error {
 		return fmt.Errorf("failed to get tags KV bucket: %w", err)
 	}
 	d.tagsService = ec2tags.NewTagsServiceImpl(d.config, tagsKV)
-	// Key pairs keep their creation tags in their own metadata; project them so
-	// describe-tags sees them, and clear them when the key pair is deleted.
-	d.keyService.SetCentralTagStore(d.tagsService)
 
 	d.eigwService, err = initServiceWithRetry("EIGW service", func() (*ec2eigw.EgressOnlyIGWServiceImpl, error) {
 		return ec2eigw.NewEgressOnlyIGWServiceImplWithNATS(d.ctx, d.config, d.natsConn)
@@ -1856,6 +1863,7 @@ func (d *Daemon) startCluster() error {
 	if err != nil {
 		return fmt.Errorf("failed to initialize launch template service: %w", err)
 	}
+	d.wirePlatformTagIndex()
 
 	d.spotInstanceService, err = initServiceWithRetry("spot instance service", func() (*ec2spotinstance.SpotInstanceServiceImpl, error) {
 		return ec2spotinstance.NewSpotInstanceServiceImplWithNATS(d.ctx, d.config, d.natsConn)
@@ -1994,6 +2002,12 @@ func (d *Daemon) startCluster() error {
 	if err != nil {
 		return fmt.Errorf("failed to initialize account settings service: %w", err)
 	}
+	accountStatusJS, err := jetstream.New(d.natsConn)
+	if err != nil {
+		return fmt.Errorf("jetstream handle for account status reader: %w", err)
+	}
+	var _ ec2account.AccountTerminationChecker = (*handlers_iam.AccountStatusReader)(nil)
+	d.accountService.SetAccountTerminationChecker(handlers_iam.NewAccountStatusReader(accountStatusJS))
 
 	// ELBv2 resolves listener certificate private keys through its own ACM
 	// Store over the same JetStream bucket the ACM service writes, so it needs
@@ -3628,3 +3642,12 @@ func probeGPU() gpuProbeResult {
 	r.Capable = len(r.Devices) > 0 && ((r.IOMMUActive && r.VFIOPresent) || hasMIG)
 	return r
 }
+
+// wirePlatformTagIndex gives the key-pair, placement-group and launch-template
+// owners the tag index their deletes clear; key pairs also project their
+// metadata-held creation tags into it so describe-tags sees them.
+func (d *Daemon) wirePlatformTagIndex() {
+	d.keyService.SetCentralTagStore(d.tagsService)
+	d.placementGroupService.SetCentralTagStore(d.tagsService)
+	d.launchTemplateService.SetCentralTagStore(d.tagsService)
+}

spinifex/daemon/daemon_handlers_account_settings_seal.go

@@ -0,0 +1,23 @@
+package daemon
+
+import (
+	"context"
+
+	ec2v1 "github.com/mulgadc/spinifex/contracts/ec2/v1"
+)
+
+// sealAccountSettings serves the irreversible EC2 settings seal. The account
+// comes from the request body, not X-Account-ID: the sealed account is part
+// of the versioned contract, so the header is ignored.
+func (d *Daemon) sealAccountSettings(ctx context.Context, req *ec2v1.SealAccountSettingsRequest, _ string) (*ec2v1.SealAccountSettingsReply, error) {
+	seal, err := d.accountService.SealAccount(ctx, req.AccountID)
+	if err != nil {
+		return nil, err
+	}
+	return &ec2v1.SealAccountSettingsReply{
+		AccountID:      req.AccountID,
+		CorrelationRef: req.CorrelationRef,
+		Disposition:    ec2v1.AccountSettingsDispositionSealed,
+		Revision:       seal.Revision,
+	}, nil
+}

spinifex/daemon/daemon_handlers_account_settings_seal_capability.go

@@ -0,0 +1,23 @@
+package daemon
+
+import (
+	"context"
+
+	ec2v1 "github.com/mulgadc/spinifex/contracts/ec2/v1"
+	awserrors "github.com/mulgadc/spinifex/spinifex/foundation/aws/errors"
+)
+
+// accountSettingsSealCapability serves this daemon's read-only seal capability.
+// It reads no KV and takes no account; a request naming another node is a
+// caller mistake, since the subject is already addressed to this node.
+func (d *Daemon) accountSettingsSealCapability(_ context.Context, req *ec2v1.AccountSettingsSealCapabilityRequest, _ string) (*ec2v1.AccountSettingsSealCapabilityReply, error) {
+	if req.Node != d.node {
+		return nil, awserrors.Errorf(awserrors.ErrorInvalidParameterValue,
+			"seal capability request for node %q answered by node %q", req.Node, d.node)
+	}
+	return &ec2v1.AccountSettingsSealCapabilityReply{
+		Node:              d.node,
+		CapabilityVersion: ec2v1.AccountSettingsSealCapabilityVersion,
+		CorrelationRef:    req.CorrelationRef,
+	}, nil
+}

spinifex/daemon/daemon_handlers_account_settings_seal_capability_test.go

@@ -0,0 +1,115 @@
+package daemon
+
+import (
+	"encoding/json"
+	"testing"
+	"time"
+
+	ec2v1 "github.com/mulgadc/spinifex/contracts/ec2/v1"
+	"github.com/mulgadc/spinifex/internal/testkit"
+	awserrors "github.com/mulgadc/spinifex/spinifex/foundation/aws/errors"
+	"github.com/nats-io/nats.go"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+// newCapabilityTestDaemon builds a daemon serving its real subscription table
+// with node as its configured node ID.
+func newCapabilityTestDaemon(t *testing.T, node string) (*nats.Conn, *nats.Conn) {
+	t.Helper()
+	ns, nc, _ := testutil.StartTestJetStream(t)
+
+	d := newSubscribeTestDaemon(t, nc, "", "")
+	d.node = node
+	require.NoError(t, d.subscribeAll())
+	require.NoError(t, nc.Flush())
+
+	caller, err := nats.Connect(ns.ClientURL())
+	require.NoError(t, err)
+	t.Cleanup(caller.Close)
+	return nc, caller
+}
+
+func capabilityRequest(t *testing.T, nc *nats.Conn, node string) *nats.Msg {
+	t.Helper()
+	data, err := json.Marshal(ec2v1.AccountSettingsSealCapabilityRequest{Node: node, CorrelationRef: "req-1"})
+	require.NoError(t, err)
+	reply, err := nc.Request(ec2v1.AccountSettingsSealCapabilitySubject(node), data, 5*time.Second)
+	require.NoError(t, err)
+	return reply
+}
+
+// The responder answers its own node with version 1, the configured node and
+// no KV access (no settings service is wired into the test daemon at all).
+func TestAccountSettingsSealCapabilityAnswersOwnNode(t *testing.T) {
+	_, caller := newCapabilityTestDaemon(t, "node-1")
+
+	raw := capabilityRequest(t, caller, "node-1")
+	assert.Equal(t, "node-1", raw.Header.Get("X-Node-ID"))
+
+	var reply ec2v1.AccountSettingsSealCapabilityReply
+	require.NoError(t, json.Unmarshal(raw.Data, &reply))
+	assert.Equal(t, "node-1", reply.Node)
+	assert.Equal(t, ec2v1.AccountSettingsSealCapabilityVersion, reply.CapabilityVersion)
+	assert.Equal(t, "req-1", reply.CorrelationRef)
+}
+
+// A request naming a node other than this daemon's own is refused, not
+// answered as if it were this node.
+func TestAccountSettingsSealCapabilityRefusesOtherNode(t *testing.T) {
+	_, caller := newCapabilityTestDaemon(t, "node-1")
+
+	data, err := json.Marshal(ec2v1.AccountSettingsSealCapabilityRequest{Node: "node-2"})
+	require.NoError(t, err)
+	raw, err := caller.Request(ec2v1.AccountSettingsSealCapabilitySubject("node-1"), data, 5*time.Second)
+	require.NoError(t, err)
+	assert.Equal(t, awserrors.ErrorInvalidParameterValue, replyErrCode(t, raw.Data))
+}
+
+// The subject is per node with no queue group, so two daemons configured
+// with the same node ID both answer a request addressed to that node.
+func TestAccountSettingsSealCapabilityHasNoQueueGroup(t *testing.T) {
+	_, nc, _ := testutil.StartTestJetStream(t)
+
+	first := newSubscribeTestDaemon(t, nc, "", "")
+	first.node = "node-1"
+	require.NoError(t, first.subscribeAll())
+
+	second := newSubscribeTestDaemon(t, nc, "", "")
+	second.node = "node-1"
+	require.NoError(t, second.subscribeAll())
+	require.NoError(t, nc.Flush())
+
+	sub, err := nc.SubscribeSync(nats.NewInbox())
+	require.NoError(t, err)
+	t.Cleanup(func() { _ = sub.Unsubscribe() })
+
+	data, err := json.Marshal(ec2v1.AccountSettingsSealCapabilityRequest{Node: "node-1"})
+	require.NoError(t, err)
+	msg := nats.NewMsg(ec2v1.AccountSettingsSealCapabilitySubject("node-1"))
+	msg.Data = data
+	msg.Reply = sub.Subject
+	require.NoError(t, nc.PublishMsg(msg))
+	require.NoError(t, nc.Flush())
+
+	seen := 0
+	for {
+		_, err := sub.NextMsg(300 * time.Millisecond)
+		if err != nil {
+			break
+		}
+		seen++
+	}
+	assert.Equal(t, 2, seen)
+}
+
+// A wildcard node ID must not become a wildcard subscription: built unchecked,
+// "node-*" would answer probes addressed to node-1 on node-1's behalf.
+func TestAccountSettingsSealCapabilityNotServedForInvalidNodeID(t *testing.T) {
+	_, caller := newCapabilityTestDaemon(t, "node-*")
+
+	data, err := json.Marshal(ec2v1.AccountSettingsSealCapabilityRequest{Node: "node-1"})
+	require.NoError(t, err)
+	_, err = caller.Request(ec2v1.AccountSettingsSealCapabilitySubject("node-1"), data, time.Second)
+	require.ErrorIs(t, err, nats.ErrNoResponders)
+}

spinifex/daemon/daemon_handlers_account_settings_seal_test.go

@@ -0,0 +1,101 @@
+package daemon
+
+import (
+	"context"
+	"encoding/json"
+	"testing"
+	"time"
+
+	ec2v1 "github.com/mulgadc/spinifex/contracts/ec2/v1"
+	"github.com/mulgadc/spinifex/internal/testkit"
+	ec2account "github.com/mulgadc/spinifex/spinifex/domains/ec2/account"
+	awserrors "github.com/mulgadc/spinifex/spinifex/foundation/aws/errors"
+	"github.com/nats-io/nats.go"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+// sealTestChecker answers AccountTerminating from a fixed set, standing in
+// for IAM in these daemon-level tests.
+type sealTestChecker struct {
+	terminating map[string]bool
+}
+
+func (c *sealTestChecker) AccountTerminating(_ context.Context, accountID string) (bool, error) {
+	return c.terminating[accountID], nil
+}
+
+// newSealTestDaemon builds a daemon serving its real subscription table over
+// a real account settings service, wired with checker as its termination
+// precondition.
+func newSealTestDaemon(t *testing.T, checker ec2account.AccountTerminationChecker) (*Daemon, *nats.Conn) {
+	t.Helper()
+	ns, nc, _ := testutil.StartTestJetStream(t)
+
+	acctSvc, err := ec2account.NewAccountSettingsServiceImplWithNATS(t.Context(), nil, nc)
+	require.NoError(t, err)
+	acctSvc.SetAccountTerminationChecker(checker)
+
+	d := newSubscribeTestDaemon(t, nc, "", "")
+	d.accountService = acctSvc
+	require.NoError(t, d.subscribeAll())
+	require.NoError(t, nc.Flush())
+
+	caller, err := nats.Connect(ns.ClientURL())
+	require.NoError(t, err)
+	t.Cleanup(caller.Close)
+	return d, caller
+}
+
+const (
+	sealTestTerminating = "111122223333"
+	sealTestActive      = "444455556666"
+)
+
+func sealRequest(t *testing.T, nc *nats.Conn, body any, header string) []byte {
+	t.Helper()
+	data, err := json.Marshal(body)
+	require.NoError(t, err)
+	reply, err := natsRequestAs(nc, ec2v1.SealAccountSettingsSubject, data, header, 5*time.Second)
+	require.NoError(t, err)
+	return reply.Data
+}
+
+// The header is ignored: the body account is what gets sealed, even when
+// X-Account-ID names a different one.
+func TestSealAccountSettingsSubjectSealsTheBodyAccount(t *testing.T) {
+	_, nc := newSealTestDaemon(t, &sealTestChecker{terminating: map[string]bool{sealTestTerminating: true}})
+
+	raw := sealRequest(t, nc, ec2v1.SealAccountSettingsRequest{AccountID: sealTestTerminating, CorrelationRef: "req-1"}, sealTestActive)
+	require.Empty(t, replyErrCode(t, raw), string(raw))
+
+	var reply ec2v1.SealAccountSettingsReply
+	require.NoError(t, json.Unmarshal(raw, &reply))
+	assert.Equal(t, sealTestTerminating, reply.AccountID)
+	assert.Equal(t, "req-1", reply.CorrelationRef)
+	assert.Equal(t, ec2v1.AccountSettingsDispositionSealed, reply.Disposition)
+	assert.NotZero(t, reply.Revision)
+}
+
+// Only an account IAM holds as TERMINATING may be sealed; ACTIVE is refused
+// before any write.
+func TestSealAccountSettingsSubjectRefusesAnActiveAccount(t *testing.T) {
+	_, nc := newSealTestDaemon(t, &sealTestChecker{terminating: map[string]bool{sealTestTerminating: true}})
+
+	raw := sealRequest(t, nc, ec2v1.SealAccountSettingsRequest{AccountID: sealTestActive}, "")
+	assert.Equal(t, awserrors.ErrorIncorrectState, replyErrCode(t, raw))
+}
+
+// A request without a valid account is refused, not sealed.
+func TestSealAccountSettingsSubjectRefusesAMissingAccount(t *testing.T) {
+	_, nc := newSealTestDaemon(t, &sealTestChecker{terminating: map[string]bool{sealTestTerminating: true}})
+
+	assert.Equal(t, awserrors.ErrorMissingParameter, replyErrCode(t, sealRequest(t, nc, ec2v1.SealAccountSettingsRequest{}, "")))
+}
+
+// A daemon that does not serve the subject leaves the request with no responder.
+func TestSealAccountSettingsSubjectHasNoResponderWithoutTheDaemon(t *testing.T) {
+	_, nc, _ := testutil.StartTestJetStream(t)
+	_, err := natsRequestAs(nc, ec2v1.SealAccountSettingsSubject, []byte(`{"account_id":"111122223333"}`), "", time.Second)
+	require.ErrorIs(t, err, nats.ErrNoResponders)
+}

spinifex/daemon/daemon_handlers_platform_inventory_test.go

@@ -164,6 +164,9 @@ func newInventoryCluster(t *testing.T) *inventoryCluster {
 	d.placementGroupService = groups
 	d.launchTemplateService = templates
 	require.NoError(t, d.subscribeAll())
+	// The caller is a separate connection, so the server must hold the SUBs
+	// before it publishes or the request can see no responders.
+	require.NoError(t, nc.Flush())
 
 	c.caller, err = nats.Connect(ns.ClientURL())
 	require.NoError(t, err)

spinifex/daemon/platform_disclosure_current_behaviour_test.go

@@ -7,7 +7,6 @@ import (
 	"bytes"
 	"encoding/json"
 	"slices"
-	"strings"
 	"testing"
 	"time"
 
@@ -56,12 +55,15 @@ func newDisclosureCluster(t *testing.T) *disclosureCluster {
 	require.NoError(t, err)
 	d.keyService = ec2key.NewKeyServiceImplWithStore(store, disclosureBucket)
 	d.tagsService = ec2tags.NewTagsServiceImplWithStore(cfg, store, tagsKV)
-	d.keyService.SetCentralTagStore(d.tagsService)
 	d.placementGroupService, err = ec2placementgroup.NewPlacementGroupServiceImplWithNATS(t.Context(), nil, nc)
 	require.NoError(t, err)
 	d.launchTemplateService, err = ec2launchtemplate.NewLaunchTemplateServiceImplWithNATS(t.Context(), nil, nc)
 	require.NoError(t, err)
+	d.wirePlatformTagIndex()
 	require.NoError(t, d.subscribeAll())
+	// The caller is a separate connection, so the server must hold the SUBs
+	// before it publishes or the request can see no responders.
+	require.NoError(t, nc.Flush())
 
 	caller, err := nats.Connect(ns.ClientURL())
 	require.NoError(t, err)
@@ -225,12 +227,13 @@ func TestCurrentBehaviour_AccountTermination_DirectNATSCallerCannotSeeHiddenResi
 	assert.Equal(t, before.taggedResources, after.taggedResources)
 }
 
-// Pins current behaviour that differs from AWS: deleting a placement group or
-// launch template leaves its tag-index entry, which DescribeTags keeps
-// returning with its values. A key-pair delete clears its entry.
-func TestCurrentBehaviour_AccountTermination_PlatformDeletesLeaveTagIndexEntries(t *testing.T) {
+// Deleting a key pair, placement group or launch template clears its
+// tag-index entry, so DescribeTags no longer returns its rows, as in AWS.
+func TestPlatformDeletesClearTagIndexEntries(t *testing.T) {
 	c := newDisclosureCluster(t)
 	s := c.seed(t, disclosureAccount, "a")
+	before := c.describe(t, disclosureAccount)
+	require.Equal(t, sortedIDs(s.keyID, s.groupID, s.templateID), before.taggedResources)
 
 	c.call(t, "ec2.DeleteKeyPair", disclosureAccount, &ec2.DeleteKeyPairInput{KeyName: aws.String(s.keyName)}, nil)
 	c.call(t, "ec2.DeletePlacementGroup", disclosureAccount, &ec2.DeletePlacementGroupInput{GroupName: aws.String(s.groupName)}, nil)
@@ -240,13 +243,8 @@ func TestCurrentBehaviour_AccountTermination_PlatformDeletesLeaveTagIndexEntries
 	assert.Empty(t, got.keyPairs)
 	assert.Empty(t, got.placementGroups)
 	assert.Empty(t, got.launchTemplates)
-	assert.Equal(t, sortedIDs(s.groupID, s.templateID), got.taggedResources)
-
+	assert.Empty(t, got.taggedResources)
 	var tags ec2.DescribeTagsOutput
 	c.call(t, "ec2.DescribeTags", disclosureAccount, &ec2.DescribeTagsInput{}, &tags)
-	for _, row := range tags.Tags {
-		assert.Equal(t, "owner", aws.StringValue(row.Key))
-		assert.Equal(t, "a", aws.StringValue(row.Value))
-		assert.True(t, strings.HasPrefix(aws.StringValue(row.ResourceId), "pg-") || strings.HasPrefix(aws.StringValue(row.ResourceId), "lt-"))
-	}
+	assert.Empty(t, tags.Tags)
 }

spinifex/daemon/vm_adapters_trace_test.go

@@ -73,7 +73,9 @@ func TestEbsRequestWithTrace_ProducerConsumerLinked(t *testing.T) {
 	sr := withRecordedSpans(t)
 	daemon := createTestDaemon(t, sharedNATSURL)
 
+	consumerEnded := make(chan struct{})
 	sub, err := daemon.natsConn.Subscribe("ebs.node-1.unmount", func(msg *nats.Msg) {
+		defer close(consumerEnded)
 		// Mirrors viperblockd's ebs.unmount handler: open a consumer span
 		// joining whatever trace context the producer injected.
 		_, span := natsmsg.StartConsumerSpan(msg)
@@ -89,6 +91,9 @@ func TestEbsRequestWithTrace_ProducerConsumerLinked(t *testing.T) {
 
 	adapter := newVolumeMounterAdapter(daemon.natsConn, daemon.node, nil)
 	require.NoError(t, adapter.UnmountOne(t.Context(), "", viperblocklegacyv1.EBSRequest{Name: "vol-linked"}))
+	// The handler replies before its deferred span.End runs, so the consumer
+	// span may not be recorded yet when UnmountOne returns.
+	<-consumerEnded
 
 	spans := sr.Ended()
 	require.Len(t, spans, 2, "expected one producer span and one consumer span")

spinifex/domains/admission/quota/live.go

@@ -39,7 +39,11 @@ func (s *Service) EnforceVPCs(ctx context.Context, natsConn *nats.Conn, accountI
 	if err != nil {
 		return err
 	}
-	return exceeds(len(out.Vpcs), want, s.limitsFor(ctx, accountID).VPCs)
+	limits, err := s.limitsFor(ctx, accountID)
+	if err != nil {
+		return err
+	}
+	return exceeds(len(out.Vpcs), want, limits.VPCs)
 }
 
 // EnforceSubnets gates CreateSubnet on the account's live DescribeSubnets count.
@@ -52,7 +56,11 @@ func (s *Service) EnforceSubnets(ctx context.Context, natsConn *nats.Conn, accou
 	if err != nil {
 		return err
 	}
-	return exceeds(len(out.Subnets), want, s.limitsFor(ctx, accountID).Subnets)
+	limits, err := s.limitsFor(ctx, accountID)
+	if err != nil {
+		return err
+	}
+	return exceeds(len(out.Subnets), want, limits.Subnets)
 }
 
 // EnforceEIPs gates AllocateAddress on the account's live DescribeAddresses count.
@@ -64,7 +72,11 @@ func (s *Service) EnforceEIPs(ctx context.Context, natsConn *nats.Conn, accountI
 	if err != nil {
 		return err
 	}
-	return exceeds(len(out.Addresses), want, s.limitsFor(ctx, accountID).EIPs)
+	limits, err := s.limitsFor(ctx, accountID)
+	if err != nil {
+		return err
+	}
+	return exceeds(len(out.Addresses), want, limits.EIPs)
 }
 
 // EnforceLoadBalancers gates CreateLoadBalancer on the account's live count.
@@ -77,7 +89,11 @@ func (s *Service) EnforceLoadBalancers(ctx context.Context, natsConn *nats.Conn,
 	if err != nil {
 		return err
 	}
-	return exceeds(len(out.LoadBalancers), want, s.limitsFor(ctx, accountID).LoadBalancers)
+	limits, err := s.limitsFor(ctx, accountID)
+	if err != nil {
+		return err
+	}
+	return exceeds(len(out.LoadBalancers), want, limits.LoadBalancers)
 }
 
 // EnforceRDSInstances gates CreateDBInstance on the account's live count. A DB
@@ -91,5 +107,9 @@ func (s *Service) EnforceRDSInstances(ctx context.Context, natsConn *nats.Conn,
 	if err != nil {
 		return err
 	}
-	return exceeds(len(out.DBInstances), want, s.limitsFor(ctx, accountID).RDSInstances)
+	limits, err := s.limitsFor(ctx, accountID)
+	if err != nil {
+		return err
+	}
+	return exceeds(len(out.DBInstances), want, limits.RDSInstances)
 }

spinifex/domains/admission/quota/override.go

@@ -8,12 +8,16 @@ import (
 	"log/slog"
 	"time"
 
+	awserrors "github.com/mulgadc/spinifex/spinifex/foundation/aws/errors"
+	"github.com/mulgadc/spinifex/spinifex/foundation/state/kvutil"
 	"github.com/nats-io/nats.go/jetstream"
 )
 
 // Overrides raises or lowers individual dimensions for one account. Every field
 // is a pointer so an explicit 0 (a limit that denies everything) is distinct
-// from an absent field, which inherits the configured value.
+// from an absent field, which inherits the configured value. This is also the
+// admin wire type: an admin request body can carry only these fields, so it can
+// never set or clear a stored record's state.
 type Overrides struct {
 	VCPUs         *int `json:"vcpus,omitempty"`
 	VPCs          *int `json:"vpcs,omitempty"`
@@ -29,11 +33,62 @@ type Overrides struct {
 	UpdatedBy string `json:"updated_by,omitempty"`
 }
 
-// Empty reports whether no dimension is overridden, in which case the record
-// carries nothing but provenance and the account resolves to the config.
-func (o Overrides) Empty() bool {
-	return o.VCPUs == nil && o.VPCs == nil && o.Subnets == nil && o.EIPs == nil &&
-		o.VolumesGiB == nil && o.Volumes == nil && o.RDSInstances == nil && o.LoadBalancers == nil
+// OverrideState is a stored override record's lifecycle state. "" is active;
+// a record written before the state field existed has none and stays active.
+type OverrideState string
+
+// OverrideStateSealed marks a record no writer may change or recreate.
+const OverrideStateSealed OverrideState = "sealed"
+
+// overrideRecord is the stored form: the admin-settable fields inline, plus a
+// state no admin request can read or write.
+type overrideRecord struct {
+	Overrides
+
+	State OverrideState `json:"state,omitempty"`
+}
+
+// errOverridesSealed marks a read or write that found a sealed record, so a
+// caller maps it to Blocked without logging it as a fault.
+var errOverridesSealed = errors.New(awserrors.ErrorBlocked)
+
+// errOverridesContended marks a write whose CAS retries were exhausted under
+// concurrent writers, bounding the retry loop instead of spinning.
+var errOverridesContended = errors.New("quota override: CAS retries exhausted")
+
+// overrideMaxAttempts bounds a write's retries against a revision that keeps
+// moving under concurrent writers, so it fails closed instead of spinning.
+const overrideMaxAttempts = 5
+
+// checkActive rejects a sealed or unknown-state record; "" is active and
+// passes through. A caller must never read or mutate a record that failed
+// this as if it were active.
+func checkActive(rec overrideRecord) error {
+	switch rec.State {
+	case "":
+		return nil
+	case OverrideStateSealed:
+		return errOverridesSealed
+	default:
+		return fmt.Errorf("quota override: unknown state %q", rec.State)
+	}
+}
+
+// mapOverrideError turns a recordFor, PutAccountQuota or clear failure into
+// the AWS error a caller returns: sealed is an expected, unlogged Blocked;
+// everything else is InternalError, logged with the account ID.
+func mapOverrideError(ctx context.Context, accountID string, err error) error {
+	switch {
+	case err == nil:
+		return nil
+	case errors.Is(err, errOverridesSealed):
+		return errors.New(awserrors.ErrorBlocked)
+	case errors.Is(err, errOverridesContended):
+		slog.ErrorContext(ctx, "Quota override CAS retries exhausted under contention", "accountID", accountID)
+	default:
+		slog.ErrorContext(ctx, "Quota override operation failed", "accountID", accountID, "err", err)
+	}
+	return errors.New(awserrors.ErrorInternalError)
 }
 
 // apply overlays the set fields onto base, leaving Enabled and the Bedrock
@@ -60,72 +115,128 @@ func (o Overrides) apply(base Limits) Limits {
 	return out
 }
 
-// limitsFor resolves the effective limits for accountID: its overrides where
-// set, the configured limits everywhere else. A read failure falls back to the
-// configured limits rather than to no limit, so a KV outage cannot lift a cap.
-func (s *Service) limitsFor(ctx context.Context, accountID string) Limits {
-	over, err := s.overridesFor(ctx, accountID)
-	if err != nil {
-		slog.WarnContext(ctx, "quota: override read failed, using configured limits",
-			"account", accountID, "err", err)
-		return s.limits
-	}
-	return over.apply(s.limits)
+// Empty reports whether no dimension is overridden, in which case the record
+// carries nothing but provenance and the account resolves to the config.
+func (o Overrides) Empty() bool {
+	return o.VCPUs == nil && o.VPCs == nil && o.Subnets == nil && o.EIPs == nil &&
+		o.VolumesGiB == nil && o.Volumes == nil && o.RDSInstances == nil && o.LoadBalancers == nil
 }
 
-// overridesFor reads accountID's override record. An absent key is the normal
-// case for nearly every account and yields the zero value, not an error; only a
-// genuine read or decode failure is returned.
-func (s *Service) overridesFor(ctx context.Context, accountID string) (Overrides, error) {
+// recordFor reads and decodes accountID's stored override record. An absent
+// key is the active zero value, not an error; only a genuine read or decode
+// failure is returned.
+func (s *Service) recordFor(ctx context.Context, accountID string) (overrideRecord, error) {
 	if s == nil || s.overrides == nil {
-		return Overrides{}, nil
+		return overrideRecord{}, nil
 	}
 	entry, err := s.overrides.Get(ctx, accountID)
 	if errors.Is(err, jetstream.ErrKeyNotFound) {
-		return Overrides{}, nil
+		return overrideRecord{}, nil
 	}
 	if err != nil {
-		return Overrides{}, fmt.Errorf("read quota override for %s: %w", accountID, err)
+		return overrideRecord{}, fmt.Errorf("read quota override for %s: %w", accountID, err)
 	}
-	var over Overrides
-	if err := json.Unmarshal(entry.Value(), &over); err != nil {
-		return Overrides{}, fmt.Errorf("decode quota override for %s: %w", accountID, err)
+	var rec overrideRecord
+	if err := json.Unmarshal(entry.Value(), &rec); err != nil {
+		return overrideRecord{}, fmt.Errorf("decode quota override for %s: %w", accountID, err)
 	}
-	return over, nil
+	return rec, nil
 }
 
-// GetAccountQuota returns accountID's stored overrides and the limits they
-// resolve to. A read failure is surfaced here rather than masked, because an
-// operator inspecting a quota must not be shown the fallback as if it were real.
+// limitsFor resolves accountID's effective limits: overrides where set, the
+// config elsewhere. A read or decode failure falls back to the config so an
+// outage cannot lift a cap; a sealed or unknown-state record refuses instead.
+func (s *Service) limitsFor(ctx context.Context, accountID string) (Limits, error) {
+	rec, err := s.recordFor(ctx, accountID)
+	if err != nil {
+		slog.WarnContext(ctx, "quota: override read failed, using configured limits",
+			"account", accountID, "err", err)
+		return s.limits, nil
+	}
+	if err := checkActive(rec); err != nil {
+		return Limits{}, mapOverrideError(ctx, accountID, err)
+	}
+	return rec.apply(s.limits), nil
+}
+
+// GetAccountQuota returns accountID's stored overrides and resolved limits.
+// Read failures surface rather than show the fallback as real; a sealed or
+// unknown-state record returns Blocked or InternalError.
 func (s *Service) GetAccountQuota(ctx context.Context, accountID string) (Overrides, Limits, error) {
-	over, err := s.overridesFor(ctx, accountID)
+	rec, err := s.recordFor(ctx, accountID)
 	if err != nil {
 		return Overrides{}, Limits{}, err
 	}
-	return over, over.apply(s.limits), nil
+	if err := checkActive(rec); err != nil {
+		return Overrides{}, Limits{}, mapOverrideError(ctx, accountID, err)
+	}
+	return rec.Overrides, rec.apply(s.limits), nil
 }
 
-// PutAccountQuota stores accountID's overrides, stamping provenance. An empty
-// set deletes the record so the account resolves to the config, which keeps
-// "has this account been touched?" a single existence check.
+// PutAccountQuota stores or clears accountID's overrides, stamping provenance
+// on a store. Both branches are revision-guarded and refuse a sealed record,
+// so a seal landing between a read and a write fails the CAS.
 func (s *Service) PutAccountQuota(ctx context.Context, accountID string, over Overrides, updatedBy string) error {
 	if s == nil || s.overrides == nil {
 		return errors.New("quota: override bucket not configured")
 	}
 	if over.Empty() {
-		if err := s.overrides.Delete(ctx, accountID); err != nil && !errors.Is(err, jetstream.ErrKeyNotFound) {
-			return fmt.Errorf("clear quota override for %s: %w", accountID, err)
-		}
-		return nil
+		return s.clearAccountQuota(ctx, accountID)
 	}
+
 	over.UpdatedAt = time.Now().UTC().Format(time.RFC3339)
 	over.UpdatedBy = updatedBy
-	data, err := json.Marshal(over)
-	if err != nil {
-		return fmt.Errorf("encode quota override for %s: %w", accountID, err)
-	}
-	if _, err := s.overrides.Put(ctx, accountID, data); err != nil {
-		return fmt.Errorf("store quota override for %s: %w", accountID, err)
-	}
-	return nil
+
+	_, err := kvutil.Update(ctx, s.overrides, accountID, kvutil.CASConfig{
+		CreateIfAbsent: true,
+		Attempts:       overrideMaxAttempts,
+		Exhausted:      func(string, int) error { return errOverridesContended },
+	}, func(rec *overrideRecord) (bool, error) {
+		if err := checkActive(*rec); err != nil {
+			return false, err
+		}
+		rec.Overrides = over
+		// Every success is a leader-accepted write over an active value: a
+		// seal landing between the read and the write fails this CAS.
+		return true, nil
+	})
+	return mapOverrideError(ctx, accountID, err)
+}
+
+// clearAccountQuota deletes an active override at the revision it read. An
+// observed-absent key writes nothing, so a stale read cannot unseal; a sealed
+// record refuses rather than being erased and later recreated active.
+func (s *Service) clearAccountQuota(ctx context.Context, accountID string) error {
+	for range overrideMaxAttempts {
+		entry, err := s.overrides.Get(ctx, accountID)
+		if errors.Is(err, jetstream.ErrKeyNotFound) {
+			return nil
+		}
+		if err != nil {
+			return mapOverrideError(ctx, accountID, fmt.Errorf("read quota override for clear: %w", err))
+		}
+
+		var rec overrideRecord
+		if err := json.Unmarshal(entry.Value(), &rec); err != nil {
+			return mapOverrideError(ctx, accountID, fmt.Errorf("decode quota override for clear: %w", err))
+		}
+		if err := checkActive(rec); err != nil {
+			return mapOverrideError(ctx, accountID, err)
+		}
+
+		err = s.overrides.Delete(ctx, accountID, jetstream.LastRevision(entry.Revision()))
+		switch {
+		case err == nil:
+			return nil
+		case errors.Is(err, jetstream.ErrKeyNotFound):
+			return nil
+		case isRevisionConflict(err):
+			// A concurrent write landed between the read and the guarded
+			// delete; retry against whatever revision now holds.
+			continue
+		default:
+			return mapOverrideError(ctx, accountID, fmt.Errorf("clear quota override: %w", err))
+		}
+	}
+	return mapOverrideError(ctx, accountID, errOverridesContended)
 }

spinifex/domains/admission/quota/override_kv_test.go

@@ -31,7 +31,8 @@ func newOverrideService(t *testing.T) (*Service, jetstream.KeyValue) {
 func TestLimitsForAbsentKeyInherits(t *testing.T) {
 	svc, _ := newOverrideService(t)
 
-	got := svc.limitsFor(t.Context(), "123456789012")
+	got, err := svc.limitsFor(t.Context(), "123456789012")
+	require.NoError(t, err)
 	require.Equal(t, baseLimits(), got)
 }
 
@@ -44,10 +45,14 @@ func TestLimitsForIsPerAccount(t *testing.T) {
 	require.NoError(t, svc.PutAccountQuota(t.Context(), upgraded,
 		Overrides{VCPUs: ptr(32), RDSInstances: ptr(8)}, "operator"))
 
-	require.Equal(t, 32, svc.limitsFor(t.Context(), upgraded).VCPUs)
-	require.Equal(t, 8, svc.limitsFor(t.Context(), upgraded).RDSInstances)
-	require.Equal(t, baseLimits().VCPUs, svc.limitsFor(t.Context(), sandbox).VCPUs)
-	require.Equal(t, baseLimits().RDSInstances, svc.limitsFor(t.Context(), sandbox).RDSInstances)
+	upgradedLimits, err := svc.limitsFor(t.Context(), upgraded)
+	require.NoError(t, err)
+	sandboxLimits, err := svc.limitsFor(t.Context(), sandbox)
+	require.NoError(t, err)
+	require.Equal(t, 32, upgradedLimits.VCPUs)
+	require.Equal(t, 8, upgradedLimits.RDSInstances)
+	require.Equal(t, baseLimits().VCPUs, sandboxLimits.VCPUs)
+	require.Equal(t, baseLimits().RDSInstances, sandboxLimits.RDSInstances)
 }
 
 // A stored override must round-trip through KV with its provenance, and an
@@ -79,7 +84,9 @@ func TestPutAccountQuotaEmptyClears(t *testing.T) {
 
 	_, err := kv.Get(t.Context(), account)
 	require.ErrorIs(t, err, jetstream.ErrKeyNotFound)
-	require.Equal(t, baseLimits(), svc.limitsFor(t.Context(), account))
+	limits, err := svc.limitsFor(t.Context(), account)
+	require.NoError(t, err)
+	require.Equal(t, baseLimits(), limits)
 }
 
 // Clearing an account that never had a record must succeed rather than fail on
@@ -97,8 +104,10 @@ func TestPutAccountQuotaExplicitZero(t *testing.T) {
 
 	require.NoError(t, svc.PutAccountQuota(t.Context(), account, Overrides{VPCs: ptr(0)}, "operator"))
 
-	require.Equal(t, 0, svc.limitsFor(t.Context(), account).VPCs)
-	require.Error(t, exceeds(0, 1, svc.limitsFor(t.Context(), account).VPCs))
+	limits, err := svc.limitsFor(t.Context(), account)
+	require.NoError(t, err)
+	require.Equal(t, 0, limits.VPCs)
+	require.Error(t, exceeds(0, 1, limits.VPCs))
 }
 
 // A corrupt record is a read failure, not an absent one: it must fall back to
@@ -110,7 +119,9 @@ func TestLimitsForCorruptRecordFailsClosed(t *testing.T) {
 	_, err := kv.Put(t.Context(), account, []byte("{not json"))
 	require.NoError(t, err)
 
-	require.Equal(t, baseLimits(), svc.limitsFor(t.Context(), account),
+	limits, err := svc.limitsFor(t.Context(), account)
+	require.NoError(t, err, "a decode failure must fall back, not error, from limitsFor")
+	require.Equal(t, baseLimits(), limits,
 		"a corrupt override must fall back to the configured limits")
 
 	_, _, err = svc.GetAccountQuota(t.Context(), account)

spinifex/domains/admission/quota/override_test.go

@@ -126,7 +126,11 @@ func TestOverridesEmpty(t *testing.T) {
 // fall through to the configured limits rather than panicking.
 func TestLimitsForWithoutBucket(t *testing.T) {
 	svc := New(baseLimits(), nil)
-	if got := svc.limitsFor(t.Context(), "123456789012"); got != baseLimits() {
+	got, err := svc.limitsFor(t.Context(), "123456789012")
+	if err != nil {
+		t.Fatalf("limitsFor error = %v, want nil", err)
+	}
+	if got != baseLimits() {
 		t.Fatalf("limitsFor = %+v, want the configured limits", got)
 	}
 }

spinifex/domains/admission/quota/override_writes_guarded_test.go

@@ -0,0 +1,75 @@
+package quota
+
+//test:in-package — pins what AccountQuotaOverrideSealCapabilityVersion 1 will
+// assert: nothing in this package writes the override bucket except through
+// PutAccountQuota's CAS loop, clearAccountQuota's guarded delete or
+// SealAccountOverrides' CAS loop.
+
+import (
+	"go/ast"
+	"go/parser"
+	"go/token"
+	"path/filepath"
+	"runtime"
+	"testing"
+)
+
+// overrideWriteFiles are the only files in this package that touch the
+// override bucket. The package also holds the vCPU usage and Bedrock usage
+// buckets, which write through their own stores and are out of scope here;
+// scanning the whole package directory, as the EC2 settings equivalent does,
+// would false-flag those unrelated writes.
+var overrideWriteFiles = []string{"override.go", "seal.go"}
+
+func quotaPackageDir(t *testing.T) string {
+	t.Helper()
+	_, file, _, ok := runtime.Caller(0)
+	if !ok {
+		t.Fatal("cannot locate test source")
+	}
+	return filepath.Dir(file)
+}
+
+// No file owning the override bucket may call Put, PutString, or Delete
+// without a revision option: every write must go through PutAccountQuota's
+// CAS loop, clearAccountQuota's guarded delete, or SealAccountOverrides' CAS
+// loop.
+func TestOverrideWritesAreRevisionGuarded(t *testing.T) {
+	root := quotaPackageDir(t)
+	fset := token.NewFileSet()
+	scanned := 0
+
+	for _, name := range overrideWriteFiles {
+		path := filepath.Join(root, name)
+		f, err := parser.ParseFile(fset, path, nil, 0)
+		if err != nil {
+			t.Fatalf("parse %s: %v", path, err)
+		}
+		scanned++
+
+		ast.Inspect(f, func(n ast.Node) bool {
+			call, ok := n.(*ast.CallExpr)
+			if !ok {
+				return true
+			}
+			sel, ok := call.Fun.(*ast.SelectorExpr)
+			if !ok {
+				return true
+			}
+			switch sel.Sel.Name {
+			case "Put", "PutString":
+				t.Errorf("%s: unconditional %s call bypasses the revision-guarded override write; use PutAccountQuota or SealAccountOverrides instead",
+					fset.Position(call.Pos()), sel.Sel.Name)
+			case "Delete":
+				if len(call.Args) < 3 {
+					t.Errorf("%s: Delete call with no revision option bypasses the guarded clear; pin it to the read revision",
+						fset.Position(call.Pos()))
+				}
+			}
+			return true
+		})
+	}
+	if scanned == 0 {
+		t.Fatalf("scanned no override-write files under %s; the invariant would pass vacuously", root)
+	}
+}

spinifex/domains/admission/quota/quota.go

@@ -90,6 +90,14 @@ type Service struct {
 	// overrides holds the per-account limit overrides. Nil leaves every account
 	// on the configured limits, which is also what an absent key means.
 	overrides jetstream.KeyValue
+
+	// js opens the override bucket for SealAccountOverrides when quotas were
+	// disabled at startup; nil makes such a seal refuse rather than create it.
+	js jetstream.KeyValueManager
+
+	// terminationChecker is the seal precondition; SealAccountOverrides
+	// refuses until it is wired.
+	terminationChecker AccountTerminationChecker
 }
 
 // New constructs a quota Service from the configured limits and the gateway-owned
@@ -125,6 +133,16 @@ func (s *Service) SetOverrides(kv jetstream.KeyValue) {
 	s.overrides = kv
 }
 
+// SetJetStream wires the client SealAccountOverrides uses to get-or-create
+// the override bucket when this gateway disabled quotas. Called once at
+// wiring time, regardless of whether quotas are enabled here.
+func (s *Service) SetJetStream(js jetstream.KeyValueManager) {
+	if s == nil {
+		return
+	}
+	s.js = js
+}
+
 // Exempt returns true for the global/system account and whenever quotas are
 // disabled, so those callers bypass every quota check. A nil Service (quotas
 // never built) is treated as disabled so callers need no nil guard.

spinifex/domains/admission/quota/seal.go

@@ -0,0 +1,163 @@
+package quota
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"log/slog"
+
+	awserrors "github.com/mulgadc/spinifex/spinifex/foundation/aws/errors"
+	awsidentifiers "github.com/mulgadc/spinifex/spinifex/foundation/aws/identifiers"
+	"github.com/mulgadc/spinifex/spinifex/foundation/state/kvutil"
+	"github.com/nats-io/nats.go/jetstream"
+)
+
+// sealMaxAttempts bounds SealAccountOverrides' retry loop against a revision
+// that keeps moving under concurrent writes, so it fails closed instead of
+// spinning.
+const sealMaxAttempts = 5
+
+// sealedRecord is the fixed value a seal writes. No dimension or provenance
+// survives, so no tenant commercial setting outlives the seal.
+var sealedRecord = &overrideRecord{State: OverrideStateSealed}
+
+// OverrideSeal is SealAccountOverrides' result: the revision at which the
+// stream leader accepted the sealed value.
+type OverrideSeal struct {
+	Revision uint64
+}
+
+// AccountTerminationChecker reports whether IAM holds accountID as
+// TERMINATING. A missing record is (false, nil); only a failed read is an
+// error.
+type AccountTerminationChecker interface {
+	AccountTerminating(ctx context.Context, accountID string) (bool, error)
+}
+
+// SetAccountTerminationChecker wires the seal precondition; until it is set,
+// SealAccountOverrides refuses.
+func (s *Service) SetAccountTerminationChecker(c AccountTerminationChecker) {
+	if s == nil {
+		return
+	}
+	s.terminationChecker = c
+}
+
+// checkTerminating runs before any override-KV call, so a refusal never
+// writes or opens the bucket. Not TERMINATING (incl. missing) is
+// IncorrectState; a read failure or unwired checker is InternalError.
+func (s *Service) checkTerminating(ctx context.Context, accountID string) error {
+	if s.terminationChecker == nil {
+		slog.ErrorContext(ctx, "Quota override seal has no termination checker wired", "accountID", accountID)
+		return errors.New(awserrors.ErrorInternalError)
+	}
+	terminating, err := s.terminationChecker.AccountTerminating(ctx, accountID)
+	if err != nil {
+		slog.ErrorContext(ctx, "Failed to read account status for seal precondition", "accountID", accountID, "err", err)
+		return errors.New(awserrors.ErrorInternalError)
+	}
+	if !terminating {
+		return awserrors.Errorf(awserrors.ErrorIncorrectState, "account %s is not TERMINATING", accountID)
+	}
+	return nil
+}
+
+// isRevisionConflict reports a lost CAS race to retry, including the bare
+// wrong-last-sequence APIError from Create's retry against a delete marker.
+// A private copy of domains/ec2/account's helper, which this package may not import.
+func isRevisionConflict(err error) bool {
+	if errors.Is(err, jetstream.ErrKeyExists) || errors.Is(err, jetstream.ErrKeyRevisionMismatch) {
+		return true
+	}
+	if apiErr, ok := errors.AsType[*jetstream.APIError](err); ok {
+		return apiErr.ErrorCode == jetstream.JSErrCodeStreamWrongLastSequence ||
+			apiErr.ErrorCode == jetstream.JSErrCodeStreamWrongLastSequenceConstant
+	}
+	return false
+}
+
+// sealBucket returns this gateway's override bucket, or get-or-creates it
+// when quotas are disabled here. Called only after checkTerminating accepts,
+// so a refused seal never creates the bucket.
+func (s *Service) sealBucket(ctx context.Context) (jetstream.KeyValue, error) {
+	if s.overrides != nil {
+		return s.overrides, nil
+	}
+	if s.js == nil {
+		slog.ErrorContext(ctx, "Quota override seal has no JetStream client wired to open the override bucket")
+		return nil, errors.New(awserrors.ErrorInternalError)
+	}
+	kv, err := kvutil.GetOrCreateBucket(ctx, s.js, KVBucketAccountQuota, 1)
+	if err != nil {
+		slog.ErrorContext(ctx, "Failed to open quota override bucket for seal", "err", err)
+		return nil, errors.New(awserrors.ErrorInternalError)
+	}
+	return kv, nil
+}
+
+// SealAccountOverrides seals accountID's override record so no later writer
+// can recreate it. Success is only a leader-accepted write, evidenced by the
+// revision; a missing key is created sealed, never reported from absence.
+func (s *Service) SealAccountOverrides(ctx context.Context, accountID string) (*OverrideSeal, error) {
+	if accountID == "" {
+		return nil, awserrors.Errorf(awserrors.ErrorMissingParameter, "accountID is required to seal quota overrides")
+	}
+	if !awsidentifiers.IsAccountID(accountID) {
+		return nil, awserrors.Errorf(awserrors.ErrorInvalidParameterValue, "accountID %q is not a valid AWS account ID", accountID)
+	}
+	if accountID == awsidentifiers.GlobalAccountID {
+		return nil, awserrors.Errorf(awserrors.ErrorInvalidParameterValue, "quota override seal must not target the global account")
+	}
+
+	if err := s.checkTerminating(ctx, accountID); err != nil {
+		return nil, err
+	}
+
+	overrides, err := s.sealBucket(ctx)
+	if err != nil {
+		return nil, err
+	}
+
+	data, err := json.Marshal(sealedRecord)
+	if err != nil {
+		slog.ErrorContext(ctx, "Failed to encode sealed quota override record", "accountID", accountID, "err", err)
+		return nil, errors.New(awserrors.ErrorInternalError)
+	}
+
+	// accountID is never resolved through any fallback here: a seal must act
+	// on exactly accountID and never fall back to the global key.
+	for range sealMaxAttempts {
+		var rev uint64
+		entry, getErr := overrides.Get(ctx, accountID)
+		switch {
+		case getErr == nil:
+			rev = entry.Revision()
+		case errors.Is(getErr, jetstream.ErrKeyNotFound):
+			rev = 0
+		default:
+			slog.ErrorContext(ctx, "Failed to read quota override for seal", "accountID", accountID, "err", getErr)
+			return nil, errors.New(awserrors.ErrorInternalError)
+		}
+
+		var newRev uint64
+		var writeErr error
+		if rev == 0 {
+			newRev, writeErr = overrides.Create(ctx, accountID, data)
+		} else {
+			newRev, writeErr = overrides.Update(ctx, accountID, data, rev)
+		}
+
+		switch {
+		case writeErr == nil:
+			return &OverrideSeal{Revision: newRev}, nil
+		case isRevisionConflict(writeErr):
+			continue
+		default:
+			slog.ErrorContext(ctx, "Failed to write sealed quota override record", "accountID", accountID, "err", writeErr)
+			return nil, errors.New(awserrors.ErrorInternalError)
+		}
+	}
+
+	slog.ErrorContext(ctx, "Quota override seal exhausted attempts against concurrent writes", "accountID", accountID, "attempts", sealMaxAttempts)
+	return nil, errors.New(awserrors.ErrorInternalError)
+}

spinifex/domains/admission/quota/seal_test.go

@@ -0,0 +1,806 @@
+package quota
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"sync"
+	"testing"
+
+	"github.com/mulgadc/spinifex/internal/testkit"
+	awserrors "github.com/mulgadc/spinifex/spinifex/foundation/aws/errors"
+	awsidentifiers "github.com/mulgadc/spinifex/spinifex/foundation/aws/identifiers"
+	"github.com/nats-io/nats.go/jetstream"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+// stubTerminationChecker answers AccountTerminating from a fixed set, for
+// tests that exercise the KV paths rather than IAM itself.
+type stubTerminationChecker struct {
+	terminating map[string]bool
+	err         error
+}
+
+func (c *stubTerminationChecker) AccountTerminating(_ context.Context, accountID string) (bool, error) {
+	if c.err != nil {
+		return false, c.err
+	}
+	return c.terminating[accountID], nil
+}
+
+// terminatingChecker answers true for exactly accountID.
+func terminatingChecker(accountID string) *stubTerminationChecker {
+	return &stubTerminationChecker{terminating: map[string]bool{accountID: true}}
+}
+
+// newSealService builds a Service over a real JetStream KV bucket with the
+// override bucket already attached, mirroring a quota-enabled gateway.
+func newSealService(t *testing.T) (*Service, jetstream.JetStream) {
+	t.Helper()
+	_, _, js := testutil.StartTestJetStream(t)
+	kv, err := js.CreateKeyValue(t.Context(), jetstream.KeyValueConfig{
+		Bucket:  KVBucketAccountQuota,
+		History: 1,
+	})
+	require.NoError(t, err)
+	usage, err := js.CreateKeyValue(t.Context(), jetstream.KeyValueConfig{
+		Bucket:  KVBucketAccountUsage,
+		History: 1,
+	})
+	require.NoError(t, err)
+
+	svc := New(baseLimits(), usage)
+	svc.SetOverrides(kv)
+	svc.SetJetStream(js)
+	return svc, js
+}
+
+// newDisabledSealService builds a Service with only the JetStream client
+// wired, mirroring a quota-disabled gateway: the override bucket does not
+// exist until a seal creates it.
+func newDisabledSealService(t *testing.T) (*Service, jetstream.JetStream) {
+	t.Helper()
+	_, _, js := testutil.StartTestJetStream(t)
+	svc := New(baseLimits(), nil)
+	svc.SetJetStream(js)
+	return svc, js
+}
+
+// lastOverrideSeq reports the override stream's last sequence, so a test can
+// assert a failed attempt wrote nothing.
+func lastOverrideSeq(t *testing.T, js jetstream.JetStream) uint64 {
+	t.Helper()
+	stream, err := js.Stream(context.Background(), "KV_"+KVBucketAccountQuota)
+	require.NoError(t, err)
+	info, err := stream.Info(context.Background())
+	require.NoError(t, err)
+	return info.State.LastSeq
+}
+
+// mutationCase is one of PutAccountQuota's two mutations, called uniformly so
+// a test can table over them.
+type mutationCase struct {
+	name string
+	call func(svc *Service, ctx context.Context) error
+}
+
+func mutationTable() []mutationCase {
+	return []mutationCase{
+		{"Set", func(svc *Service, ctx context.Context) error {
+			return svc.PutAccountQuota(ctx, testAccount, Overrides{VCPUs: ptr(32)}, "operator")
+		}},
+		{"Clear", func(svc *Service, ctx context.Context) error {
+			return svc.PutAccountQuota(ctx, testAccount, Overrides{}, "operator")
+		}},
+	}
+}
+
+// seedOverrideState prepares testAccount's record for a race test: present
+// (an active record), absent (never written) or tombstone (written then
+// deleted, leaving a delete marker Get hides).
+func seedOverrideState(t *testing.T, svc *Service, state string) {
+	t.Helper()
+	ctx := context.Background()
+	switch state {
+	case "present":
+		require.NoError(t, svc.PutAccountQuota(ctx, testAccount, Overrides{VCPUs: ptr(8)}, "operator"))
+	case "tombstone":
+		require.NoError(t, svc.PutAccountQuota(ctx, testAccount, Overrides{VCPUs: ptr(8)}, "operator"))
+		require.NoError(t, svc.overrides.Delete(ctx, testAccount))
+	case "absent":
+	default:
+		t.Fatalf("unknown seed state %q", state)
+	}
+}
+
+// sealRaceKV wraps a real KeyValue so a test can inject a deterministic race
+// between a caller's read and its guarded write (Create, Update or Delete),
+// instead of relying on goroutine timing.
+type sealRaceKV struct {
+	jetstream.KeyValue
+
+	writeCalls  int
+	beforeWrite func(attempt int) error
+}
+
+func (f *sealRaceKV) Create(ctx context.Context, key string, value []byte, opts ...jetstream.KVCreateOpt) (uint64, error) {
+	f.writeCalls++
+	if f.beforeWrite != nil {
+		if err := f.beforeWrite(f.writeCalls); err != nil {
+			return 0, err
+		}
+	}
+	return f.KeyValue.Create(ctx, key, value, opts...)
+}
+
+func (f *sealRaceKV) Update(ctx context.Context, key string, value []byte, revision uint64) (uint64, error) {
+	f.writeCalls++
+	if f.beforeWrite != nil {
+		if err := f.beforeWrite(f.writeCalls); err != nil {
+			return 0, err
+		}
+	}
+	return f.KeyValue.Update(ctx, key, value, revision)
+}
+
+func (f *sealRaceKV) Delete(ctx context.Context, key string, opts ...jetstream.KVDeleteOpt) error {
+	f.writeCalls++
+	if f.beforeWrite != nil {
+		if err := f.beforeWrite(f.writeCalls); err != nil {
+			return err
+		}
+	}
+	return f.KeyValue.Delete(ctx, key, opts...)
+}
+
+// staleGetKV always answers Get with a fixed snapshot taken at construction,
+// so a caller's CAS keeps retrying against a revision that can never win.
+type staleGetKV struct {
+	jetstream.KeyValue
+
+	entry jetstream.KeyValueEntry
+}
+
+func (f *staleGetKV) Get(context.Context, string) (jetstream.KeyValueEntry, error) {
+	return f.entry, nil
+}
+
+// staleNotFoundKV always answers Get as absent, regardless of what the real
+// bucket holds.
+type staleNotFoundKV struct {
+	jetstream.KeyValue
+}
+
+func (f *staleNotFoundKV) Get(context.Context, string) (jetstream.KeyValueEntry, error) {
+	return nil, jetstream.ErrKeyNotFound
+}
+
+// errGetKV always fails Get with a fixed store error.
+type errGetKV struct {
+	jetstream.KeyValue
+
+	err error
+}
+
+func (f *errGetKV) Get(context.Context, string) (jetstream.KeyValueEntry, error) {
+	return nil, f.err
+}
+
+func decodeOverride(t *testing.T, entry jetstream.KeyValueEntry) overrideRecord {
+	t.Helper()
+	var rec overrideRecord
+	require.NoError(t, json.Unmarshal(entry.Value(), &rec))
+	return rec
+}
+
+// A seal landing between a mutation's read and its guarded write is not
+// outlived: the mutation's retry sees the seal and is rejected, and nothing
+// it wrote survives. Clear on an absent or tombstoned key never reaches a
+// write at all (see TestClearOnMissingKeyWritesNothing), so only the four
+// cells with a real write are exercised here.
+func TestSealRaceMutation_SealLandsBetweenReadAndWrite(t *testing.T) {
+	cases := []struct {
+		name  string
+		state string
+		call  func(svc *Service, ctx context.Context) error
+	}{
+		{"Set_present", "present", mutationTable()[0].call},
+		{"Set_absent", "absent", mutationTable()[0].call},
+		{"Set_tombstone", "tombstone", mutationTable()[0].call},
+		{"Clear_present", "present", mutationTable()[1].call},
+	}
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			svc, _ := newSealService(t)
+			ctx := context.Background()
+			seedOverrideState(t, svc, tc.state)
+
+			realKV := svc.overrides
+			sealer := New(baseLimits(), nil)
+			sealer.SetOverrides(realKV)
+			sealer.SetAccountTerminationChecker(terminatingChecker(testAccount))
+
+			var sealed *OverrideSeal
+			race := &sealRaceKV{KeyValue: realKV}
+			race.beforeWrite = func(attempt int) error {
+				if attempt != 1 {
+					return nil
+				}
+				var sealErr error
+				sealed, sealErr = sealer.SealAccountOverrides(ctx, testAccount)
+				return sealErr
+			}
+			svc.overrides = race
+
+			err := tc.call(svc, ctx)
+			require.Error(t, err)
+			assert.True(t, awserrors.IsErrorCode(err, awserrors.ErrorBlocked), "got %v", err)
+			require.NotNil(t, sealed)
+
+			entry, err := realKV.Get(ctx, testAccount)
+			require.NoError(t, err)
+			assert.Equal(t, OverrideStateSealed, decodeOverride(t, entry).State)
+			assert.Equal(t, sealed.Revision, entry.Revision(), "no write landed after the seal")
+		})
+	}
+}
+
+// A mutation landing inside a seal's read-write window does not survive it:
+// the seal retries against the mutation's write and wins, leaving the sealed
+// value with no overridden dimension.
+func TestSealRaceMutation_MutationLandsBetweenSealReadAndWrite(t *testing.T) {
+	for _, m := range mutationTable() {
+		for _, state := range []string{"present", "absent", "tombstone"} {
+			t.Run(m.name+"_"+state, func(t *testing.T) {
+				svc, _ := newSealService(t)
+				ctx := context.Background()
+				seedOverrideState(t, svc, state)
+
+				realKV := svc.overrides
+				race := &sealRaceKV{KeyValue: realKV}
+				race.beforeWrite = func(attempt int) error {
+					if attempt != 1 {
+						return nil
+					}
+					_ = m.call(svc, ctx)
+					return nil
+				}
+
+				sealer := New(baseLimits(), nil)
+				sealer.SetOverrides(race)
+				sealer.SetAccountTerminationChecker(terminatingChecker(testAccount))
+
+				seal, err := sealer.SealAccountOverrides(ctx, testAccount)
+				require.NoError(t, err)
+
+				entry, err := realKV.Get(ctx, testAccount)
+				require.NoError(t, err)
+				rec := decodeOverride(t, entry)
+				assert.Equal(t, OverrideStateSealed, rec.State)
+				assert.Nil(t, rec.VCPUs)
+				assert.Equal(t, seal.Revision, entry.Revision())
+			})
+		}
+	}
+}
+
+// Neither mutation can act on an already-sealed record.
+func TestMutationOnSealedRecordIsRejected(t *testing.T) {
+	for _, m := range mutationTable() {
+		t.Run(m.name, func(t *testing.T) {
+			svc, _ := newSealService(t)
+			ctx := context.Background()
+			svc.SetAccountTerminationChecker(terminatingChecker(testAccount))
+			_, err := svc.SealAccountOverrides(ctx, testAccount)
+			require.NoError(t, err)
+
+			before, err := svc.overrides.Get(ctx, testAccount)
+			require.NoError(t, err)
+
+			err = m.call(svc, ctx)
+			require.Error(t, err)
+			assert.True(t, awserrors.IsErrorCode(err, awserrors.ErrorBlocked), "got %v", err)
+
+			after, err := svc.overrides.Get(ctx, testAccount)
+			require.NoError(t, err)
+			assert.Equal(t, before.Revision(), after.Revision())
+		})
+	}
+}
+
+// Every override mutation races one seal on a real bucket; afterwards the
+// record is sealed and stays sealed, so no success among them could have
+// landed after the seal.
+func TestSealConcurrentMutations(t *testing.T) {
+	svc, _ := newSealService(t)
+	ctx := context.Background()
+	svc.SetAccountTerminationChecker(terminatingChecker(testAccount))
+	require.NoError(t, svc.PutAccountQuota(ctx, testAccount, Overrides{VCPUs: ptr(8)}, "operator"))
+
+	mutations := mutationTable()
+	errs := make([]error, len(mutations))
+	var wg sync.WaitGroup
+	for i, m := range mutations {
+		wg.Add(1)
+		go func(i int, m mutationCase) {
+			defer wg.Done()
+			errs[i] = m.call(svc, ctx)
+		}(i, m)
+	}
+
+	var sealResult *OverrideSeal
+	var sealErr error
+	wg.Go(func() {
+		sealResult, sealErr = svc.SealAccountOverrides(ctx, testAccount)
+	})
+	wg.Wait()
+
+	require.NoError(t, sealErr)
+	require.NotNil(t, sealResult)
+	for i, err := range errs {
+		if err != nil {
+			assert.True(t, awserrors.IsErrorCode(err, awserrors.ErrorBlocked), "mutation %s: got %v", mutations[i].name, err)
+		}
+	}
+
+	entry, err := svc.overrides.Get(ctx, testAccount)
+	require.NoError(t, err)
+	assert.Equal(t, OverrideStateSealed, decodeOverride(t, entry).State)
+	assert.Equal(t, sealResult.Revision, entry.Revision())
+
+	_, _, err = svc.GetAccountQuota(ctx, testAccount)
+	assert.True(t, awserrors.IsErrorCode(err, awserrors.ErrorBlocked))
+}
+
+// A Get pinned to a revision the record has already moved past exhausts the
+// mutation's retries without ever writing.
+func TestMutationStaleReadExhaustsWithoutWriting(t *testing.T) {
+	svc, js := newSealService(t)
+	ctx := context.Background()
+
+	require.NoError(t, svc.PutAccountQuota(ctx, testAccount, Overrides{VCPUs: ptr(8)}, "operator"))
+	stale, err := svc.overrides.Get(ctx, testAccount)
+	require.NoError(t, err)
+
+	require.NoError(t, svc.PutAccountQuota(ctx, testAccount, Overrides{VCPUs: ptr(16)}, "operator"))
+
+	before := lastOverrideSeq(t, js)
+	svc.overrides = &staleGetKV{KeyValue: svc.overrides, entry: stale}
+
+	err = svc.PutAccountQuota(ctx, testAccount, Overrides{VCPUs: ptr(32)}, "operator")
+	assert.True(t, awserrors.IsErrorCode(err, awserrors.ErrorInternalError), "got %v", err)
+	assert.Equal(t, before, lastOverrideSeq(t, js))
+}
+
+// A Get that always reports absent, while a sealed record actually exists,
+// never gets to recreate it: every Create attempt conflicts.
+func TestMutationStaleAbsentDoesNotRecreate(t *testing.T) {
+	svc, js := newSealService(t)
+	ctx := context.Background()
+	svc.SetAccountTerminationChecker(terminatingChecker(testAccount))
+	_, err := svc.SealAccountOverrides(ctx, testAccount)
+	require.NoError(t, err)
+
+	before := lastOverrideSeq(t, js)
+	svc.overrides = &staleNotFoundKV{KeyValue: svc.overrides}
+
+	err = svc.PutAccountQuota(ctx, testAccount, Overrides{VCPUs: ptr(8)}, "operator")
+	assert.True(t, awserrors.IsErrorCode(err, awserrors.ErrorInternalError), "got %v", err)
+	assert.Equal(t, before, lastOverrideSeq(t, js))
+}
+
+// A revision that keeps moving under sustained concurrent writes must not let
+// the seal spin forever: it gives up, bounded by its attempt budget, and its
+// own write never lands.
+func TestSealSustainedConflictFailsWithoutWriting(t *testing.T) {
+	svc, _ := newSealService(t)
+	ctx := context.Background()
+	svc.SetAccountTerminationChecker(terminatingChecker(testAccount))
+	require.NoError(t, svc.PutAccountQuota(ctx, testAccount, Overrides{VCPUs: ptr(8)}, "operator"))
+
+	realKV := svc.overrides
+	race := &sealRaceKV{KeyValue: realKV}
+	race.beforeWrite = func(int) error {
+		data, merr := json.Marshal(&overrideRecord{Overrides: Overrides{VCPUs: ptr(16)}})
+		require.NoError(t, merr)
+		_, perr := realKV.Put(ctx, testAccount, data)
+		return perr
+	}
+	svc.overrides = race
+
+	_, err := svc.SealAccountOverrides(ctx, testAccount)
+	assert.True(t, awserrors.IsErrorCode(err, awserrors.ErrorInternalError), "got %v", err)
+	assert.Equal(t, sealMaxAttempts, race.writeCalls)
+
+	entry, err := realKV.Get(ctx, testAccount)
+	require.NoError(t, err)
+	assert.NotEqual(t, OverrideStateSealed, decodeOverride(t, entry).State, "the seal's own write never landed")
+}
+
+// A store error on the seal's read reports failed, never written.
+func TestSealStoreErrorWritesNothing(t *testing.T) {
+	svc, js := newSealService(t)
+	ctx := context.Background()
+	svc.SetAccountTerminationChecker(terminatingChecker(testAccount))
+
+	before := lastOverrideSeq(t, js)
+	svc.overrides = &errGetKV{KeyValue: svc.overrides, err: errors.New("kv unavailable")}
+
+	_, err := svc.SealAccountOverrides(ctx, testAccount)
+	assert.True(t, awserrors.IsErrorCode(err, awserrors.ErrorInternalError), "got %v", err)
+	assert.Equal(t, before, lastOverrideSeq(t, js))
+}
+
+func TestIsRevisionConflictRecognisesTombstoneRetry(t *testing.T) {
+	assert.True(t, isRevisionConflict(jetstream.ErrKeyExists))
+	assert.True(t, isRevisionConflict(jetstream.ErrKeyRevisionMismatch))
+	assert.True(t, isRevisionConflict(&jetstream.APIError{ErrorCode: jetstream.JSErrCodeStreamWrongLastSequence}))
+	assert.True(t, isRevisionConflict(&jetstream.APIError{ErrorCode: jetstream.JSErrCodeStreamWrongLastSequenceConstant}))
+	assert.True(t, isRevisionConflict(fmt.Errorf("wrapped: %w", &jetstream.APIError{ErrorCode: jetstream.JSErrCodeStreamWrongLastSequence})))
+	assert.False(t, isRevisionConflict(errors.New("unrelated failure")))
+	assert.False(t, isRevisionConflict(nil))
+}
+
+// Clearing never writes when there is nothing active to clear: an absent key,
+// a tombstoned key, and a sealed key hidden behind a stale not-found Get all
+// leave the stream exactly where they found it.
+func TestClearOnMissingKeyWritesNothing(t *testing.T) {
+	t.Run("absent", func(t *testing.T) {
+		svc, js := newSealService(t)
+		ctx := context.Background()
+		before := lastOverrideSeq(t, js)
+
+		require.NoError(t, svc.PutAccountQuota(ctx, testAccount, Overrides{}, "operator"))
+		assert.Equal(t, before, lastOverrideSeq(t, js))
+	})
+
+	t.Run("tombstone", func(t *testing.T) {
+		svc, js := newSealService(t)
+		ctx := context.Background()
+		require.NoError(t, svc.PutAccountQuota(ctx, testAccount, Overrides{VCPUs: ptr(8)}, "operator"))
+		require.NoError(t, svc.overrides.Delete(ctx, testAccount))
+		before := lastOverrideSeq(t, js)
+
+		require.NoError(t, svc.PutAccountQuota(ctx, testAccount, Overrides{}, "operator"))
+		assert.Equal(t, before, lastOverrideSeq(t, js))
+	})
+
+	t.Run("sealed_hidden_by_stale_not_found", func(t *testing.T) {
+		svc, js := newSealService(t)
+		ctx := context.Background()
+		svc.SetAccountTerminationChecker(terminatingChecker(testAccount))
+		sealed, err := svc.SealAccountOverrides(ctx, testAccount)
+		require.NoError(t, err)
+		before := lastOverrideSeq(t, js)
+
+		realKV := svc.overrides
+		svc.overrides = &staleNotFoundKV{KeyValue: realKV}
+
+		require.NoError(t, svc.PutAccountQuota(ctx, testAccount, Overrides{}, "operator"))
+		assert.Equal(t, before, lastOverrideSeq(t, js))
+
+		entry, err := realKV.Get(ctx, testAccount)
+		require.NoError(t, err)
+		rec := decodeOverride(t, entry)
+		assert.Equal(t, OverrideStateSealed, rec.State)
+		assert.Equal(t, sealed.Revision, entry.Revision())
+	})
+}
+
+// A seal against an absent or a tombstoned key creates the sealed record.
+func TestSealOnMissingKeyCreatesSealedRecord(t *testing.T) {
+	for _, state := range []string{"absent", "tombstone"} {
+		t.Run(state, func(t *testing.T) {
+			svc, _ := newSealService(t)
+			ctx := context.Background()
+			svc.SetAccountTerminationChecker(terminatingChecker(testAccount))
+			if state == "tombstone" {
+				require.NoError(t, svc.PutAccountQuota(ctx, testAccount, Overrides{VCPUs: ptr(8)}, "operator"))
+				require.NoError(t, svc.overrides.Delete(ctx, testAccount))
+			}
+			_, err := svc.overrides.Get(ctx, testAccount)
+			require.ErrorIs(t, err, jetstream.ErrKeyNotFound)
+
+			seal, err := svc.SealAccountOverrides(ctx, testAccount)
+			require.NoError(t, err)
+
+			entry, err := svc.overrides.Get(ctx, testAccount)
+			require.NoError(t, err)
+			assert.Equal(t, seal.Revision, entry.Revision())
+			assert.Equal(t, OverrideStateSealed, decodeOverride(t, entry).State)
+		})
+	}
+}
+
+// alwaysConflictThenErrorKV answers Get as absent and every Create attempt
+// with a conflict, then a store error on the final attempt: a recording
+// wrapper proving the seal tried a real write on every attempt before giving
+// up, rather than ever reporting success from an observed absence.
+type alwaysConflictThenErrorKV struct {
+	jetstream.KeyValue
+
+	calls int
+}
+
+func (f *alwaysConflictThenErrorKV) Get(context.Context, string) (jetstream.KeyValueEntry, error) {
+	return nil, jetstream.ErrKeyNotFound
+}
+
+func (f *alwaysConflictThenErrorKV) Create(context.Context, string, []byte, ...jetstream.KVCreateOpt) (uint64, error) {
+	f.calls++
+	if f.calls < sealMaxAttempts {
+		return 0, jetstream.ErrKeyExists
+	}
+	return 0, errors.New("store unavailable")
+}
+
+func TestSealNeverSucceedsWithoutAWrite(t *testing.T) {
+	svc, _ := newSealService(t)
+	ctx := context.Background()
+	svc.SetAccountTerminationChecker(terminatingChecker(testAccount))
+	fake := &alwaysConflictThenErrorKV{}
+	svc.overrides = fake
+
+	seal, err := svc.SealAccountOverrides(ctx, testAccount)
+	assert.Error(t, err)
+	assert.Nil(t, seal)
+	assert.Equal(t, sealMaxAttempts, fake.calls, "every attempt must have tried a real write before giving up")
+}
+
+// An account IAM does not hold as TERMINATING refuses the seal before any
+// override-KV write, whether the key is present or absent.
+func TestSealRefusedForActiveAccount(t *testing.T) {
+	for _, state := range []string{"present", "absent"} {
+		t.Run(state, func(t *testing.T) {
+			svc, js := newSealService(t)
+			ctx := context.Background()
+			if state == "present" {
+				require.NoError(t, svc.PutAccountQuota(ctx, testAccount, Overrides{VCPUs: ptr(8)}, "operator"))
+			}
+			svc.SetAccountTerminationChecker(&stubTerminationChecker{terminating: map[string]bool{}})
+
+			before := lastOverrideSeq(t, js)
+			_, err := svc.SealAccountOverrides(ctx, testAccount)
+			assert.True(t, awserrors.IsErrorCode(err, awserrors.ErrorIncorrectState), "got %v", err)
+			assert.Equal(t, before, lastOverrideSeq(t, js))
+		})
+	}
+}
+
+func TestSealRefusedForMissingIAMRecord(t *testing.T) {
+	svc, js := newSealService(t)
+	ctx := context.Background()
+	svc.SetAccountTerminationChecker(&stubTerminationChecker{terminating: map[string]bool{}})
+
+	before := lastOverrideSeq(t, js)
+	_, err := svc.SealAccountOverrides(ctx, testAccount)
+	assert.True(t, awserrors.IsErrorCode(err, awserrors.ErrorIncorrectState), "got %v", err)
+	assert.Equal(t, before, lastOverrideSeq(t, js))
+}
+
+func TestSealRefusedWhenIAMReadFails(t *testing.T) {
+	svc, js := newSealService(t)
+	ctx := context.Background()
+	svc.SetAccountTerminationChecker(&stubTerminationChecker{err: errors.New("iam unavailable")})
+
+	before := lastOverrideSeq(t, js)
+	_, err := svc.SealAccountOverrides(ctx, testAccount)
+	assert.True(t, awserrors.IsErrorCode(err, awserrors.ErrorInternalError), "got %v", err)
+	assert.Equal(t, before, lastOverrideSeq(t, js))
+}
+
+func TestSealRefusedWithNoCheckerWired(t *testing.T) {
+	svc, js := newSealService(t)
+	ctx := context.Background()
+
+	before := lastOverrideSeq(t, js)
+	_, err := svc.SealAccountOverrides(ctx, testAccount)
+	assert.True(t, awserrors.IsErrorCode(err, awserrors.ErrorInternalError), "got %v", err)
+	assert.Equal(t, before, lastOverrideSeq(t, js))
+}
+
+func TestSealAcceptedForTerminatingAccount(t *testing.T) {
+	svc, _ := newSealService(t)
+	ctx := context.Background()
+	svc.SetAccountTerminationChecker(terminatingChecker(testAccount))
+
+	seal, err := svc.SealAccountOverrides(ctx, testAccount)
+	require.NoError(t, err)
+	assert.NotZero(t, seal.Revision)
+
+	entry, err := svc.overrides.Get(ctx, testAccount)
+	require.NoError(t, err)
+	assert.Equal(t, seal.Revision, entry.Revision())
+}
+
+// A refused seal, on a gateway that disabled quotas and so holds no override
+// bucket, never creates one.
+func TestSealRefusalDoesNotCreateBucket(t *testing.T) {
+	t.Run("not_terminating", func(t *testing.T) {
+		svc, js := newDisabledSealService(t)
+		ctx := context.Background()
+		svc.SetAccountTerminationChecker(&stubTerminationChecker{terminating: map[string]bool{}})
+
+		_, err := svc.SealAccountOverrides(ctx, testAccount)
+		assert.True(t, awserrors.IsErrorCode(err, awserrors.ErrorIncorrectState), "got %v", err)
+		_, err = js.KeyValue(context.Background(), KVBucketAccountQuota)
+		assert.ErrorIs(t, err, jetstream.ErrBucketNotFound)
+	})
+
+	t.Run("iam_read_fails", func(t *testing.T) {
+		svc, js := newDisabledSealService(t)
+		ctx := context.Background()
+		svc.SetAccountTerminationChecker(&stubTerminationChecker{err: errors.New("iam unavailable")})
+
+		_, err := svc.SealAccountOverrides(ctx, testAccount)
+		assert.True(t, awserrors.IsErrorCode(err, awserrors.ErrorInternalError), "got %v", err)
+		_, err = js.KeyValue(context.Background(), KVBucketAccountQuota)
+		assert.ErrorIs(t, err, jetstream.ErrBucketNotFound)
+	})
+}
+
+// A seal accepted on a gateway whose quotas were disabled at startup creates
+// the override bucket at seal time, never at startup.
+func TestSealOnQuotaDisabledGatewayCreatesBucket(t *testing.T) {
+	svc, js := newDisabledSealService(t)
+	ctx := context.Background()
+	svc.SetAccountTerminationChecker(terminatingChecker(testAccount))
+
+	seal, err := svc.SealAccountOverrides(ctx, testAccount)
+	require.NoError(t, err)
+
+	kv, err := js.KeyValue(context.Background(), KVBucketAccountQuota)
+	require.NoError(t, err)
+	entry, err := kv.Get(context.Background(), testAccount)
+	require.NoError(t, err)
+	assert.Equal(t, seal.Revision, entry.Revision())
+	assert.Equal(t, OverrideStateSealed, decodeOverride(t, entry).State)
+}
+
+// A sealed record stays sealed after the service is rebuilt over the same
+// JetStream, every read and write refuses it, and a re-seal gets a higher
+// revision.
+func TestSealSurvivesServiceReconstruction(t *testing.T) {
+	_, _, js := testutil.StartTestJetStream(t)
+	ctx := context.Background()
+
+	kv, err := js.CreateKeyValue(ctx, jetstream.KeyValueConfig{Bucket: KVBucketAccountQuota, History: 1})
+	require.NoError(t, err)
+
+	svc := New(baseLimits(), nil)
+	svc.SetOverrides(kv)
+	svc.SetJetStream(js)
+	svc.SetAccountTerminationChecker(terminatingChecker(testAccount))
+	first, err := svc.SealAccountOverrides(ctx, testAccount)
+	require.NoError(t, err)
+
+	rebuiltKV, err := js.KeyValue(ctx, KVBucketAccountQuota)
+	require.NoError(t, err)
+	rebuilt := New(baseLimits(), nil)
+	rebuilt.SetOverrides(rebuiltKV)
+	rebuilt.SetJetStream(js)
+
+	entry, err := rebuilt.overrides.Get(ctx, testAccount)
+	require.NoError(t, err)
+	assert.Equal(t, OverrideStateSealed, decodeOverride(t, entry).State)
+
+	_, _, err = rebuilt.GetAccountQuota(ctx, testAccount)
+	assert.True(t, awserrors.IsErrorCode(err, awserrors.ErrorBlocked))
+	assert.True(t, awserrors.IsErrorCode(rebuilt.PutAccountQuota(ctx, testAccount, Overrides{VCPUs: ptr(8)}, "operator"), awserrors.ErrorBlocked))
+	assert.True(t, awserrors.IsErrorCode(rebuilt.PutAccountQuota(ctx, testAccount, Overrides{}, "operator"), awserrors.ErrorBlocked))
+
+	limits, err := rebuilt.limitsFor(ctx, testAccount)
+	assert.True(t, awserrors.IsErrorCode(err, awserrors.ErrorBlocked))
+	assert.Zero(t, limits)
+
+	rebuilt.SetAccountTerminationChecker(terminatingChecker(testAccount))
+	second, err := rebuilt.SealAccountOverrides(ctx, testAccount)
+	require.NoError(t, err)
+	assert.Greater(t, second.Revision, first.Revision)
+}
+
+func TestSealRefusesEmptyInvalidAndGlobalAccount(t *testing.T) {
+	svc, _ := newSealService(t)
+	ctx := context.Background()
+
+	_, err := svc.SealAccountOverrides(ctx, "")
+	assert.True(t, awserrors.IsErrorCode(err, awserrors.ErrorMissingParameter), "got %v", err)
+
+	_, err = svc.SealAccountOverrides(ctx, "not-an-account")
+	assert.True(t, awserrors.IsErrorCode(err, awserrors.ErrorInvalidParameterValue), "got %v", err)
+
+	_, err = svc.SealAccountOverrides(ctx, awsidentifiers.GlobalAccountID)
+	assert.True(t, awserrors.IsErrorCode(err, awserrors.ErrorInvalidParameterValue), "got %v", err)
+}
+
+func TestSealLeavesOtherAccountsUntouched(t *testing.T) {
+	svc, _ := newSealService(t)
+	ctx := context.Background()
+	accountB := "222222222222"
+	require.NoError(t, svc.PutAccountQuota(ctx, accountB, Overrides{VCPUs: ptr(8)}, "operator"))
+
+	svc.SetAccountTerminationChecker(terminatingChecker(testAccount))
+	_, err := svc.SealAccountOverrides(ctx, testAccount)
+	require.NoError(t, err)
+
+	over, _, err := svc.GetAccountQuota(ctx, accountB)
+	require.NoError(t, err)
+	require.NotNil(t, over.VCPUs)
+	assert.Equal(t, 8, *over.VCPUs)
+}
+
+// A record written before the state field existed has none and is treated as
+// active by both reads and writes; a write on it preserves the old shape.
+func TestRecordWithoutStateIsActive(t *testing.T) {
+	svc, _ := newSealService(t)
+	ctx := context.Background()
+
+	data, err := json.Marshal(Overrides{VCPUs: ptr(8)})
+	require.NoError(t, err)
+	_, err = svc.overrides.Put(ctx, testAccount, data)
+	require.NoError(t, err)
+
+	over, limits, err := svc.GetAccountQuota(ctx, testAccount)
+	require.NoError(t, err)
+	require.NotNil(t, over.VCPUs)
+	assert.Equal(t, 8, limits.VCPUs)
+
+	require.NoError(t, svc.PutAccountQuota(ctx, testAccount, Overrides{VCPUs: ptr(16)}, "operator"))
+	entry, err := svc.overrides.Get(ctx, testAccount)
+	require.NoError(t, err)
+	assert.NotContains(t, string(entry.Value()), `"state"`)
+}
+
+// A record whose state is neither active nor sealed fails every read and
+// write closed, rather than being treated as active.
+func TestUnknownStateFailsClosed(t *testing.T) {
+	svc, _ := newSealService(t)
+	ctx := context.Background()
+
+	_, err := svc.overrides.Put(ctx, testAccount, []byte(`{"state":"frozen"}`))
+	require.NoError(t, err)
+
+	_, _, err = svc.GetAccountQuota(ctx, testAccount)
+	assert.True(t, awserrors.IsErrorCode(err, awserrors.ErrorInternalError), "got %v", err)
+
+	limits, err := svc.limitsFor(ctx, testAccount)
+	assert.True(t, awserrors.IsErrorCode(err, awserrors.ErrorInternalError), "got %v", err)
+	assert.Zero(t, limits)
+
+	assert.True(t, awserrors.IsErrorCode(svc.PutAccountQuota(ctx, testAccount, Overrides{VCPUs: ptr(8)}, "operator"), awserrors.ErrorInternalError))
+	assert.True(t, awserrors.IsErrorCode(svc.PutAccountQuota(ctx, testAccount, Overrides{}, "operator"), awserrors.ErrorInternalError))
+}
+
+// A sealed record is Blocked for admission's own resolver, which every
+// caller (admin read, ELBv2 DescribeAccountLimits, every Enforce* method)
+// shares.
+func TestAdmissionOnSealedRecordIsBlocked(t *testing.T) {
+	svc, _ := newSealService(t)
+	ctx := context.Background()
+	svc.SetAccountTerminationChecker(terminatingChecker(testAccount))
+	_, err := svc.SealAccountOverrides(ctx, testAccount)
+	require.NoError(t, err)
+
+	limits, err := svc.limitsFor(ctx, testAccount)
+	assert.True(t, awserrors.IsErrorCode(err, awserrors.ErrorBlocked), "got %v", err)
+	assert.Zero(t, limits)
+
+	assert.Error(t, svc.CheckVCPU(ctx, testAccount, 1))
+	assert.True(t, awserrors.IsErrorCode(svc.CheckVCPU(ctx, testAccount, 1), awserrors.ErrorBlocked))
+}
+
+// The existing read-failure fallback in limitsFor (a decode or store error
+// falls back to the configured limits) is unchanged by the seal work: it is
+// distinct from a successfully read sealed record, which refuses instead.
+func TestAdmissionReadFailureFallbackUnchanged(t *testing.T) {
+	svc, kv := newOverrideService(t)
+
+	_, err := kv.Put(t.Context(), testAccount, []byte("{not json"))
+	require.NoError(t, err)
+
+	limits, err := svc.limitsFor(t.Context(), testAccount)
+	require.NoError(t, err, "a decode failure must still fall back, not error, from limitsFor")
+	require.Equal(t, baseLimits(), limits)
+}

spinifex/domains/admission/quota/storage.go

@@ -23,7 +23,10 @@ func (s *Service) EnforceVolumeCreate(ctx context.Context, natsConn *nats.Conn,
 	if err != nil {
 		return err
 	}
-	limits := s.limitsFor(ctx, accountID)
+	limits, err := s.limitsFor(ctx, accountID)
+	if err != nil {
+		return err
+	}
 	if err := exceeds(len(out.Volumes), 1, limits.Volumes); err != nil {
 		return err
 	}
@@ -43,7 +46,11 @@ func (s *Service) EnforceVolumeModify(ctx context.Context, natsConn *nats.Conn,
 	if err != nil {
 		return err
 	}
-	return exceeds(total-oldGiB, newGiB, s.limitsFor(ctx, accountID).VolumesGiB)
+	limits, err := s.limitsFor(ctx, accountID)
+	if err != nil {
+		return err
+	}
+	return exceeds(total-oldGiB, newGiB, limits.VolumesGiB)
 }
 
 // volumeUsage sums the account's volume sizes in GiB via a single DescribeVolumes

spinifex/domains/admission/quota/vcpu.go

@@ -37,7 +37,11 @@ func (s *Service) CheckVCPU(ctx context.Context, accountID string, want int) err
 	if err != nil {
 		return err
 	}
-	return exceeds(current, want, s.limitsFor(ctx, accountID).VCPUs)
+	limits, err := s.limitsFor(ctx, accountID)
+	if err != nil {
+		return err
+	}
+	return exceeds(current, want, limits.VCPUs)
 }
 
 // casVCPU applies next to accountID's counter under bounded JetStream CAS so

spinifex/domains/ec2/account/seal.go

@@ -0,0 +1,135 @@
+package account
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"log/slog"
+
+	awserrors "github.com/mulgadc/spinifex/spinifex/foundation/aws/errors"
+	awsidentifiers "github.com/mulgadc/spinifex/spinifex/foundation/aws/identifiers"
+	"github.com/nats-io/nats.go/jetstream"
+)
+
+// sealMaxAttempts bounds SealAccount's retry loop against a revision that
+// keeps moving under concurrent writes, so it fails closed instead of
+// spinning.
+const sealMaxAttempts = 5
+
+// sealedRecord is the fixed value a seal writes. Clearing both booleans
+// retains no tenant setting; an older reader that ignores state sees the
+// AWS defaults.
+var sealedRecord = &AccountSettingsRecord{State: AccountSettingsStateSealed}
+
+// AccountSettingsSeal is SealAccount's result: the revision at which the
+// stream leader accepted the sealed value.
+type AccountSettingsSeal struct {
+	Revision uint64
+}
+
+// AccountTerminationChecker reports whether IAM holds accountID as TERMINATING.
+// A missing record is (false, nil); only a failed read is an error.
+type AccountTerminationChecker interface {
+	AccountTerminating(ctx context.Context, accountID string) (bool, error)
+}
+
+// SetAccountTerminationChecker wires the precondition; until it is set, SealAccount refuses.
+func (s *AccountSettingsServiceImpl) SetAccountTerminationChecker(c AccountTerminationChecker) {
+	s.terminationChecker = c
+}
+
+// checkTerminating runs before any settings-KV call, so a refusal never writes.
+// Not TERMINATING, including a missing IAM record, is IncorrectState; a read
+// failure or an unwired checker is InternalError.
+func (s *AccountSettingsServiceImpl) checkTerminating(ctx context.Context, accountID string) error {
+	if s.terminationChecker == nil {
+		slog.ErrorContext(ctx, "Account settings seal has no termination checker wired", "accountID", accountID)
+		return errors.New(awserrors.ErrorInternalError)
+	}
+
+	terminating, err := s.terminationChecker.AccountTerminating(ctx, accountID)
+	if err != nil {
+		slog.ErrorContext(ctx, "Failed to read account status for seal precondition", "accountID", accountID, "err", err)
+		return errors.New(awserrors.ErrorInternalError)
+	}
+	if !terminating {
+		return awserrors.Errorf(awserrors.ErrorIncorrectState, "account %s is not TERMINATING", accountID)
+	}
+	return nil
+}
+
+// isRevisionConflict reports a lost CAS race SealAccount should retry. Besides
+// the two sentinels, Create's retry against a delete marker returns a bare
+// *jetstream.APIError carrying a wrong-last-sequence code.
+func isRevisionConflict(err error) bool {
+	if errors.Is(err, jetstream.ErrKeyExists) || errors.Is(err, jetstream.ErrKeyRevisionMismatch) {
+		return true
+	}
+	if apiErr, ok := errors.AsType[*jetstream.APIError](err); ok {
+		return apiErr.ErrorCode == jetstream.JSErrCodeStreamWrongLastSequence ||
+			apiErr.ErrorCode == jetstream.JSErrCodeStreamWrongLastSequenceConstant
+	}
+	return false
+}
+
+// SealAccount seals accountID's settings record so no later mutation can
+// recreate it. It never relies on a read for its claim: the only success is a
+// leader-accepted write, evidenced by the returned revision.
+func (s *AccountSettingsServiceImpl) SealAccount(ctx context.Context, accountID string) (*AccountSettingsSeal, error) {
+	if accountID == "" {
+		return nil, awserrors.Errorf(awserrors.ErrorMissingParameter, "accountID is required to seal settings")
+	}
+	if !awsidentifiers.IsAccountID(accountID) {
+		return nil, awserrors.Errorf(awserrors.ErrorInvalidParameterValue, "accountID %q is not a valid AWS account ID", accountID)
+	}
+	if accountID == awsidentifiers.GlobalAccountID {
+		return nil, awserrors.Errorf(awserrors.ErrorInvalidParameterValue, "settings seal must not target the global account")
+	}
+
+	if err := s.checkTerminating(ctx, accountID); err != nil {
+		return nil, err
+	}
+
+	data, err := json.Marshal(sealedRecord)
+	if err != nil {
+		slog.ErrorContext(ctx, "Failed to encode sealed account settings record", "accountID", accountID, "err", err)
+		return nil, errors.New(awserrors.ErrorInternalError)
+	}
+
+	// settingsKey is never used here: a seal must act on exactly accountID
+	// and never fall back to the global key.
+	for range sealMaxAttempts {
+		var rev uint64
+		entry, getErr := s.settingsKV.Get(ctx, accountID)
+		switch {
+		case getErr == nil:
+			rev = entry.Revision()
+		case errors.Is(getErr, jetstream.ErrKeyNotFound):
+			rev = 0
+		default:
+			slog.ErrorContext(ctx, "Failed to read account settings for seal", "accountID", accountID, "err", getErr)
+			return nil, errors.New(awserrors.ErrorInternalError)
+		}
+
+		var newRev uint64
+		var writeErr error
+		if rev == 0 {
+			newRev, writeErr = s.settingsKV.Create(ctx, accountID, data)
+		} else {
+			newRev, writeErr = s.settingsKV.Update(ctx, accountID, data, rev)
+		}
+
+		switch {
+		case writeErr == nil:
+			return &AccountSettingsSeal{Revision: newRev}, nil
+		case isRevisionConflict(writeErr):
+			continue
+		default:
+			slog.ErrorContext(ctx, "Failed to write sealed account settings record", "accountID", accountID, "err", writeErr)
+			return nil, errors.New(awserrors.ErrorInternalError)
+		}
+	}
+
+	slog.ErrorContext(ctx, "Account settings seal exhausted attempts against concurrent writes", "accountID", accountID, "attempts", sealMaxAttempts)
+	return nil, errors.New(awserrors.ErrorInternalError)
+}

spinifex/domains/ec2/account/seal_test.go

@@ -0,0 +1,559 @@
+package account
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"sync"
+	"testing"
+
+	"github.com/aws/aws-sdk-go/service/ec2"
+	"github.com/mulgadc/spinifex/internal/testkit"
+	awserrors "github.com/mulgadc/spinifex/spinifex/foundation/aws/errors"
+	awsidentifiers "github.com/mulgadc/spinifex/spinifex/foundation/aws/identifiers"
+	"github.com/nats-io/nats.go/jetstream"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+// stubTerminationChecker answers AccountTerminating from a fixed set, for
+// tests that exercise the KV paths rather than IAM itself.
+type stubTerminationChecker struct {
+	terminating map[string]bool
+	err         error
+}
+
+func (c *stubTerminationChecker) AccountTerminating(_ context.Context, accountID string) (bool, error) {
+	if c.err != nil {
+		return false, c.err
+	}
+	return c.terminating[accountID], nil
+}
+
+// terminatingChecker answers true for exactly accountID.
+func terminatingChecker(accountID string) *stubTerminationChecker {
+	return &stubTerminationChecker{terminating: map[string]bool{accountID: true}}
+}
+
+// lastSettingsSeq reports the account-settings stream's last sequence, so a
+// test can assert a failed attempt wrote nothing.
+func lastSettingsSeq(t *testing.T, svc *AccountSettingsServiceImpl) uint64 {
+	t.Helper()
+	stream, err := svc.js.Stream(context.Background(), "KV_"+KVBucketAccountSettings)
+	require.NoError(t, err)
+	info, err := stream.Info(context.Background())
+	require.NoError(t, err)
+	return info.State.LastSeq
+}
+
+// mutationCase is one of the four settings mutations, called uniformly so a
+// test can table over them.
+type mutationCase struct {
+	name string
+	call func(svc *AccountSettingsServiceImpl, ctx context.Context) error
+}
+
+func mutationTable() []mutationCase {
+	return []mutationCase{
+		{"EnableEbs", func(svc *AccountSettingsServiceImpl, ctx context.Context) error {
+			_, err := svc.EnableEbsEncryptionByDefault(ctx, &ec2.EnableEbsEncryptionByDefaultInput{}, testAccountID)
+			return err
+		}},
+		{"DisableEbs", func(svc *AccountSettingsServiceImpl, ctx context.Context) error {
+			_, err := svc.DisableEbsEncryptionByDefault(ctx, &ec2.DisableEbsEncryptionByDefaultInput{}, testAccountID)
+			return err
+		}},
+		{"EnableSerial", func(svc *AccountSettingsServiceImpl, ctx context.Context) error {
+			_, err := svc.EnableSerialConsoleAccess(ctx, &ec2.EnableSerialConsoleAccessInput{}, testAccountID)
+			return err
+		}},
+		{"DisableSerial", func(svc *AccountSettingsServiceImpl, ctx context.Context) error {
+			_, err := svc.DisableSerialConsoleAccess(ctx, &ec2.DisableSerialConsoleAccessInput{}, testAccountID)
+			return err
+		}},
+	}
+}
+
+// seedSettingsState prepares testAccountID's record for a race test: present
+// (an active record), absent (never written) or tombstone (written then
+// deleted, leaving a delete marker Get hides).
+func seedSettingsState(t *testing.T, svc *AccountSettingsServiceImpl, state string) {
+	t.Helper()
+	ctx := context.Background()
+	switch state {
+	case "present":
+		_, err := svc.EnableEbsEncryptionByDefault(ctx, &ec2.EnableEbsEncryptionByDefaultInput{}, testAccountID)
+		require.NoError(t, err)
+	case "tombstone":
+		_, err := svc.EnableEbsEncryptionByDefault(ctx, &ec2.EnableEbsEncryptionByDefaultInput{}, testAccountID)
+		require.NoError(t, err)
+		require.NoError(t, svc.settingsKV.Delete(ctx, testAccountID))
+	case "absent":
+	default:
+		t.Fatalf("unknown seed state %q", state)
+	}
+}
+
+// sealRaceKV wraps a real KeyValue so a test can inject a deterministic race
+// between a caller's read and its guarded write (Create or Update), instead
+// of relying on goroutine timing.
+type sealRaceKV struct {
+	jetstream.KeyValue
+
+	writeCalls  int
+	beforeWrite func(attempt int) error
+}
+
+func (f *sealRaceKV) Create(ctx context.Context, key string, value []byte, opts ...jetstream.KVCreateOpt) (uint64, error) {
+	f.writeCalls++
+	if f.beforeWrite != nil {
+		if err := f.beforeWrite(f.writeCalls); err != nil {
+			return 0, err
+		}
+	}
+	return f.KeyValue.Create(ctx, key, value, opts...)
+}
+
+func (f *sealRaceKV) Update(ctx context.Context, key string, value []byte, revision uint64) (uint64, error) {
+	f.writeCalls++
+	if f.beforeWrite != nil {
+		if err := f.beforeWrite(f.writeCalls); err != nil {
+			return 0, err
+		}
+	}
+	return f.KeyValue.Update(ctx, key, value, revision)
+}
+
+// staleGetKV always answers Get with a fixed snapshot taken at construction,
+// so a caller's CAS keeps retrying against a revision that can never win.
+type staleGetKV struct {
+	jetstream.KeyValue
+
+	entry jetstream.KeyValueEntry
+}
+
+func (f *staleGetKV) Get(context.Context, string) (jetstream.KeyValueEntry, error) {
+	return f.entry, nil
+}
+
+// staleNotFoundKV always answers Get as absent, regardless of what the real
+// bucket holds.
+type staleNotFoundKV struct {
+	jetstream.KeyValue
+}
+
+func (f *staleNotFoundKV) Get(context.Context, string) (jetstream.KeyValueEntry, error) {
+	return nil, jetstream.ErrKeyNotFound
+}
+
+// errGetKV always fails Get with a fixed store error.
+type errGetKV struct {
+	jetstream.KeyValue
+
+	err error
+}
+
+func (f *errGetKV) Get(context.Context, string) (jetstream.KeyValueEntry, error) {
+	return nil, f.err
+}
+
+func decodeSettings(t *testing.T, entry jetstream.KeyValueEntry) AccountSettingsRecord {
+	t.Helper()
+	var rec AccountSettingsRecord
+	require.NoError(t, json.Unmarshal(entry.Value(), &rec))
+	return rec
+}
+
+// A seal landing between a mutation's read and its guarded write is not
+// outlived: the mutation's retry sees the seal and is rejected, and nothing
+// it wrote survives.
+func TestSealRaceMutation_SealLandsBetweenReadAndWrite(t *testing.T) {
+	for _, m := range mutationTable() {
+		for _, state := range []string{"present", "absent", "tombstone"} {
+			t.Run(m.name+"_"+state, func(t *testing.T) {
+				svc := setupTestAccountService(t)
+				ctx := context.Background()
+				seedSettingsState(t, svc, state)
+
+				realKV := svc.settingsKV
+				sealer := &AccountSettingsServiceImpl{js: svc.js, settingsKV: realKV}
+				sealer.SetAccountTerminationChecker(terminatingChecker(testAccountID))
+
+				var sealed *AccountSettingsSeal
+				race := &sealRaceKV{KeyValue: realKV}
+				race.beforeWrite = func(attempt int) error {
+					if attempt != 1 {
+						return nil
+					}
+					var sealErr error
+					sealed, sealErr = sealer.SealAccount(ctx, testAccountID)
+					return sealErr
+				}
+				svc.settingsKV = race
+
+				err := m.call(svc, ctx)
+				require.Error(t, err)
+				assert.True(t, awserrors.IsErrorCode(err, awserrors.ErrorBlocked), "got %v", err)
+				require.NotNil(t, sealed)
+
+				entry, err := realKV.Get(ctx, testAccountID)
+				require.NoError(t, err)
+				assert.Equal(t, AccountSettingsStateSealed, decodeSettings(t, entry).State)
+				assert.Equal(t, sealed.Revision, entry.Revision(), "no write landed after the seal")
+			})
+		}
+	}
+}
+
+// A mutation landing inside a seal's read-write window does not survive it:
+// the seal retries against the mutation's write and wins, leaving the
+// sealed value with both booleans false.
+func TestSealRaceMutation_MutationLandsBetweenSealReadAndWrite(t *testing.T) {
+	for _, m := range mutationTable() {
+		for _, state := range []string{"present", "absent", "tombstone"} {
+			t.Run(m.name+"_"+state, func(t *testing.T) {
+				svc := setupTestAccountService(t)
+				ctx := context.Background()
+				seedSettingsState(t, svc, state)
+
+				realKV := svc.settingsKV
+				race := &sealRaceKV{KeyValue: realKV}
+				race.beforeWrite = func(attempt int) error {
+					if attempt != 1 {
+						return nil
+					}
+					_ = m.call(svc, ctx)
+					return nil
+				}
+
+				sealer := &AccountSettingsServiceImpl{js: svc.js, settingsKV: race}
+				sealer.SetAccountTerminationChecker(terminatingChecker(testAccountID))
+
+				seal, err := sealer.SealAccount(ctx, testAccountID)
+				require.NoError(t, err)
+
+				entry, err := realKV.Get(ctx, testAccountID)
+				require.NoError(t, err)
+				rec := decodeSettings(t, entry)
+				assert.Equal(t, AccountSettingsStateSealed, rec.State)
+				assert.False(t, rec.EbsEncryptionByDefault)
+				assert.False(t, rec.SerialConsoleAccess)
+				assert.Equal(t, seal.Revision, entry.Revision())
+			})
+		}
+	}
+}
+
+// None of the four mutations can act on an already-sealed record.
+func TestMutationOnSealedRecordIsRejected(t *testing.T) {
+	for _, m := range mutationTable() {
+		t.Run(m.name, func(t *testing.T) {
+			svc := setupTestAccountService(t)
+			ctx := context.Background()
+			svc.SetAccountTerminationChecker(terminatingChecker(testAccountID))
+			_, err := svc.SealAccount(ctx, testAccountID)
+			require.NoError(t, err)
+
+			before, err := svc.settingsKV.Get(ctx, testAccountID)
+			require.NoError(t, err)
+
+			err = m.call(svc, ctx)
+			require.Error(t, err)
+			assert.True(t, awserrors.IsErrorCode(err, awserrors.ErrorBlocked), "got %v", err)
+
+			after, err := svc.settingsKV.Get(ctx, testAccountID)
+			require.NoError(t, err)
+			assert.Equal(t, before.Revision(), after.Revision())
+		})
+	}
+}
+
+// Every settings mutation races one seal on a real bucket; afterwards the
+// record is sealed and stays sealed, so no success among them could have
+// landed after the seal.
+func TestSealConcurrentMutations(t *testing.T) {
+	svc := setupTestAccountService(t)
+	ctx := context.Background()
+	svc.SetAccountTerminationChecker(terminatingChecker(testAccountID))
+	_, err := svc.EnableEbsEncryptionByDefault(ctx, &ec2.EnableEbsEncryptionByDefaultInput{}, testAccountID)
+	require.NoError(t, err)
+
+	mutations := mutationTable()
+	errs := make([]error, len(mutations))
+	var wg sync.WaitGroup
+	for i, m := range mutations {
+		wg.Add(1)
+		go func(i int, m mutationCase) {
+			defer wg.Done()
+			errs[i] = m.call(svc, ctx)
+		}(i, m)
+	}
+
+	var sealResult *AccountSettingsSeal
+	var sealErr error
+	wg.Go(func() {
+		sealResult, sealErr = svc.SealAccount(ctx, testAccountID)
+	})
+	wg.Wait()
+
+	require.NoError(t, sealErr)
+	require.NotNil(t, sealResult)
+	for i, err := range errs {
+		if err != nil {
+			assert.True(t, awserrors.IsErrorCode(err, awserrors.ErrorBlocked), "mutation %s: got %v", mutations[i].name, err)
+		}
+	}
+
+	entry, err := svc.settingsKV.Get(ctx, testAccountID)
+	require.NoError(t, err)
+	assert.Equal(t, AccountSettingsStateSealed, decodeSettings(t, entry).State)
+	assert.Equal(t, sealResult.Revision, entry.Revision())
+
+	_, err = svc.GetEbsEncryptionByDefault(ctx, &ec2.GetEbsEncryptionByDefaultInput{}, testAccountID)
+	assert.True(t, awserrors.IsErrorCode(err, awserrors.ErrorBlocked))
+}
+
+// A Get pinned to a revision the record has already moved past exhausts the
+// mutation's retries without ever writing.
+func TestMutationStaleReadExhaustsWithoutWriting(t *testing.T) {
+	svc := setupTestAccountService(t)
+	ctx := context.Background()
+
+	_, err := svc.EnableEbsEncryptionByDefault(ctx, &ec2.EnableEbsEncryptionByDefaultInput{}, testAccountID)
+	require.NoError(t, err)
+	stale, err := svc.settingsKV.Get(ctx, testAccountID)
+	require.NoError(t, err)
+
+	_, err = svc.DisableEbsEncryptionByDefault(ctx, &ec2.DisableEbsEncryptionByDefaultInput{}, testAccountID)
+	require.NoError(t, err)
+
+	before := lastSettingsSeq(t, svc)
+	svc.settingsKV = &staleGetKV{KeyValue: svc.settingsKV, entry: stale}
+
+	_, err = svc.EnableSerialConsoleAccess(ctx, &ec2.EnableSerialConsoleAccessInput{}, testAccountID)
+	assert.True(t, awserrors.IsErrorCode(err, awserrors.ErrorInternalError), "got %v", err)
+	assert.Equal(t, before, lastSettingsSeq(t, svc))
+}
+
+// A Get that always reports absent, while a sealed record actually exists,
+// never gets to recreate it: every Create attempt conflicts.
+func TestMutationStaleAbsentDoesNotRecreate(t *testing.T) {
+	svc := setupTestAccountService(t)
+	ctx := context.Background()
+	svc.SetAccountTerminationChecker(terminatingChecker(testAccountID))
+	_, err := svc.SealAccount(ctx, testAccountID)
+	require.NoError(t, err)
+
+	before := lastSettingsSeq(t, svc)
+	svc.settingsKV = &staleNotFoundKV{KeyValue: svc.settingsKV}
+
+	_, err = svc.EnableEbsEncryptionByDefault(ctx, &ec2.EnableEbsEncryptionByDefaultInput{}, testAccountID)
+	assert.True(t, awserrors.IsErrorCode(err, awserrors.ErrorInternalError), "got %v", err)
+	assert.Equal(t, before, lastSettingsSeq(t, svc))
+}
+
+// A revision that keeps moving under sustained concurrent writes must not
+// let the seal spin forever: it gives up, bounded by its attempt budget,
+// and its own write never lands.
+func TestSealSustainedConflictFailsWithoutWriting(t *testing.T) {
+	svc := setupTestAccountService(t)
+	ctx := context.Background()
+	svc.SetAccountTerminationChecker(terminatingChecker(testAccountID))
+	_, err := svc.EnableEbsEncryptionByDefault(ctx, &ec2.EnableEbsEncryptionByDefaultInput{}, testAccountID)
+	require.NoError(t, err)
+
+	realKV := svc.settingsKV
+	race := &sealRaceKV{KeyValue: realKV}
+	race.beforeWrite = func(int) error {
+		data, merr := json.Marshal(&AccountSettingsRecord{SerialConsoleAccess: true})
+		require.NoError(t, merr)
+		_, perr := realKV.Put(ctx, testAccountID, data)
+		return perr
+	}
+	svc.settingsKV = race
+
+	_, err = svc.SealAccount(ctx, testAccountID)
+	assert.True(t, awserrors.IsErrorCode(err, awserrors.ErrorInternalError), "got %v", err)
+	assert.Equal(t, sealMaxAttempts, race.writeCalls)
+
+	entry, err := realKV.Get(ctx, testAccountID)
+	require.NoError(t, err)
+	assert.NotEqual(t, AccountSettingsStateSealed, decodeSettings(t, entry).State, "the seal's own write never landed")
+}
+
+// A store error on the seal's read reports failed, never written.
+func TestSealStoreErrorWritesNothing(t *testing.T) {
+	svc := setupTestAccountService(t)
+	ctx := context.Background()
+	svc.SetAccountTerminationChecker(terminatingChecker(testAccountID))
+
+	before := lastSettingsSeq(t, svc)
+	svc.settingsKV = &errGetKV{KeyValue: svc.settingsKV, err: errors.New("kv unavailable")}
+
+	_, err := svc.SealAccount(ctx, testAccountID)
+	assert.True(t, awserrors.IsErrorCode(err, awserrors.ErrorInternalError), "got %v", err)
+	assert.Equal(t, before, lastSettingsSeq(t, svc))
+}
+
+func TestIsRevisionConflictRecognisesTombstoneRetry(t *testing.T) {
+	assert.True(t, isRevisionConflict(jetstream.ErrKeyExists))
+	assert.True(t, isRevisionConflict(jetstream.ErrKeyRevisionMismatch))
+	assert.True(t, isRevisionConflict(&jetstream.APIError{ErrorCode: jetstream.JSErrCodeStreamWrongLastSequence}))
+	assert.True(t, isRevisionConflict(&jetstream.APIError{ErrorCode: jetstream.JSErrCodeStreamWrongLastSequenceConstant}))
+	assert.True(t, isRevisionConflict(fmt.Errorf("wrapped: %w", &jetstream.APIError{ErrorCode: jetstream.JSErrCodeStreamWrongLastSequence})))
+	assert.False(t, isRevisionConflict(errors.New("unrelated failure")))
+	assert.False(t, isRevisionConflict(nil))
+}
+
+// An account IAM does not hold as TERMINATING refuses the seal before any
+// settings-KV write, whether the key is present or absent.
+func TestSealRefusedForActiveAccount(t *testing.T) {
+	for _, state := range []string{"present", "absent"} {
+		t.Run(state, func(t *testing.T) {
+			svc := setupTestAccountService(t)
+			ctx := context.Background()
+			if state == "present" {
+				_, err := svc.EnableEbsEncryptionByDefault(ctx, &ec2.EnableEbsEncryptionByDefaultInput{}, testAccountID)
+				require.NoError(t, err)
+			}
+			svc.SetAccountTerminationChecker(&stubTerminationChecker{terminating: map[string]bool{}})
+
+			before := lastSettingsSeq(t, svc)
+			_, err := svc.SealAccount(ctx, testAccountID)
+			assert.True(t, awserrors.IsErrorCode(err, awserrors.ErrorIncorrectState), "got %v", err)
+			assert.Equal(t, before, lastSettingsSeq(t, svc))
+		})
+	}
+}
+
+func TestSealRefusedForMissingIAMRecord(t *testing.T) {
+	svc := setupTestAccountService(t)
+	ctx := context.Background()
+	svc.SetAccountTerminationChecker(&stubTerminationChecker{terminating: map[string]bool{}})
+
+	before := lastSettingsSeq(t, svc)
+	_, err := svc.SealAccount(ctx, testAccountID)
+	assert.True(t, awserrors.IsErrorCode(err, awserrors.ErrorIncorrectState), "got %v", err)
+	assert.Equal(t, before, lastSettingsSeq(t, svc))
+}
+
+func TestSealRefusedWhenIAMReadFails(t *testing.T) {
+	svc := setupTestAccountService(t)
+	ctx := context.Background()
+	svc.SetAccountTerminationChecker(&stubTerminationChecker{err: errors.New("iam unavailable")})
+
+	before := lastSettingsSeq(t, svc)
+	_, err := svc.SealAccount(ctx, testAccountID)
+	assert.True(t, awserrors.IsErrorCode(err, awserrors.ErrorInternalError), "got %v", err)
+	assert.Equal(t, before, lastSettingsSeq(t, svc))
+}
+
+func TestSealRefusedWithNoCheckerWired(t *testing.T) {
+	svc := setupTestAccountService(t)
+	ctx := context.Background()
+
+	before := lastSettingsSeq(t, svc)
+	_, err := svc.SealAccount(ctx, testAccountID)
+	assert.True(t, awserrors.IsErrorCode(err, awserrors.ErrorInternalError), "got %v", err)
+	assert.Equal(t, before, lastSettingsSeq(t, svc))
+}
+
+func TestSealAcceptedForTerminatingAccount(t *testing.T) {
+	svc := setupTestAccountService(t)
+	ctx := context.Background()
+	svc.SetAccountTerminationChecker(terminatingChecker(testAccountID))
+
+	seal, err := svc.SealAccount(ctx, testAccountID)
+	require.NoError(t, err)
+	assert.NotZero(t, seal.Revision)
+
+	entry, err := svc.settingsKV.Get(ctx, testAccountID)
+	require.NoError(t, err)
+	assert.Equal(t, seal.Revision, entry.Revision())
+}
+
+// A sealed record stays sealed after the service is rebuilt over the same
+// JetStream, and every one of the six settings actions still refuses it.
+func TestSealSurvivesServiceReconstruction(t *testing.T) {
+	_, nc, _ := testutil.StartTestJetStream(t)
+	ctx := context.Background()
+
+	svc, err := NewAccountSettingsServiceImplWithNATS(ctx, nil, nc)
+	require.NoError(t, err)
+	svc.SetAccountTerminationChecker(terminatingChecker(testAccountID))
+	first, err := svc.SealAccount(ctx, testAccountID)
+	require.NoError(t, err)
+
+	rebuilt, err := NewAccountSettingsServiceImplWithNATS(ctx, nil, nc)
+	require.NoError(t, err)
+
+	entry, err := rebuilt.settingsKV.Get(ctx, testAccountID)
+	require.NoError(t, err)
+	assert.Equal(t, AccountSettingsStateSealed, decodeSettings(t, entry).State)
+
+	_, err = rebuilt.GetEbsEncryptionByDefault(ctx, &ec2.GetEbsEncryptionByDefaultInput{}, testAccountID)
+	assert.True(t, awserrors.IsErrorCode(err, awserrors.ErrorBlocked))
+	_, err = rebuilt.GetSerialConsoleAccessStatus(ctx, &ec2.GetSerialConsoleAccessStatusInput{}, testAccountID)
+	assert.True(t, awserrors.IsErrorCode(err, awserrors.ErrorBlocked))
+	_, err = rebuilt.EnableEbsEncryptionByDefault(ctx, &ec2.EnableEbsEncryptionByDefaultInput{}, testAccountID)
+	assert.True(t, awserrors.IsErrorCode(err, awserrors.ErrorBlocked))
+	_, err = rebuilt.DisableEbsEncryptionByDefault(ctx, &ec2.DisableEbsEncryptionByDefaultInput{}, testAccountID)
+	assert.True(t, awserrors.IsErrorCode(err, awserrors.ErrorBlocked))
+	_, err = rebuilt.EnableSerialConsoleAccess(ctx, &ec2.EnableSerialConsoleAccessInput{}, testAccountID)
+	assert.True(t, awserrors.IsErrorCode(err, awserrors.ErrorBlocked))
+	_, err = rebuilt.DisableSerialConsoleAccess(ctx, &ec2.DisableSerialConsoleAccessInput{}, testAccountID)
+	assert.True(t, awserrors.IsErrorCode(err, awserrors.ErrorBlocked))
+
+	rebuilt.SetAccountTerminationChecker(terminatingChecker(testAccountID))
+	second, err := rebuilt.SealAccount(ctx, testAccountID)
+	require.NoError(t, err)
+	assert.Greater(t, second.Revision, first.Revision)
+}
+
+func TestSealAbsentKeyCreatesSealedRecord(t *testing.T) {
+	svc := setupTestAccountService(t)
+	ctx := context.Background()
+	svc.SetAccountTerminationChecker(terminatingChecker(testAccountID))
+
+	_, err := svc.settingsKV.Get(ctx, testAccountID)
+	require.ErrorIs(t, err, jetstream.ErrKeyNotFound)
+
+	seal, err := svc.SealAccount(ctx, testAccountID)
+	require.NoError(t, err)
+
+	entry, err := svc.settingsKV.Get(ctx, testAccountID)
+	require.NoError(t, err)
+	assert.Equal(t, seal.Revision, entry.Revision())
+	assert.Equal(t, AccountSettingsStateSealed, decodeSettings(t, entry).State)
+}
+
+func TestSealRefusesEmptyInvalidAndGlobalAccount(t *testing.T) {
+	svc := setupTestAccountService(t)
+	ctx := context.Background()
+
+	_, err := svc.SealAccount(ctx, "")
+	assert.True(t, awserrors.IsErrorCode(err, awserrors.ErrorMissingParameter), "got %v", err)
+
+	_, err = svc.SealAccount(ctx, "not-an-account")
+	assert.True(t, awserrors.IsErrorCode(err, awserrors.ErrorInvalidParameterValue), "got %v", err)
+
+	_, err = svc.SealAccount(ctx, awsidentifiers.GlobalAccountID)
+	assert.True(t, awserrors.IsErrorCode(err, awserrors.ErrorInvalidParameterValue), "got %v", err)
+}
+
+func TestSealLeavesOtherAccountsUntouched(t *testing.T) {
+	svc := setupTestAccountService(t)
+	ctx := context.Background()
+	accountB := "222222222222"
+	_, err := svc.EnableEbsEncryptionByDefault(ctx, &ec2.EnableEbsEncryptionByDefaultInput{}, accountB)
+	require.NoError(t, err)
+
+	svc.SetAccountTerminationChecker(terminatingChecker(testAccountID))
+	_, err = svc.SealAccount(ctx, testAccountID)
+	require.NoError(t, err)
+
+	out, err := svc.GetEbsEncryptionByDefault(ctx, &ec2.GetEbsEncryptionByDefaultInput{}, accountB)
+	require.NoError(t, err)
+	assert.True(t, *out.EbsEncryptionByDefault)
+}

spinifex/domains/ec2/account/service_impl.go

@@ -5,12 +5,13 @@ import (
 	"encoding/json"
 	"errors"
 	"fmt"
-	awsidentifiers "github.com/mulgadc/spinifex/spinifex/foundation/aws/identifiers"
 	"log/slog"
 
 	"github.com/aws/aws-sdk-go/aws"
 	"github.com/aws/aws-sdk-go/service/ec2"
 	"github.com/mulgadc/spinifex/spinifex/bootstrap/config"
+	awserrors "github.com/mulgadc/spinifex/spinifex/foundation/aws/errors"
+	awsidentifiers "github.com/mulgadc/spinifex/spinifex/foundation/aws/identifiers"
 	"github.com/mulgadc/spinifex/spinifex/foundation/state/kvutil"
 	"github.com/mulgadc/spinifex/spinifex/foundation/state/migrate"
 	"github.com/nats-io/nats.go"
@@ -20,14 +21,19 @@ import (
 const (
 	KVBucketAccountSettings        = "spinifex-ec2-account-settings"
 	KVBucketAccountSettingsVersion = 1
-	KeyEbsEncryptionDefault        = "ebs-encryption-default"
-	KeySerialConsoleAccess         = "serial-console-access"
 )
 
+// AccountSettingsState is a settings record's lifecycle state. "" is active;
+// a record written before the state field existed has none and stays active.
+type AccountSettingsState string
+
+const AccountSettingsStateSealed AccountSettingsState = "sealed"
+
 // AccountSettingsRecord represents stored account settings.
 type AccountSettingsRecord struct {
-	EbsEncryptionByDefault bool `json:"ebs_encryption_by_default"`
-	SerialConsoleAccess    bool `json:"serial_console_access"`
+	EbsEncryptionByDefault bool                 `json:"ebs_encryption_by_default"`
+	SerialConsoleAccess    bool                 `json:"serial_console_access"`
+	State                  AccountSettingsState `json:"state,omitempty"`
 }
 
 // AccountSettingsServiceImpl implements account settings operations with NATS JetStream persistence.
@@ -35,6 +41,10 @@ type AccountSettingsServiceImpl struct {
 	config     *config.Config
 	js         jetstream.JetStream
 	settingsKV jetstream.KeyValue
+
+	// terminationChecker is the seal precondition; SealAccount refuses until
+	// it is wired.
+	terminationChecker AccountTerminationChecker
 }
 
 var _ AccountSettingsService = (*AccountSettingsServiceImpl)(nil)
@@ -72,50 +82,99 @@ func settingsKey(accountID string) string {
 	return accountID
 }
 
-// getSettings retrieves current account settings.
+// errSettingsSealed marks a mutation or read that found a sealed record, so
+// callers map it to Blocked without logging it as a fault.
+var errSettingsSealed = errors.New(awserrors.ErrorBlocked)
+
+// errSettingsContended marks a mutation whose CAS retries were exhausted
+// under concurrent writers, bounding the retry loop instead of spinning.
+var errSettingsContended = errors.New("account settings: CAS retries exhausted")
+
+// checkActive rejects a sealed or unknown-state record; "" is active and
+// passes through. A mutation must never act on a record that failed this.
+func checkActive(rec *AccountSettingsRecord) error {
+	switch rec.State {
+	case "":
+		return nil
+	case AccountSettingsStateSealed:
+		return errSettingsSealed
+	default:
+		return fmt.Errorf("account settings: unknown state %q", rec.State)
+	}
+}
+
+// mapSettingsError turns a mutateSettings or getSettings failure into the AWS
+// error a caller returns: sealed is an expected, unlogged Blocked; everything
+// else is InternalError, logged with the account ID.
+func mapSettingsError(ctx context.Context, accountID string, err error) error {
+	switch {
+	case err == nil:
+		return nil
+	case errors.Is(err, errSettingsSealed):
+		return errors.New(awserrors.ErrorBlocked)
+	case errors.Is(err, errSettingsContended):
+		slog.ErrorContext(ctx, "Account settings CAS retries exhausted under contention", "accountID", accountID)
+	case errors.Is(err, kvutil.ErrRead):
+		slog.ErrorContext(ctx, "Failed to read account settings from KV", "accountID", accountID, "err", err)
+	case errors.Is(err, kvutil.ErrDecode):
+		slog.ErrorContext(ctx, "Corrupt account settings record in KV", "accountID", accountID, "err", err)
+	case errors.Is(err, kvutil.ErrEncode):
+		slog.ErrorContext(ctx, "Failed to marshal account settings record", "accountID", accountID, "err", err)
+	case errors.Is(err, kvutil.ErrWrite):
+		slog.ErrorContext(ctx, "Failed to write account settings to KV", "accountID", accountID, "err", err)
+	default:
+		slog.ErrorContext(ctx, "Account settings operation failed", "accountID", accountID, "err", err)
+	}
+	return errors.New(awserrors.ErrorInternalError)
+}
+
+// mutateSettings applies set to accountID's record at the revision it read.
+// changed is always true, so every success is a leader-accepted write over an
+// active value; a seal landing between the read and the write fails the CAS.
+func (s *AccountSettingsServiceImpl) mutateSettings(ctx context.Context, accountID string, set func(*AccountSettingsRecord)) error {
+	_, err := kvutil.Update(ctx, s.settingsKV, settingsKey(accountID), kvutil.CASConfig{
+		CreateIfAbsent: true,
+		Exhausted:      func(string, int) error { return errSettingsContended },
+	}, func(rec *AccountSettingsRecord) (bool, error) {
+		if err := checkActive(rec); err != nil {
+			return false, err
+		}
+		set(rec)
+		return true, nil
+	})
+	return mapSettingsError(ctx, accountID, err)
+}
+
+// getSettings retrieves current account settings, rejecting a sealed or
+// unknown-state record so a read never answers from behind a seal. Absent is
+// the AWS defaults; a read is a direct get and can be stale.
 func (s *AccountSettingsServiceImpl) getSettings(ctx context.Context, accountID string) (*AccountSettingsRecord, error) {
 	entry, err := s.settingsKV.Get(ctx, settingsKey(accountID))
 	if err != nil {
 		if errors.Is(err, jetstream.ErrKeyNotFound) {
-			return &AccountSettingsRecord{
-				EbsEncryptionByDefault: false,
-				SerialConsoleAccess:    false,
-			}, nil
+			return &AccountSettingsRecord{}, nil
 		}
-		return nil, fmt.Errorf("failed to get account settings: %w", err)
+		slog.ErrorContext(ctx, "Failed to read account settings from KV", "accountID", accountID, "err", err)
+		return nil, errors.New(awserrors.ErrorInternalError)
 	}
 
 	var settings AccountSettingsRecord
 	if err := json.Unmarshal(entry.Value(), &settings); err != nil {
-		return nil, err
+		slog.ErrorContext(ctx, "Corrupt account settings record in KV", "accountID", accountID, "err", err)
+		return nil, errors.New(awserrors.ErrorInternalError)
+	}
+	if err := checkActive(&settings); err != nil {
+		return nil, mapSettingsError(ctx, accountID, err)
 	}
 
 	return &settings, nil
 }
 
-// saveSettings saves current account settings.
-func (s *AccountSettingsServiceImpl) saveSettings(ctx context.Context, settings *AccountSettingsRecord, accountID string) error {
-	data, err := json.Marshal(settings)
-	if err != nil {
-		return err
-	}
-
-	_, err = s.settingsKV.Put(ctx, settingsKey(accountID), data)
-	return err
-}
-
 // EnableEbsEncryptionByDefault enables EBS encryption by default for the account.
 func (s *AccountSettingsServiceImpl) EnableEbsEncryptionByDefault(ctx context.Context, input *ec2.EnableEbsEncryptionByDefaultInput, accountID string) (*ec2.EnableEbsEncryptionByDefaultOutput, error) {
 	slog.InfoContext(ctx, "EnableEbsEncryptionByDefault called", "accountID", accountID)
 
-	settings, err := s.getSettings(ctx, accountID)
-	if err != nil {
-		return nil, err
-	}
-
-	settings.EbsEncryptionByDefault = true
-
-	if err := s.saveSettings(ctx, settings, accountID); err != nil {
+	if err := s.mutateSettings(ctx, accountID, func(rec *AccountSettingsRecord) { rec.EbsEncryptionByDefault = true }); err != nil {
 		return nil, err
 	}
 
@@ -128,14 +187,7 @@ func (s *AccountSettingsServiceImpl) EnableEbsEncryptionByDefault(ctx context.Co
 func (s *AccountSettingsServiceImpl) DisableEbsEncryptionByDefault(ctx context.Context, input *ec2.DisableEbsEncryptionByDefaultInput, accountID string) (*ec2.DisableEbsEncryptionByDefaultOutput, error) {
 	slog.InfoContext(ctx, "DisableEbsEncryptionByDefault called", "accountID", accountID)
 
-	settings, err := s.getSettings(ctx, accountID)
-	if err != nil {
-		return nil, err
-	}
-
-	settings.EbsEncryptionByDefault = false
-
-	if err := s.saveSettings(ctx, settings, accountID); err != nil {
+	if err := s.mutateSettings(ctx, accountID, func(rec *AccountSettingsRecord) { rec.EbsEncryptionByDefault = false }); err != nil {
 		return nil, err
 	}
 
@@ -176,14 +228,7 @@ func (s *AccountSettingsServiceImpl) GetSerialConsoleAccessStatus(ctx context.Co
 func (s *AccountSettingsServiceImpl) EnableSerialConsoleAccess(ctx context.Context, input *ec2.EnableSerialConsoleAccessInput, accountID string) (*ec2.EnableSerialConsoleAccessOutput, error) {
 	slog.InfoContext(ctx, "EnableSerialConsoleAccess called", "accountID", accountID)
 
-	settings, err := s.getSettings(ctx, accountID)
-	if err != nil {
-		return nil, err
-	}
-
-	settings.SerialConsoleAccess = true
-
-	if err := s.saveSettings(ctx, settings, accountID); err != nil {
+	if err := s.mutateSettings(ctx, accountID, func(rec *AccountSettingsRecord) { rec.SerialConsoleAccess = true }); err != nil {
 		return nil, err
 	}
 
@@ -196,14 +241,7 @@ func (s *AccountSettingsServiceImpl) EnableSerialConsoleAccess(ctx context.Conte
 func (s *AccountSettingsServiceImpl) DisableSerialConsoleAccess(ctx context.Context, input *ec2.DisableSerialConsoleAccessInput, accountID string) (*ec2.DisableSerialConsoleAccessOutput, error) {
 	slog.InfoContext(ctx, "DisableSerialConsoleAccess called", "accountID", accountID)
 
-	settings, err := s.getSettings(ctx, accountID)
-	if err != nil {
-		return nil, err
-	}
-
-	settings.SerialConsoleAccess = false
-
-	if err := s.saveSettings(ctx, settings, accountID); err != nil {
+	if err := s.mutateSettings(ctx, accountID, func(rec *AccountSettingsRecord) { rec.SerialConsoleAccess = false }); err != nil {
 		return nil, err
 	}
 

spinifex/domains/ec2/account/service_impl_test.go

@@ -2,11 +2,13 @@ package account
 
 import (
 	"context"
+	"encoding/json"
 	awsidentifiers "github.com/mulgadc/spinifex/spinifex/foundation/aws/identifiers"
 	"testing"
 
 	"github.com/aws/aws-sdk-go/service/ec2"
 	"github.com/mulgadc/spinifex/internal/testkit"
+	awserrors "github.com/mulgadc/spinifex/spinifex/foundation/aws/errors"
 	"github.com/stretchr/testify/assert"
 	"github.com/stretchr/testify/require"
 )
@@ -143,3 +145,44 @@ func TestSettingsKey_EmptyAccountIDFallback(t *testing.T) {
 	assert.Equal(t, awsidentifiers.GlobalAccountID, settingsKey(""))
 	assert.Equal(t, "123456789012", settingsKey("123456789012"))
 }
+
+// A raw pre-seal record, with no state field at all, is readable and
+// mutable, and a mutation writes it back without adding one.
+func TestRecordWithoutStateIsActive(t *testing.T) {
+	svc := setupTestAccountService(t)
+	ctx := context.Background()
+	_, err := svc.settingsKV.Put(ctx, testAccountID, []byte(`{"ebs_encryption_by_default":true,"serial_console_access":false}`))
+	require.NoError(t, err)
+
+	out, err := svc.GetEbsEncryptionByDefault(ctx, &ec2.GetEbsEncryptionByDefaultInput{}, testAccountID)
+	require.NoError(t, err)
+	assert.True(t, *out.EbsEncryptionByDefault)
+
+	_, err = svc.EnableSerialConsoleAccess(ctx, &ec2.EnableSerialConsoleAccessInput{}, testAccountID)
+	require.NoError(t, err)
+
+	entry, err := svc.settingsKV.Get(ctx, testAccountID)
+	require.NoError(t, err)
+	assert.NotContains(t, string(entry.Value()), "state")
+}
+
+// A record with an unrecognised state fails closed rather than being treated
+// as active or sealed, for both a read and a mutation.
+func TestUnknownStateFailsClosed(t *testing.T) {
+	svc := setupTestAccountService(t)
+	ctx := context.Background()
+	_, err := svc.settingsKV.Put(ctx, testAccountID, []byte(`{"ebs_encryption_by_default":false,"serial_console_access":false,"state":"quarantined"}`))
+	require.NoError(t, err)
+
+	_, err = svc.GetEbsEncryptionByDefault(ctx, &ec2.GetEbsEncryptionByDefaultInput{}, testAccountID)
+	assert.True(t, awserrors.IsErrorCode(err, awserrors.ErrorInternalError), "got %v", err)
+
+	_, err = svc.EnableEbsEncryptionByDefault(ctx, &ec2.EnableEbsEncryptionByDefaultInput{}, testAccountID)
+	assert.True(t, awserrors.IsErrorCode(err, awserrors.ErrorInternalError), "got %v", err)
+
+	entry, err := svc.settingsKV.Get(ctx, testAccountID)
+	require.NoError(t, err)
+	var rec AccountSettingsRecord
+	require.NoError(t, json.Unmarshal(entry.Value(), &rec))
+	assert.EqualValues(t, "quarantined", rec.State, "the unknown-state record was never overwritten")
+}

spinifex/domains/ec2/account/settings_writes_guarded_test.go

@@ -0,0 +1,68 @@
+package account
+
+//test:in-package — pins what AccountSettingsSealCapabilityVersion 1 asserts:
+// nothing in this package writes the settings KV except through the
+// revision-guarded helpers.
+
+import (
+	"go/ast"
+	"go/parser"
+	"go/token"
+	"io/fs"
+	"path/filepath"
+	"runtime"
+	"strings"
+	"testing"
+)
+
+func accountPackageDir(t *testing.T) string {
+	t.Helper()
+	_, file, _, ok := runtime.Caller(0)
+	if !ok {
+		t.Fatal("cannot locate test source")
+	}
+	return filepath.Dir(file)
+}
+
+// The capability daemons report as version 1 is a claim about every writer in
+// the binary, not just the responder: no non-test file in this package may
+// call Put or PutString, which bypass the revision-guarded mutate/seal paths.
+func TestAccountSettingsWritesAreRevisionGuarded(t *testing.T) {
+	root := accountPackageDir(t)
+	fset := token.NewFileSet()
+	scanned := 0
+
+	err := filepath.WalkDir(root, func(path string, d fs.DirEntry, err error) error {
+		if err != nil || d.IsDir() || !strings.HasSuffix(path, ".go") || strings.HasSuffix(path, "_test.go") {
+			return err
+		}
+		f, perr := parser.ParseFile(fset, path, nil, 0)
+		if perr != nil {
+			return perr
+		}
+		scanned++
+
+		ast.Inspect(f, func(n ast.Node) bool {
+			call, ok := n.(*ast.CallExpr)
+			if !ok {
+				return true
+			}
+			sel, ok := call.Fun.(*ast.SelectorExpr)
+			if !ok {
+				return true
+			}
+			if sel.Sel.Name == "Put" || sel.Sel.Name == "PutString" {
+				t.Errorf("%s: unconditional %s call bypasses the revision-guarded settings write; use mutateSettings or SealAccount's CAS loop instead",
+					fset.Position(call.Pos()), sel.Sel.Name)
+			}
+			return true
+		})
+		return nil
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if scanned == 0 {
+		t.Fatalf("scanned no Go files under %s; the invariant would pass vacuously", root)
+	}
+}

spinifex/domains/ec2/accountcleanup/cleanup.go

@@ -0,0 +1,277 @@
+// Package accountcleanup deletes the proven-safe subset of one account's EC2
+// platform resources through the owners' own delete paths: placement groups
+// with no members that no instance, spot request or live launch-template
+// version names; key pairs none of those name; and launch-template residue no
+// header owns. Live launch templates are never deleted: an EKS nodegroup may
+// name one and EC2 cannot see it. A reference check that cannot complete
+// deletes nothing of the kinds it guards. Every disposition is proved by an
+// owner-store re-scan, never by Describe.
+//
+// It adds no caller, checks no authority and holds no store or transport of
+// its own: the owners and reference readers are injected. Whoever binds it
+// decides who may invoke it. It is not exposed on any NATS subject.
+package accountcleanup
+
+import (
+	"context"
+	"log/slog"
+	"slices"
+	"strings"
+	"time"
+
+	ec2v1 "github.com/mulgadc/spinifex/contracts/ec2/v1"
+	ec2key "github.com/mulgadc/spinifex/spinifex/domains/ec2/key"
+	ec2launchtemplate "github.com/mulgadc/spinifex/spinifex/domains/ec2/launchtemplate"
+	ec2placementgroup "github.com/mulgadc/spinifex/spinifex/domains/ec2/placementgroup"
+	ec2tags "github.com/mulgadc/spinifex/spinifex/domains/ec2/tags"
+	awserrors "github.com/mulgadc/spinifex/spinifex/foundation/aws/errors"
+	awsidentifiers "github.com/mulgadc/spinifex/spinifex/foundation/aws/identifiers"
+	otelsetup "github.com/mulgadc/spinifex/spinifex/foundation/telemetry"
+)
+
+// PlatformCleaner deletes one account's EC2 platform resources and reports
+// what the owner stores hold afterwards. It is destructive.
+type PlatformCleaner interface {
+	CleanupPlatform(ctx context.Context, req *Request) (*Result, error)
+}
+
+var _ PlatformCleaner = (*Cleanup)(nil)
+
+// KeyPairOwner is the key-pair owner's account scan and cleanup.
+type KeyPairOwner interface {
+	ScanAccount(ctx context.Context, accountID string) (*ec2key.AccountScan, error)
+	CleanupAccount(ctx context.Context, accountID string, settledBefore time.Time, referenced func(keyName string) string) (*ec2key.AccountCleanup, error)
+}
+
+// PlacementGroupOwner is the placement-group owner's account scan and cleanup.
+type PlacementGroupOwner interface {
+	ScanAccount(ctx context.Context, accountID string) (*ec2placementgroup.AccountScan, error)
+	CleanupAccount(ctx context.Context, accountID string, referenced func(groupID, groupName string) string) (*ec2placementgroup.AccountCleanup, error)
+}
+
+// LaunchTemplateOwner is the launch-template owner's account scan and residue
+// cleanup; live templates are never deleted by account cleanup.
+type LaunchTemplateOwner interface {
+	ScanAccount(ctx context.Context, accountID string) (*ec2launchtemplate.AccountScan, error)
+	CleanupAccountResidue(ctx context.Context, accountID string, settledBefore time.Time) (*ec2launchtemplate.AccountCleanup, error)
+}
+
+// TagIndexOwner lists and clears the account's tag-index entries.
+type TagIndexOwner interface {
+	ScanAccountResourceIDs(ctx context.Context, accountID string, prefixes ...string) ([]string, error)
+	DeleteAllTags(ctx context.Context, accountID, resourceID string) error
+}
+
+var (
+	_ KeyPairOwner        = (*ec2key.KeyServiceImpl)(nil)
+	_ PlacementGroupOwner = (*ec2placementgroup.PlacementGroupServiceImpl)(nil)
+	_ LaunchTemplateOwner = (*ec2launchtemplate.LaunchTemplateServiceImpl)(nil)
+	_ TagIndexOwner       = (*ec2tags.TagsServiceImpl)(nil)
+)
+
+// DefaultSettleWindow is how old headerless residue must be before it is
+// removed; anything younger may be a create still in flight.
+const DefaultSettleWindow = 10 * time.Minute
+
+// Cleanup composes the owner cleanups and re-scans.
+type Cleanup struct {
+	keys      KeyPairOwner
+	groups    PlacementGroupOwner
+	templates LaunchTemplateOwner
+	tags      TagIndexOwner
+	refs      ReferenceReaders
+
+	// SettleWindow overrides DefaultSettleWindow when positive.
+	SettleWindow time.Duration
+	now          func() time.Time
+}
+
+// New builds a cleanup over the four owners and the reference readers. A nil
+// reader counts as a failed check, so the kinds it guards delete nothing.
+func New(keys KeyPairOwner, groups PlacementGroupOwner, templates LaunchTemplateOwner, tags TagIndexOwner, refs ReferenceReaders) *Cleanup {
+	return &Cleanup{keys: keys, groups: groups, templates: templates, tags: tags, refs: refs, now: time.Now}
+}
+
+// Resource-ID prefixes of the three kinds in the tag index.
+const (
+	keyPairIDPrefix        = "key-"
+	placementGroupIDPrefix = "pg-"
+	launchTemplateIDPrefix = "lt-"
+)
+
+// CleanupPlatform deletes the account's key pairs, empty placement groups and
+// launch templates, removes settled residue and orphaned tag-index entries,
+// then re-scans every owner store to build the result. It is safe to re-run.
+func (c *Cleanup) CleanupPlatform(ctx context.Context, req *Request) (*Result, error) {
+	if req == nil || req.AccountID == "" {
+		return nil, awserrors.Errorf(awserrors.ErrorMissingParameter, "The request must specify an account ID.")
+	}
+	if !awsidentifiers.IsAccountID(req.AccountID) {
+		return nil, awserrors.Errorf(awserrors.ErrorInvalidParameterValue, "The account ID must be 12 digits.")
+	}
+	start := time.Now()
+	window := c.SettleWindow
+	if window <= 0 {
+		window = DefaultSettleWindow
+	}
+	settledBefore := c.now().Add(-window)
+	acct := req.AccountID
+	res := &Result{AccountID: acct, CorrelationRef: req.CorrelationRef}
+
+	refs := c.readReferences(ctx, acct)
+	keyRun := c.cleanKeyPairs(ctx, acct, settledBefore, refs)
+	groupRun := c.cleanPlacementGroups(ctx, acct, refs)
+	templateRun := c.cleanLaunchTemplates(ctx, acct, settledBefore)
+
+	// Tags are listed before the owner re-scans, so a resource created in
+	// between is seen as owned and keeps its tags.
+	tagKind := c.startKind(ec2v1.PlatformKindTagIndex)
+	tagged, tagErr := c.tags.ScanAccountResourceIDs(ctx, acct, keyPairIDPrefix, placementGroupIDPrefix, launchTemplateIDPrefix)
+	if tagErr != nil {
+		slog.WarnContext(ctx, "Platform cleanup: tag index listing failed", "err", tagErr)
+		tagKind.incomplete(ec2v1.PlatformReasonStoreListFailed)
+	}
+
+	owners := map[string]ownedIDs{
+		keyPairIDPrefix:        c.finishKeyPairs(ctx, acct, keyRun, res),
+		placementGroupIDPrefix: c.finishPlacementGroups(ctx, acct, groupRun, res),
+		launchTemplateIDPrefix: c.finishLaunchTemplates(ctx, acct, templateRun, res),
+	}
+	res.Kinds = []KindReport{keyRun.kind.report, groupRun.kind.report, templateRun.kind.report}
+	res.Kinds = append(res.Kinds, c.cleanTagIndex(ctx, acct, tagKind, tagged, tagErr == nil, owners, res))
+
+	slog.InfoContext(ctx, "Platform cleanup completed", "accountID", acct, "correlationRef", req.CorrelationRef,
+		"items", len(res.Items), "duration_ms", otelsetup.Millis(time.Since(start)))
+	return res, nil
+}
+
+// ownedIDs is the set of IDs whose owner record a re-scan found, live or
+// unreadable; resolved is false unless the re-scan listed and read every
+// record, so an absent ID is not proof its tags are orphaned.
+type ownedIDs struct {
+	ids      map[string]bool
+	resolved bool
+}
+
+// kindRun accumulates one kind's report across its cleanup and re-scan.
+type kindRun struct {
+	report KindReport
+	found  bool
+}
+
+func (c *Cleanup) startKind(kind string) *kindRun {
+	return &kindRun{report: KindReport{Kind: kind, StartedAt: c.now().UTC()}}
+}
+
+func (k *kindRun) incomplete(reason string) {
+	if !slices.Contains(k.report.IncompleteReasons, reason) {
+		k.report.IncompleteReasons = append(k.report.IncompleteReasons, reason)
+	}
+}
+
+func (c *Cleanup) finishKind(k *kindRun) {
+	k.report.EndedAt = c.now().UTC()
+	switch {
+	case len(k.report.IncompleteReasons) > 0:
+		k.report.Status = ec2v1.PlatformStatusScanIncomplete
+	case k.found:
+		k.report.Status = ec2v1.PlatformStatusRecordsFound
+	default:
+		k.report.Status = ec2v1.PlatformStatusObservedClean
+	}
+}
+
+// attempts is what an owner said it did, by resource ID. kept holds items
+// left on purpose, with the fixed reason.
+type attempts struct {
+	deleted, failed map[string]bool
+	blocked         map[string]int
+	kept            map[string]string
+}
+
+func newAttempts() attempts {
+	return attempts{deleted: map[string]bool{}, failed: map[string]bool{}, blocked: map[string]int{}, kept: map[string]string{}}
+}
+
+// itemDisposition decides one ID from what the owner did and whether the
+// re-scan still found it. Only a complete re-scan that misses an ID is deleted;
+// an item kept on purpose stays blocked unless that re-scan shows it gone.
+func itemDisposition(id string, did attempts, present, rescanned bool) (string, string) {
+	switch {
+	case did.kept[id] != "" && (present || !rescanned):
+		return DispositionBlocked, did.kept[id]
+	case !rescanned:
+		return DispositionFailed, ReasonAbsenceUnproven
+	case !present:
+		return DispositionDeleted, ""
+	case did.blocked[id] > 0:
+		return DispositionBlocked, ReasonInstanceMembers
+	case did.failed[id]:
+		return DispositionFailed, ReasonOwnerDeleteFailed
+	case did.deleted[id]:
+		return DispositionFailed, ReasonStillPresent
+	default:
+		return DispositionResidue, ReasonAppearedDuringCleanup
+	}
+}
+
+func (r *Result) addItem(k *kindRun, id, disposition, reason string, members int) {
+	if disposition != DispositionDeleted {
+		k.found = true
+	}
+	r.Items = append(r.Items, Item{ResourceID: id, Kind: k.report.Kind, Disposition: disposition, Reason: reason, MemberCount: members})
+}
+
+func (r *Result) addCount(k *kindRun, class, disposition, reason string, n int) {
+	if n <= 0 {
+		return
+	}
+	if disposition != DispositionDeleted {
+		k.found = true
+	}
+	r.Counts = append(r.Counts, Count{Kind: k.report.Kind, Class: class, Disposition: disposition, Reason: reason, Count: n})
+}
+
+// residueCounts reports one headerless class: the proven decrease as deleted
+// (never more than the owner removed), and what remains split by cause, kept
+// records first, by reason in a fixed order.
+func (r *Result) residueCounts(k *kindRun, class string, before, after, removed, failed, unsettled int, kept map[string]int) {
+	r.addCount(k, class, DispositionDeleted, "", min(removed, max(0, before-after)))
+	left := after
+	for _, reason := range keptReasons {
+		n := min(left, kept[reason])
+		r.addCount(k, class, DispositionBlocked, reason, n)
+		left -= n
+	}
+	failedLeft := min(left, failed)
+	unsettledLeft := min(left-failedLeft, unsettled)
+	r.addCount(k, class, DispositionFailed, ReasonOwnerDeleteFailed, failedLeft)
+	r.addCount(k, class, DispositionResidue, ReasonRecentlyWritten, unsettledLeft)
+	r.addCount(k, class, DispositionFailed, ReasonStillPresent, left-failedLeft-unsettledLeft)
+}
+
+// keptReasons are the reasons an owner may keep a record, in report order.
+var keptReasons = []string{ReasonReferenceCheckFailed, ReasonInstanceReference, ReasonSpotRequestReference, ReasonLaunchTemplateReference}
+
+// sortedUnion returns the distinct IDs of all lists, sorted.
+func sortedUnion(lists ...[]string) []string {
+	var out []string
+	for _, l := range lists {
+		for _, id := range l {
+			if !slices.Contains(out, id) {
+				out = append(out, id)
+			}
+		}
+	}
+	slices.Sort(out)
+	return out
+}
+
+func idPrefix(id string) string {
+	for _, prefix := range []string{keyPairIDPrefix, placementGroupIDPrefix, launchTemplateIDPrefix} {
+		if strings.HasPrefix(id, prefix) {
+			return prefix
+		}
+	}
+	return ""
+}

spinifex/domains/ec2/accountcleanup/cleanup_test.go

@@ -0,0 +1,777 @@
+package accountcleanup
+
+//test:in-package — the clock is injected to place residue outside the settle
+// window without waiting for it.
+
+import (
+	"context"
+	"encoding/base64"
+	"encoding/json"
+	"errors"
+	"strings"
+	"testing"
+	"time"
+
+	"github.com/aws/aws-sdk-go/aws"
+	"github.com/aws/aws-sdk-go/service/ec2"
+	"github.com/aws/aws-sdk-go/service/s3"
+	ec2v1 "github.com/mulgadc/spinifex/contracts/ec2/v1"
+	testutil "github.com/mulgadc/spinifex/internal/testkit"
+	"github.com/mulgadc/spinifex/spinifex/bootstrap/config"
+	ec2key "github.com/mulgadc/spinifex/spinifex/domains/ec2/key"
+	ec2launchtemplate "github.com/mulgadc/spinifex/spinifex/domains/ec2/launchtemplate"
+	ec2placementgroup "github.com/mulgadc/spinifex/spinifex/domains/ec2/placementgroup"
+	ec2spotinstance "github.com/mulgadc/spinifex/spinifex/domains/ec2/spotinstance"
+	ec2tags "github.com/mulgadc/spinifex/spinifex/domains/ec2/tags"
+	awserrors "github.com/mulgadc/spinifex/spinifex/foundation/aws/errors"
+	"github.com/mulgadc/spinifex/spinifex/providers/objectstore"
+	"github.com/nats-io/nats.go/jetstream"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+const (
+	testAccount  = "111122223333"
+	otherAccount = "999988887777"
+	testBucket   = "test-bucket"
+	testPubKey   = "ssh-ed25519 AAAAC3NzaC1lZDI1NTE5AAAAIOMqqnkVzrm0SdG6UOoqKLsabgH5C9okWi0dh2l9GKJl"
+	memberID     = "i-0123456789abcdef0"
+	staleTagID   = "pg-0stale000000000"
+)
+
+var errInjected = errors.New("injected failure")
+
+// faultyStore fails object deletes its predicate selects.
+type faultyStore struct {
+	*objectstore.MemoryObjectStore
+
+	failDelete func(key string) bool
+}
+
+var _ objectstore.ObjectStore = (*faultyStore)(nil)
+
+func (f *faultyStore) DeleteObject(ctx context.Context, in *s3.DeleteObjectInput) (*s3.DeleteObjectOutput, error) {
+	if f.failDelete != nil && f.failDelete(aws.StringValue(in.Key)) {
+		return nil, errInjected
+	}
+	return f.MemoryObjectStore.DeleteObject(ctx, in)
+}
+
+// owners are the real EC2 owner services over an embedded JetStream, wired to
+// the tag index as the daemon wires them.
+type owners struct {
+	store     *faultyStore
+	keys      *ec2key.KeyServiceImpl
+	groups    *ec2placementgroup.PlacementGroupServiceImpl
+	templates *ec2launchtemplate.LaunchTemplateServiceImpl
+	tags      *ec2tags.TagsServiceImpl
+	spot      *ec2spotinstance.SpotInstanceServiceImpl
+	instances *fakeInstances
+	js        jetstream.JetStream
+}
+
+// fakeInstances stands in for the cluster-wide instance listing, which needs
+// every daemon to answer; its own refusal is covered separately.
+type fakeInstances struct {
+	refs *References
+	err  error
+}
+
+func (f *fakeInstances) AccountInstanceReferences(context.Context, string) (*References, error) {
+	if f.err != nil {
+		return nil, f.err
+	}
+	return f.refs, nil
+}
+
+func newOwners(t *testing.T) *owners {
+	t.Helper()
+	_, nc, js := testutil.StartTestJetStream(t)
+	o := &owners{store: &faultyStore{MemoryObjectStore: objectstore.NewMemoryObjectStore()}}
+	tagsKV, err := ec2tags.GetOrCreateTagsBucket(t.Context(), js)
+	require.NoError(t, err)
+	o.tags = ec2tags.NewTagsServiceImplWithStore(&config.Config{Predastore: config.PredastoreConfig{Bucket: testBucket}}, o.store, tagsKV)
+	o.keys = ec2key.NewKeyServiceImplWithStore(o.store, testBucket)
+	o.keys.SetCentralTagStore(o.tags)
+	o.groups, err = ec2placementgroup.NewPlacementGroupServiceImplWithNATS(t.Context(), nil, nc)
+	require.NoError(t, err)
+	o.groups.SetCentralTagStore(o.tags)
+	o.templates, err = ec2launchtemplate.NewLaunchTemplateServiceImplWithNATS(t.Context(), nil, nc)
+	require.NoError(t, err)
+	o.templates.SetCentralTagStore(o.tags)
+	o.spot, err = ec2spotinstance.NewSpotInstanceServiceImplWithNATS(t.Context(), nil, nc)
+	require.NoError(t, err)
+	o.instances = &fakeInstances{refs: newReferences()}
+	o.js = js
+	return o
+}
+
+// readers are the real spot and launch-template checks and the instance fake.
+func (o *owners) readers() ReferenceReaders {
+	return ReferenceReaders{Instances: o.instances, Spot: o.spot, Templates: o.templates}
+}
+
+func (o *owners) newCleanup() *Cleanup {
+	return New(o.keys, o.groups, o.templates, o.tags, o.readers())
+}
+
+// cleanup runs over the owners with residue of any age treated as settled.
+func (o *owners) cleanup(t *testing.T) *Result {
+	t.Helper()
+	return run(t, o.newCleanup())
+}
+
+func run(t *testing.T, c *Cleanup) *Result {
+	t.Helper()
+	c.now = func() time.Time { return time.Now().Add(time.Hour) }
+	res, err := c.CleanupPlatform(t.Context(), &Request{AccountID: testAccount, CorrelationRef: "ref-1"})
+	require.NoError(t, err)
+	return res
+}
+
+func (o *owners) importKey(t *testing.T, account, name string) string {
+	t.Helper()
+	out, err := o.keys.ImportKeyPair(t.Context(), &ec2.ImportKeyPairInput{
+		KeyName: aws.String(name), PublicKeyMaterial: []byte(testPubKey),
+		TagSpecifications: []*ec2.TagSpecification{{ResourceType: aws.String("key-pair"), Tags: secretTags()}},
+	}, account)
+	require.NoError(t, err)
+	return aws.StringValue(out.KeyPairId)
+}
+
+func (o *owners) createGroup(t *testing.T, account, name string) string {
+	t.Helper()
+	out, err := o.groups.CreatePlacementGroup(t.Context(), &ec2.CreatePlacementGroupInput{
+		GroupName: aws.String(name), Strategy: aws.String(ec2.PlacementStrategyCluster),
+	}, account)
+	require.NoError(t, err)
+	id := aws.StringValue(out.PlacementGroup.GroupId)
+	o.tag(t, account, id)
+	return id
+}
+
+func (o *owners) addMember(t *testing.T, group string) {
+	t.Helper()
+	reserved, err := o.groups.ReserveClusterNode(t.Context(), &ec2placementgroup.ReserveClusterNodeInput{
+		GroupName: group, EligibleNodes: []string{"node-1"},
+	}, testAccount)
+	require.NoError(t, err)
+	_, err = o.groups.FinalizeClusterInstances(t.Context(), &ec2placementgroup.FinalizeClusterInstancesInput{
+		GroupName: group, NodeInstances: map[string][]string{reserved.TargetNode: {memberID}},
+	}, testAccount)
+	require.NoError(t, err)
+}
+
+func (o *owners) removeMember(t *testing.T, group string) {
+	t.Helper()
+	_, err := o.groups.RemoveInstance(t.Context(), &ec2placementgroup.RemoveInstanceInput{
+		GroupName: group, NodeName: "node-1", InstanceID: memberID,
+	}, testAccount)
+	require.NoError(t, err)
+}
+
+func (o *owners) createTemplate(t *testing.T, account, name string) string {
+	t.Helper()
+	out, err := o.templates.CreateLaunchTemplate(t.Context(), &ec2.CreateLaunchTemplateInput{
+		LaunchTemplateName: aws.String(name),
+		LaunchTemplateData: &ec2.RequestLaunchTemplateData{
+			ImageId:  aws.String("ami-123"),
+			UserData: aws.String(base64.StdEncoding.EncodeToString([]byte("secret-user-data"))),
+		},
+	}, account)
+	require.NoError(t, err)
+	id := aws.StringValue(out.LaunchTemplate.LaunchTemplateId)
+	o.tag(t, account, id)
+	return id
+}
+
+func secretTags() []*ec2.Tag {
+	return []*ec2.Tag{{Key: aws.String("secret-tag-key"), Value: aws.String("secret-tag-value")}}
+}
+
+func (o *owners) tag(t *testing.T, account, id string) {
+	t.Helper()
+	_, err := o.tags.CreateTags(t.Context(), &ec2.CreateTagsInput{Resources: []*string{aws.String(id)}, Tags: secretTags()}, account)
+	require.NoError(t, err)
+}
+
+func (o *owners) putObject(t *testing.T, key, body string) {
+	t.Helper()
+	_, err := o.store.PutObject(t.Context(), &s3.PutObjectInput{
+		Bucket: aws.String(testBucket), Key: aws.String("keys/" + testAccount + "/" + key), Body: strings.NewReader(body),
+	})
+	require.NoError(t, err)
+}
+
+func statuses(res *Result) map[string]string {
+	out := map[string]string{}
+	for _, k := range res.Kinds {
+		out[k.Kind] = k.Status
+	}
+	return out
+}
+
+func allStatus(status string) map[string]string {
+	return map[string]string{
+		ec2v1.PlatformKindKeyPair: status, ec2v1.PlatformKindPlacementGroup: status,
+		ec2v1.PlatformKindLaunchTemplate: status, ec2v1.PlatformKindTagIndex: status,
+	}
+}
+
+func item(id, kind, disposition, reason string) Item {
+	return Item{ResourceID: id, Kind: kind, Disposition: disposition, Reason: reason}
+}
+
+// One run deletes every unreferenced key pair and empty placement group with
+// its tag-index entry, removes settled hidden residue and orphaned tags,
+// blocks on a group holding an instance, and leaves another account alone.
+// The next run, once the member leaves, converges to clean.
+func TestCleanupPlatformDeletesBlocksAndConverges(t *testing.T) {
+	o := newOwners(t)
+	keyID := o.importKey(t, testAccount, "secret-key-name")
+	emptyID := o.createGroup(t, testAccount, "secret-empty-group")
+	heldID := o.createGroup(t, testAccount, "secret-held-group")
+	o.addMember(t, "secret-held-group")
+	o.putObject(t, "secret-orphan-key", testPubKey)
+	o.tag(t, testAccount, staleTagID)
+	theirKey := o.importKey(t, otherAccount, "theirs")
+	o.createGroup(t, otherAccount, "theirs")
+	o.createTemplate(t, otherAccount, "theirs")
+
+	res := o.cleanup(t)
+
+	assert.Equal(t, testAccount, res.AccountID)
+	assert.Equal(t, "ref-1", res.CorrelationRef)
+	assert.ElementsMatch(t, []Item{
+		item(keyID, ec2v1.PlatformKindKeyPair, DispositionDeleted, ""),
+		item(emptyID, ec2v1.PlatformKindPlacementGroup, DispositionDeleted, ""),
+		{ResourceID: heldID, Kind: ec2v1.PlatformKindPlacementGroup, Disposition: DispositionBlocked, Reason: ReasonInstanceMembers, MemberCount: 1},
+		item(staleTagID, ec2v1.PlatformKindTagIndex, DispositionDeleted, ""),
+	}, res.Items)
+	assert.Equal(t, []Count{{
+		Kind: ec2v1.PlatformKindKeyPair, Class: ec2v1.PlatformClassPublicKeyWithoutMetadata, Disposition: DispositionDeleted, Count: 1,
+	}}, res.Counts)
+	// The held group's own tags are its, not residue.
+	assert.Equal(t, map[string]string{
+		ec2v1.PlatformKindKeyPair:        ec2v1.PlatformStatusObservedClean,
+		ec2v1.PlatformKindPlacementGroup: ec2v1.PlatformStatusRecordsFound,
+		ec2v1.PlatformKindLaunchTemplate: ec2v1.PlatformStatusObservedClean,
+		ec2v1.PlatformKindTagIndex:       ec2v1.PlatformStatusRecordsFound,
+	}, statuses(res))
+
+	theirs, err := o.keys.ScanAccount(t.Context(), otherAccount)
+	require.NoError(t, err)
+	assert.Equal(t, []string{theirKey}, theirs.LiveKeyPairIDs)
+	theirTags, err := o.tags.ScanAccountResourceIDs(t.Context(), otherAccount, "key-", "pg-", "lt-")
+	require.NoError(t, err)
+	assert.Len(t, theirTags, 3)
+
+	o.removeMember(t, "secret-held-group")
+	res = o.cleanup(t)
+	assert.Equal(t, []Item{item(heldID, ec2v1.PlatformKindPlacementGroup, DispositionDeleted, "")}, res.Items,
+		"items finished on an earlier run are not reported again")
+	assert.Empty(t, res.Counts)
+	assert.Equal(t, allStatus(ec2v1.PlatformStatusObservedClean), statuses(res))
+
+	res = o.cleanup(t)
+	assert.Empty(t, res.Items)
+	assert.Equal(t, allStatus(ec2v1.PlatformStatusObservedClean), statuses(res))
+}
+
+// The result carries IDs, kinds, dispositions, fixed reasons and counts only:
+// no names, tags, key material, user data, member instance IDs or error text.
+func TestCleanupPlatformResultDisclosesNoTenantData(t *testing.T) {
+	o := newOwners(t)
+	o.importKey(t, testAccount, "secret-key-name")
+	o.createGroup(t, testAccount, "secret-held-group")
+	o.addMember(t, "secret-held-group")
+	o.createTemplate(t, testAccount, "secret-template-name")
+	o.putObject(t, "secret-orphan-key", testPubKey)
+	o.putObject(t, "key-0badbadbadbadbad0.json", "{secret-broken")
+	failing := o.importKey(t, testAccount, "secret-failing-key")
+	o.store.failDelete = func(key string) bool { return strings.HasSuffix(key, failing+".json") }
+	o.importKey(t, testAccount, "secret-instance-key")
+	o.instances.refs.addKeys("secret-instance-key")
+	o.importKey(t, testAccount, "secret-spot-key")
+	o.openSpotRequest(t, "sir-0secret000000001", "secret-spot-key", "")
+	o.importKey(t, testAccount, "secret-template-key")
+	_, err := o.templates.CreateLaunchTemplate(t.Context(), &ec2.CreateLaunchTemplateInput{
+		LaunchTemplateName: aws.String("secret-referencing-template"),
+		LaunchTemplateData: &ec2.RequestLaunchTemplateData{KeyName: aws.String("secret-template-key")},
+	}, testAccount)
+	require.NoError(t, err)
+
+	body, err := json.Marshal(o.cleanup(t))
+	require.NoError(t, err)
+	for _, leak := range []string{
+		"secret", "AAAAC3Nz", memberID, "node-1", "injected", "sir-",
+		base64.StdEncoding.EncodeToString([]byte("secret-user-data")),
+	} {
+		assert.NotContains(t, string(body), leak)
+	}
+	assert.Contains(t, string(body), failing, "the IDs themselves are reported")
+}
+
+// A delete that fails part way is failed, not deleted; the re-run completes
+// the partly deleted key pair and reports it deleted exactly once.
+func TestCleanupPlatformInterruptedRunConverges(t *testing.T) {
+	o := newOwners(t)
+	keyID := o.importKey(t, testAccount, "deploy")
+	o.store.failDelete = func(key string) bool { return strings.HasSuffix(key, ".json") }
+
+	res := o.cleanup(t)
+	assert.Equal(t, []Item{item(keyID, ec2v1.PlatformKindKeyPair, DispositionFailed, ReasonOwnerDeleteFailed)}, res.Items)
+	assert.Equal(t, ec2v1.PlatformStatusRecordsFound, statuses(res)[ec2v1.PlatformKindKeyPair])
+	scan, err := o.keys.ScanAccount(t.Context(), testAccount)
+	require.NoError(t, err)
+	assert.Equal(t, []string{keyID}, scan.LiveKeyPairIDs, "the metadata anchors the unfinished delete")
+
+	o.store.failDelete = nil
+	res = o.cleanup(t)
+	assert.Equal(t, []Item{item(keyID, ec2v1.PlatformKindKeyPair, DispositionDeleted, "")}, res.Items)
+	assert.Equal(t, allStatus(ec2v1.PlatformStatusObservedClean), statuses(res))
+}
+
+// Records the owner cannot read are never deleted and stay reported as residue,
+// and orphaned tags whose owner may be one of them are left.
+func TestCleanupPlatformLeavesUnreadableRecords(t *testing.T) {
+	o := newOwners(t)
+	o.putObject(t, "key-0badbadbadbadbad0.json", "{not json")
+	o.putObject(t, "orphan", testPubKey)
+
+	res := o.cleanup(t)
+	assert.Empty(t, res.Items)
+	assert.ElementsMatch(t, []Count{
+		{Kind: ec2v1.PlatformKindKeyPair, Class: ec2v1.PlatformClassUnreadableRecord, Disposition: DispositionResidue, Reason: ReasonUnreadableRecord, Count: 1},
+		{Kind: ec2v1.PlatformKindKeyPair, Class: ec2v1.PlatformClassPublicKeyUnresolved, Disposition: DispositionResidue, Reason: ReasonOwnerRecordsUnresolved, Count: 1},
+	}, res.Counts)
+	assert.Equal(t, ec2v1.PlatformStatusRecordsFound, statuses(res)[ec2v1.PlatformKindKeyPair])
+	scan, err := o.keys.ScanAccount(t.Context(), testAccount)
+	require.NoError(t, err)
+	assert.Equal(t, []string{"key-0badbadbadbadbad0"}, scan.UnreadableKeyPairIDs)
+}
+
+// Residue younger than the settle window may be a create in flight: it is
+// reported, not removed.
+func TestCleanupPlatformKeepsUnsettledResidue(t *testing.T) {
+	o := newOwners(t)
+	o.putObject(t, "orphan", testPubKey)
+	c := o.newCleanup()
+
+	res, err := c.CleanupPlatform(t.Context(), &Request{AccountID: testAccount})
+	require.NoError(t, err)
+	assert.Equal(t, []Count{{
+		Kind: ec2v1.PlatformKindKeyPair, Class: ec2v1.PlatformClassPublicKeyWithoutMetadata,
+		Disposition: DispositionResidue, Reason: ReasonRecentlyWritten, Count: 1,
+	}}, res.Counts)
+}
+
+// lyingKeys reports every key pair deleted without deleting anything.
+type lyingKeys struct{ *ec2key.KeyServiceImpl }
+
+func (l lyingKeys) CleanupAccount(ctx context.Context, account string, _ time.Time, _ func(string) string) (*ec2key.AccountCleanup, error) {
+	scan, err := l.ScanAccount(ctx, account)
+	if err != nil {
+		return nil, err
+	}
+	return &ec2key.AccountCleanup{DeletedKeyPairIDs: scan.LiveKeyPairIDs}, nil
+}
+
+// An owner that reports a delete is not believed: the re-scan still finds the
+// record, so it is failed and still present.
+func TestCleanupPlatformProvesDeletesByRescan(t *testing.T) {
+	o := newOwners(t)
+	keyID := o.importKey(t, testAccount, "deploy")
+
+	res := run(t, New(lyingKeys{o.keys}, o.groups, o.templates, o.tags, o.readers()))
+	assert.Contains(t, res.Items, item(keyID, ec2v1.PlatformKindKeyPair, DispositionFailed, ReasonStillPresent))
+	assert.Equal(t, ec2v1.PlatformStatusRecordsFound, statuses(res)[ec2v1.PlatformKindKeyPair])
+}
+
+// failingRescan fails every scan after the first, so the cleanup deletes but
+// cannot prove it.
+type failingRescan struct {
+	*ec2launchtemplate.LaunchTemplateServiceImpl
+
+	scans int
+}
+
+func (f *failingRescan) ScanAccount(ctx context.Context, account string) (*ec2launchtemplate.AccountScan, error) {
+	if f.scans++; f.scans > 1 {
+		return nil, errInjected
+	}
+	return f.LaunchTemplateServiceImpl.ScanAccount(ctx, account)
+}
+
+// A failed re-scan is incomplete, never clean: the template stays blocked, and
+// a tag of that kind no reported resource owns stays, as residue.
+func TestCleanupPlatformFailedRescanIsIncomplete(t *testing.T) {
+	o := newOwners(t)
+	ltID := o.createTemplate(t, testAccount, "web")
+	o.tag(t, testAccount, "lt-0stale000000000")
+
+	res := run(t, New(o.keys, o.groups, &failingRescan{LaunchTemplateServiceImpl: o.templates}, o.tags, o.readers()))
+	assert.Equal(t, []Item{
+		item(ltID, ec2v1.PlatformKindLaunchTemplate, DispositionBlocked, ReasonReferenceUnknown),
+		item("lt-0stale000000000", ec2v1.PlatformKindTagIndex, DispositionResidue, ReasonOwnerRecordsUnresolved),
+	}, res.Items)
+	assert.Equal(t, ec2v1.PlatformStatusScanIncomplete, statuses(res)[ec2v1.PlatformKindLaunchTemplate])
+	assert.Equal(t, ec2v1.PlatformStatusScanIncomplete, statuses(res)[ec2v1.PlatformKindTagIndex])
+	tagged, err := o.tags.ScanAccountResourceIDs(t.Context(), testAccount, "lt-")
+	require.NoError(t, err)
+	assert.ElementsMatch(t, []string{"lt-0stale000000000", ltID}, tagged)
+}
+
+// failingTags fails every tag-index clear.
+type failingTags struct{ *ec2tags.TagsServiceImpl }
+
+func (failingTags) DeleteAllTags(context.Context, string, string) error { return errInjected }
+
+// A tag-index entry that cannot be cleared is failed and the kind keeps
+// records, never clean.
+func TestCleanupPlatformReportsUnclearedTags(t *testing.T) {
+	o := newOwners(t)
+	o.tag(t, testAccount, staleTagID)
+
+	res := run(t, New(o.keys, o.groups, o.templates, failingTags{o.tags}, o.readers()))
+	assert.Equal(t, []Item{item(staleTagID, ec2v1.PlatformKindTagIndex, DispositionFailed, ReasonOwnerDeleteFailed)}, res.Items)
+	assert.Equal(t, ec2v1.PlatformStatusRecordsFound, statuses(res)[ec2v1.PlatformKindTagIndex])
+}
+
+// failingCleanup fails the owner cleanup's own listing.
+type failingCleanup struct {
+	*ec2placementgroup.PlacementGroupServiceImpl
+}
+
+func (failingCleanup) CleanupAccount(context.Context, string, func(string, string) string) (*ec2placementgroup.AccountCleanup, error) {
+	return nil, errInjected
+}
+
+// An owner cleanup that cannot list deletes nothing and is incomplete; the
+// group it could not reach is still reported, from the re-scan.
+func TestCleanupPlatformOwnerListingFailureIsIncomplete(t *testing.T) {
+	o := newOwners(t)
+	groupID := o.createGroup(t, testAccount, "web")
+
+	res := run(t, New(o.keys, failingCleanup{o.groups}, o.templates, o.tags, o.readers()))
+	assert.Equal(t, ec2v1.PlatformStatusScanIncomplete, statuses(res)[ec2v1.PlatformKindPlacementGroup])
+	assert.Contains(t, res.Items, item(groupID, ec2v1.PlatformKindPlacementGroup, DispositionResidue, ReasonAppearedDuringCleanup))
+}
+
+func TestCleanupPlatformValidatesTheRequest(t *testing.T) {
+	c := New(nil, nil, nil, nil, ReferenceReaders{})
+	for name, tc := range map[string]struct {
+		req  *Request
+		code string
+	}{
+		"nil":       {nil, awserrors.ErrorMissingParameter},
+		"empty":     {&Request{}, awserrors.ErrorMissingParameter},
+		"malformed": {&Request{AccountID: "acct-1"}, awserrors.ErrorInvalidParameterValue},
+	} {
+		t.Run(name, func(t *testing.T) {
+			_, err := c.CleanupPlatform(t.Context(), tc.req)
+			require.Error(t, err)
+			got, _ := awserrors.ResolveErrorCode(err)
+			assert.Equal(t, tc.code, got)
+		})
+	}
+}
+
+func (o *owners) openSpotRequest(t *testing.T, id, keyName, groupName string) {
+	t.Helper()
+	spec := &ec2.LaunchSpecification{ImageId: aws.String("ami-123"), KeyName: aws.String(keyName)}
+	if groupName != "" {
+		spec.Placement = &ec2.SpotPlacement{GroupName: aws.String(groupName)}
+	}
+	_, err := o.spot.PutSpotInstanceRequests(t.Context(), &ec2spotinstance.PutSpotRequestsInput{Requests: []*ec2.SpotInstanceRequest{{
+		SpotInstanceRequestId: aws.String(id), State: aws.String(ec2.SpotInstanceStateOpen), LaunchSpecification: spec,
+	}}}, testAccount)
+	require.NoError(t, err)
+}
+
+func (o *owners) putTemplateRecord(t *testing.T, key, value string) {
+	t.Helper()
+	kv, err := o.js.KeyValue(t.Context(), ec2launchtemplate.KVBucketLaunchTemplates)
+	require.NoError(t, err)
+	_, err = kv.Put(t.Context(), key, []byte(value))
+	require.NoError(t, err)
+}
+
+// A live launch template is never deleted: an EKS nodegroup may name it and
+// EC2 cannot read EKS records. It stays blocked, described and tagged on every
+// run; headerless version bodies and names, which nothing can resolve, are
+// removed once settled.
+func TestCleanupPlatformNeverDeletesLiveLaunchTemplates(t *testing.T) {
+	o := newOwners(t)
+	ltID := o.createTemplate(t, testAccount, "web")
+	o.putTemplateRecord(t, testAccount+".lt-0orphan000000000.v1", `{"version_number":1}`)
+
+	for range 2 {
+		res := o.cleanup(t)
+		assert.Equal(t, []Item{item(ltID, ec2v1.PlatformKindLaunchTemplate, DispositionBlocked, ReasonReferenceUnknown)}, res.Items)
+		assert.Equal(t, ec2v1.PlatformStatusRecordsFound, statuses(res)[ec2v1.PlatformKindLaunchTemplate])
+		assert.Equal(t, ec2v1.PlatformStatusRecordsFound, statuses(res)[ec2v1.PlatformKindTagIndex], "the template keeps its tags")
+	}
+	scan, err := o.templates.ScanAccount(t.Context(), testAccount)
+	require.NoError(t, err)
+	assert.Equal(t, []string{ltID}, scan.LiveTemplateIDs)
+	assert.Zero(t, scan.OrphanVersionBodies)
+	out, err := o.templates.DescribeLaunchTemplateVersions(t.Context(), &ec2.DescribeLaunchTemplateVersionsInput{LaunchTemplateId: aws.String(ltID)}, testAccount)
+	require.NoError(t, err)
+	assert.Len(t, out.LaunchTemplateVersions, 1, "the version body is untouched")
+}
+
+// A key pair a non-terminated instance names is blocked and kept, and so is an
+// orphan public key the instance names, because IMDS serves it by name. A
+// placement group a stopped instance names is blocked though it has no members.
+func TestCleanupPlatformBlocksOnInstanceReferences(t *testing.T) {
+	o := newOwners(t)
+	usedKey := o.importKey(t, testAccount, "in-use")
+	freeKey := o.importKey(t, testAccount, "free")
+	o.putObject(t, "served-orphan", testPubKey)
+	usedGroup := o.createGroup(t, testAccount, "stopped-member")
+	o.instances.refs.addKeys("in-use", "served-orphan")
+	o.instances.refs.addGroups("stopped-member")
+
+	res := o.cleanup(t)
+	assert.ElementsMatch(t, []Item{
+		item(usedKey, ec2v1.PlatformKindKeyPair, DispositionBlocked, ReasonInstanceReference),
+		item(freeKey, ec2v1.PlatformKindKeyPair, DispositionDeleted, ""),
+		item(usedGroup, ec2v1.PlatformKindPlacementGroup, DispositionBlocked, ReasonInstanceReference),
+	}, res.Items)
+	assert.Equal(t, []Count{{
+		Kind: ec2v1.PlatformKindKeyPair, Class: ec2v1.PlatformClassPublicKeyWithoutMetadata,
+		Disposition: DispositionBlocked, Reason: ReasonInstanceReference, Count: 1,
+	}}, res.Counts)
+	scan, err := o.keys.ScanAccount(t.Context(), testAccount)
+	require.NoError(t, err)
+	assert.Equal(t, []string{usedKey}, scan.LiveKeyPairIDs)
+	assert.Equal(t, 1, scan.OrphanPublicKeys)
+	material, err := o.keys.GetPublicKeyMaterial(testAccount, "served-orphan")
+	require.NoError(t, err)
+	assert.Equal(t, testPubKey, material)
+}
+
+// An open spot request's launch specification blocks the key pair and the
+// placement group it names until the request is cancelled.
+func TestCleanupPlatformBlocksOnSpotRequests(t *testing.T) {
+	o := newOwners(t)
+	keyID := o.importKey(t, testAccount, "spot-key")
+	groupID := o.createGroup(t, testAccount, "spot-group")
+	o.openSpotRequest(t, "sir-0000000000000001", "spot-key", "spot-group")
+
+	res := o.cleanup(t)
+	assert.ElementsMatch(t, []Item{
+		item(keyID, ec2v1.PlatformKindKeyPair, DispositionBlocked, ReasonSpotRequestReference),
+		item(groupID, ec2v1.PlatformKindPlacementGroup, DispositionBlocked, ReasonSpotRequestReference),
+	}, res.Items)
+
+	_, err := o.spot.CancelSpotInstanceRequests(t.Context(), &ec2.CancelSpotInstanceRequestsInput{
+		SpotInstanceRequestIds: []*string{aws.String("sir-0000000000000001")},
+	}, testAccount)
+	require.NoError(t, err)
+	res = o.cleanup(t)
+	assert.ElementsMatch(t, []Item{
+		item(keyID, ec2v1.PlatformKindKeyPair, DispositionDeleted, ""),
+		item(groupID, ec2v1.PlatformKindPlacementGroup, DispositionDeleted, ""),
+	}, res.Items)
+}
+
+// failingTemplates fails the launch-template reference check.
+type failingTemplates struct{}
+
+func (failingTemplates) AccountTemplateReferences(context.Context, string) (*ec2launchtemplate.TemplateReferences, error) {
+	return nil, errInjected
+}
+
+// failingSpot fails the spot reference check.
+type failingSpot struct{}
+
+func (failingSpot) AccountLaunchReferences(context.Context, string) (*ec2spotinstance.LaunchReferences, error) {
+	return nil, errInjected
+}
+
+// A reference check that cannot complete is never read as "no references":
+// key pairs, orphan public keys and placement groups are all kept, both kinds
+// are incomplete, and launch-template residue, which no check guards, is
+// still removed.
+func TestCleanupPlatformFailedReferenceCheckDeletesNothing(t *testing.T) {
+	for name, build := range map[string]func(o *owners) *Cleanup{
+		"instance listing fails": func(o *owners) *Cleanup {
+			o.instances.err = errInjected
+			return o.newCleanup()
+		},
+		"spot listing fails": func(o *owners) *Cleanup {
+			r := o.readers()
+			r.Spot = failingSpot{}
+			return New(o.keys, o.groups, o.templates, o.tags, r)
+		},
+		"launch-template reading fails": func(o *owners) *Cleanup {
+			r := o.readers()
+			r.Templates = failingTemplates{}
+			return New(o.keys, o.groups, o.templates, o.tags, r)
+		},
+		"instance reader absent": func(o *owners) *Cleanup {
+			r := o.readers()
+			r.Instances = nil
+			return New(o.keys, o.groups, o.templates, o.tags, r)
+		},
+		"launch-template reader absent": func(o *owners) *Cleanup {
+			r := o.readers()
+			r.Templates = nil
+			return New(o.keys, o.groups, o.templates, o.tags, r)
+		},
+	} {
+		t.Run(name, func(t *testing.T) {
+			o := newOwners(t)
+			keyID := o.importKey(t, testAccount, "deploy")
+			o.putObject(t, "orphan", testPubKey)
+			groupID := o.createGroup(t, testAccount, "web")
+			o.putTemplateRecord(t, testAccount+".lt-0orphan000000000.v1", `{"version_number":1}`)
+
+			res := run(t, build(o))
+			assert.ElementsMatch(t, []Item{
+				item(keyID, ec2v1.PlatformKindKeyPair, DispositionBlocked, ReasonReferenceCheckFailed),
+				item(groupID, ec2v1.PlatformKindPlacementGroup, DispositionBlocked, ReasonReferenceCheckFailed),
+			}, res.Items)
+			assert.ElementsMatch(t, []Count{
+				{Kind: ec2v1.PlatformKindKeyPair, Class: ec2v1.PlatformClassPublicKeyWithoutMetadata, Disposition: DispositionBlocked, Reason: ReasonReferenceCheckFailed, Count: 1},
+				{Kind: ec2v1.PlatformKindLaunchTemplate, Class: ec2v1.PlatformClassVersionBodyWithoutHeader, Disposition: DispositionDeleted, Count: 1},
+			}, res.Counts)
+			for _, k := range res.Kinds {
+				if k.Kind == ec2v1.PlatformKindKeyPair || k.Kind == ec2v1.PlatformKindPlacementGroup {
+					assert.Equal(t, ec2v1.PlatformStatusScanIncomplete, k.Status, k.Kind)
+					assert.Equal(t, []string{ReasonReferenceCheckFailed}, k.IncompleteReasons, k.Kind)
+				}
+			}
+			keys, err := o.keys.ScanAccount(t.Context(), testAccount)
+			require.NoError(t, err)
+			assert.Equal(t, []string{keyID}, keys.LiveKeyPairIDs)
+			assert.Equal(t, 1, keys.OrphanPublicKeys)
+			groups, err := o.groups.ScanAccount(t.Context(), testAccount)
+			require.NoError(t, err)
+			assert.Len(t, groups.Groups, 1)
+		})
+	}
+}
+
+// A key pair or placement group named by any version of a live launch
+// template, by name or group ID, is blocked: RunInstances may resolve any
+// version by $Default, $Latest or number. The template itself stays blocked.
+func TestCleanupPlatformBlocksOnLaunchTemplateReferences(t *testing.T) {
+	o := newOwners(t)
+	usedKey := o.importKey(t, testAccount, "template-key")
+	freeKey := o.importKey(t, testAccount, "free")
+	byName := o.createGroup(t, testAccount, "template-group")
+	byID := o.createGroup(t, testAccount, "by-id")
+	ltID := o.createTemplate(t, testAccount, "web")
+	_, err := o.templates.CreateLaunchTemplateVersion(t.Context(), &ec2.CreateLaunchTemplateVersionInput{
+		LaunchTemplateId: aws.String(ltID),
+		LaunchTemplateData: &ec2.RequestLaunchTemplateData{
+			KeyName:   aws.String("template-key"),
+			Placement: &ec2.LaunchTemplatePlacementRequest{GroupName: aws.String("template-group")},
+		},
+	}, testAccount)
+	require.NoError(t, err)
+	_, err = o.templates.CreateLaunchTemplateVersion(t.Context(), &ec2.CreateLaunchTemplateVersionInput{
+		LaunchTemplateId:   aws.String(ltID),
+		LaunchTemplateData: &ec2.RequestLaunchTemplateData{Placement: &ec2.LaunchTemplatePlacementRequest{GroupId: aws.String(byID)}},
+	}, testAccount)
+	require.NoError(t, err)
+
+	res := o.cleanup(t)
+	assert.ElementsMatch(t, []Item{
+		item(usedKey, ec2v1.PlatformKindKeyPair, DispositionBlocked, ReasonLaunchTemplateReference),
+		item(freeKey, ec2v1.PlatformKindKeyPair, DispositionDeleted, ""),
+		item(byName, ec2v1.PlatformKindPlacementGroup, DispositionBlocked, ReasonLaunchTemplateReference),
+		item(byID, ec2v1.PlatformKindPlacementGroup, DispositionBlocked, ReasonLaunchTemplateReference),
+		item(ltID, ec2v1.PlatformKindLaunchTemplate, DispositionBlocked, ReasonReferenceUnknown),
+	}, res.Items)
+	keys, err := o.keys.ScanAccount(t.Context(), testAccount)
+	require.NoError(t, err)
+	assert.Equal(t, []string{usedKey}, keys.LiveKeyPairIDs)
+	groups, err := o.groups.ScanAccount(t.Context(), testAccount)
+	require.NoError(t, err)
+	assert.Len(t, groups.Groups, 2)
+}
+
+// A live template version the owner cannot decode may name any key pair or
+// group, so both kinds are blocked and incomplete and nothing is deleted.
+func TestCleanupPlatformUndecodableTemplateVersionDeletesNothing(t *testing.T) {
+	o := newOwners(t)
+	keyID := o.importKey(t, testAccount, "deploy")
+	groupID := o.createGroup(t, testAccount, "web")
+	ltID := o.createTemplate(t, testAccount, "web")
+	o.putTemplateRecord(t, testAccount+"."+ltID+".v2", "{not json")
+
+	res := o.cleanup(t)
+	assert.Contains(t, res.Items, item(keyID, ec2v1.PlatformKindKeyPair, DispositionBlocked, ReasonReferenceCheckFailed))
+	assert.Contains(t, res.Items, item(groupID, ec2v1.PlatformKindPlacementGroup, DispositionBlocked, ReasonReferenceCheckFailed))
+	assert.Equal(t, ec2v1.PlatformStatusScanIncomplete, statuses(res)[ec2v1.PlatformKindKeyPair])
+	assert.Equal(t, ec2v1.PlatformStatusScanIncomplete, statuses(res)[ec2v1.PlatformKindPlacementGroup])
+	keys, err := o.keys.ScanAccount(t.Context(), testAccount)
+	require.NoError(t, err)
+	assert.Equal(t, []string{keyID}, keys.LiveKeyPairIDs)
+}
+
+// A tag-index entry whose owner re-scan holds an unreadable record is not
+// cleared, because that record may be its owner; it is residue, not deleted.
+func TestCleanupPlatformUnresolvedOwnerKeepsTags(t *testing.T) {
+	for name, seed := range map[string]func(t *testing.T, o *owners) string{
+		"key pair": func(t *testing.T, o *owners) string {
+			o.putObject(t, "key-0badbadbadbadbad0.json", "{not json")
+			return "key-0stale000000000"
+		},
+		"placement group": func(t *testing.T, o *owners) string {
+			kv, err := o.js.KeyValue(t.Context(), ec2placementgroup.KVBucketPlacementGroups)
+			require.NoError(t, err)
+			_, err = kv.Put(t.Context(), testAccount+".broken", []byte("{not json"))
+			require.NoError(t, err)
+			return staleTagID
+		},
+		"launch template": func(t *testing.T, o *owners) string {
+			o.putTemplateRecord(t, testAccount+".lt-0broken000000000", "{not json")
+			return "lt-0stale000000000"
+		},
+	} {
+		t.Run(name, func(t *testing.T) {
+			o := newOwners(t)
+			stale := seed(t, o)
+			o.tag(t, testAccount, stale)
+
+			res := o.cleanup(t)
+			assert.Contains(t, res.Items, item(stale, ec2v1.PlatformKindTagIndex, DispositionResidue, ReasonOwnerRecordsUnresolved))
+			assert.Equal(t, ec2v1.PlatformStatusScanIncomplete, statuses(res)[ec2v1.PlatformKindTagIndex])
+			tagged, err := o.tags.ScanAccountResourceIDs(t.Context(), testAccount, stale[:3])
+			require.NoError(t, err)
+			assert.Equal(t, []string{stale}, tagged)
+		})
+	}
+}
+
+// Settled launch-template residue the owner fails to remove is reported
+// failed from the re-scan, never deleted.
+func TestCleanupPlatformResidueRemovalFailureIsFailed(t *testing.T) {
+	o := newOwners(t)
+	o.putTemplateRecord(t, testAccount+".lt-0orphan000000000.v1", `{"version_number":1}`)
+
+	res := run(t, New(o.keys, o.groups, failingResidue{o.templates}, o.tags, o.readers()))
+	assert.Equal(t, []Count{{
+		Kind: ec2v1.PlatformKindLaunchTemplate, Class: ec2v1.PlatformClassVersionBodyWithoutHeader,
+		Disposition: DispositionFailed, Reason: ReasonOwnerDeleteFailed, Count: 1,
+	}}, res.Counts)
+	assert.Equal(t, ec2v1.PlatformStatusRecordsFound, statuses(res)[ec2v1.PlatformKindLaunchTemplate])
+}
+
+// failingResidue reports every settled orphan as a failed removal and removes nothing.
+type failingResidue struct {
+	*ec2launchtemplate.LaunchTemplateServiceImpl
+}
+
+func (f failingResidue) CleanupAccountResidue(ctx context.Context, acct string, _ time.Time) (*ec2launchtemplate.AccountCleanup, error) {
+	scan, err := f.ScanAccount(ctx, acct)
+	if err != nil {
+		return nil, err
+	}
+	return &ec2launchtemplate.AccountCleanup{KeptTemplateIDs: scan.LiveTemplateIDs, OrphanBodiesFailed: scan.OrphanVersionBodies}, nil
+}

spinifex/domains/ec2/accountcleanup/kinds.go

@@ -0,0 +1,411 @@
+package accountcleanup
+
+import (
+	"context"
+	"log/slog"
+	"maps"
+	"slices"
+	"time"
+
+	ec2v1 "github.com/mulgadc/spinifex/contracts/ec2/v1"
+	ec2key "github.com/mulgadc/spinifex/spinifex/domains/ec2/key"
+	ec2launchtemplate "github.com/mulgadc/spinifex/spinifex/domains/ec2/launchtemplate"
+	ec2placementgroup "github.com/mulgadc/spinifex/spinifex/domains/ec2/placementgroup"
+)
+
+// Each kind runs scan, owner cleanup, then re-scan. The first scan only bounds
+// how much nameless residue a run can claim to have removed.
+
+type keyPairRun struct {
+	kind *kindRun
+	// unchecked: the reference check failed, so nothing was attempted.
+	unchecked bool
+	before    *ec2key.AccountScan
+	cleaned   *ec2key.AccountCleanup
+	did       attempts
+}
+
+// cleanKeyPairs deletes only when the reference check completed; otherwise
+// every key pair and orphan public key is kept and the kind is incomplete.
+func (c *Cleanup) cleanKeyPairs(ctx context.Context, acct string, settledBefore time.Time, refs referenceCheck) *keyPairRun {
+	run := &keyPairRun{kind: c.startKind(ec2v1.PlatformKindKeyPair), did: newAttempts()}
+	if before, err := c.keys.ScanAccount(ctx, acct); err == nil {
+		run.before = before
+	}
+	if !refs.complete {
+		run.unchecked = true
+		run.kind.incomplete(ReasonReferenceCheckFailed)
+		return run
+	}
+	cleaned, err := c.keys.CleanupAccount(ctx, acct, settledBefore, refs.keyReason)
+	if err != nil {
+		slog.WarnContext(ctx, "Platform cleanup: key-pair cleanup failed", "err", err)
+		run.kind.incomplete(ec2v1.PlatformReasonStoreListFailed)
+		return run
+	}
+	run.cleaned = cleaned
+	markAll(run.did.deleted, cleaned.DeletedKeyPairIDs)
+	markAll(run.did.failed, cleaned.FailedKeyPairIDs)
+	maps.Copy(run.did.kept, cleaned.Referenced)
+	return run
+}
+
+func (c *Cleanup) finishKeyPairs(ctx context.Context, acct string, run *keyPairRun, res *Result) ownedIDs {
+	defer c.finishKind(run.kind)
+	after, err := c.keys.ScanAccount(ctx, acct)
+	if err != nil {
+		slog.WarnContext(ctx, "Platform cleanup: key-pair re-scan failed", "err", err)
+		run.kind.incomplete(ec2v1.PlatformReasonStoreListFailed)
+		after = nil
+	}
+	present := map[string]bool{}
+	if after != nil {
+		markAll(present, after.LiveKeyPairIDs)
+		markAll(present, after.UnreadableKeyPairIDs)
+		if after.ReadFailures > 0 {
+			run.kind.incomplete(ec2v1.PlatformReasonRecordReadFailed)
+		}
+	}
+	var beforeLive, afterLive, deleted, failed []string
+	if run.before != nil {
+		beforeLive = run.before.LiveKeyPairIDs
+	}
+	if after != nil {
+		afterLive = after.LiveKeyPairIDs
+	}
+	if run.cleaned != nil {
+		deleted, failed = run.cleaned.DeletedKeyPairIDs, run.cleaned.FailedKeyPairIDs
+	}
+	if run.unchecked {
+		keepAll(run.did.kept, beforeLive, ReasonReferenceCheckFailed)
+		keepAll(run.did.kept, afterLive, ReasonReferenceCheckFailed)
+	}
+	for _, id := range sortedUnion(beforeLive, deleted, failed, afterLive) {
+		disposition, reason := itemDisposition(id, run.did, present[id], after != nil)
+		res.addItem(run.kind, id, disposition, reason, 0)
+	}
+	if after == nil {
+		return ownedIDs{}
+	}
+	res.addCount(run.kind, ec2v1.PlatformClassUnreadableRecord, DispositionResidue, ReasonUnreadableRecord, len(after.UnreadableKeyPairIDs))
+	res.addCount(run.kind, ec2v1.PlatformClassPublicKeyUnresolved, DispositionResidue, ReasonOwnerRecordsUnresolved, after.UnresolvedPublicKeys)
+	var before, removed, failedN, unsettled int
+	kept := map[string]int{}
+	if run.before != nil {
+		before = run.before.OrphanPublicKeys
+	}
+	if run.cleaned != nil {
+		removed, failedN, unsettled = run.cleaned.OrphansRemoved, run.cleaned.OrphansFailed, run.cleaned.OrphansUnsettled
+		kept = run.cleaned.OrphansReferenced
+	} else if run.unchecked {
+		kept[ReasonReferenceCheckFailed] = after.OrphanPublicKeys
+	}
+	res.residueCounts(run.kind, ec2v1.PlatformClassPublicKeyWithoutMetadata, before, after.OrphanPublicKeys, removed, failedN, unsettled, kept)
+	return ownedIDs{ids: present, resolved: after.ReadFailures == 0 && len(after.UnreadableKeyPairIDs) == 0}
+}
+
+type placementGroupRun struct {
+	kind      *kindRun
+	unchecked bool
+	before    *ec2placementgroup.AccountScan
+	cleaned   *ec2placementgroup.AccountCleanup
+	did       attempts
+}
+
+// cleanPlacementGroups deletes only when the reference check completed.
+func (c *Cleanup) cleanPlacementGroups(ctx context.Context, acct string, refs referenceCheck) *placementGroupRun {
+	run := &placementGroupRun{kind: c.startKind(ec2v1.PlatformKindPlacementGroup), did: newAttempts()}
+	if before, err := c.groups.ScanAccount(ctx, acct); err == nil {
+		run.before = before
+	}
+	if !refs.complete {
+		run.unchecked = true
+		run.kind.incomplete(ReasonReferenceCheckFailed)
+		return run
+	}
+	cleaned, err := c.groups.CleanupAccount(ctx, acct, refs.groupReason)
+	if err != nil {
+		slog.WarnContext(ctx, "Platform cleanup: placement-group cleanup failed", "err", err)
+		run.kind.incomplete(ec2v1.PlatformReasonStoreListFailed)
+		return run
+	}
+	run.cleaned = cleaned
+	markAll(run.did.deleted, cleaned.DeletedGroupIDs)
+	markAll(run.did.failed, cleaned.FailedGroupIDs)
+	for _, b := range cleaned.Blocked {
+		run.did.blocked[b.GroupID] = b.MemberCount
+	}
+	maps.Copy(run.did.kept, cleaned.Referenced)
+	return run
+}
+
+// finishPlacementGroups reports a group that still names instances as
+// blocked with its member count, from the re-scan when it can.
+func (c *Cleanup) finishPlacementGroups(ctx context.Context, acct string, run *placementGroupRun, res *Result) ownedIDs {
+	defer c.finishKind(run.kind)
+	after, err := c.groups.ScanAccount(ctx, acct)
+	if err != nil {
+		slog.WarnContext(ctx, "Platform cleanup: placement-group re-scan failed", "err", err)
+		run.kind.incomplete(ec2v1.PlatformReasonStoreListFailed)
+		after = nil
+	}
+	present := map[string]bool{}
+	members := map[string]int{}
+	var ids []string
+	if run.before != nil {
+		for _, g := range run.before.Groups {
+			ids = append(ids, g.GroupID)
+		}
+	}
+	if run.cleaned != nil {
+		ids = append(ids, run.cleaned.DeletedGroupIDs...)
+		ids = append(ids, run.cleaned.FailedGroupIDs...)
+		for _, b := range run.cleaned.Blocked {
+			ids = append(ids, b.GroupID)
+		}
+		for id := range run.cleaned.Referenced {
+			ids = append(ids, id)
+		}
+	}
+	if after != nil {
+		for _, g := range after.Groups {
+			present[g.GroupID] = true
+			members[g.GroupID] = g.MemberCount
+			ids = append(ids, g.GroupID)
+		}
+		if after.ReadFailures > 0 {
+			run.kind.incomplete(ec2v1.PlatformReasonRecordReadFailed)
+		}
+	}
+	if run.unchecked {
+		keepAll(run.did.kept, ids, ReasonReferenceCheckFailed)
+	}
+	for _, id := range sortedUnion(ids) {
+		if present[id] && (members[id] > 0 || run.did.blocked[id] > 0) {
+			count := members[id]
+			if count == 0 {
+				count = run.did.blocked[id]
+			}
+			res.addItem(run.kind, id, DispositionBlocked, ReasonInstanceMembers, count)
+			continue
+		}
+		disposition, reason := itemDisposition(id, run.did, present[id], after != nil)
+		res.addItem(run.kind, id, disposition, reason, 0)
+	}
+	if after == nil {
+		return ownedIDs{}
+	}
+	res.addCount(run.kind, ec2v1.PlatformClassUnreadableRecord, DispositionResidue, ReasonUnreadableRecord, after.Unreadable)
+	return ownedIDs{ids: present, resolved: after.ReadFailures == 0 && after.Unreadable == 0}
+}
+
+type launchTemplateRun struct {
+	kind    *kindRun
+	before  *ec2launchtemplate.AccountScan
+	cleaned *ec2launchtemplate.AccountCleanup
+	did     attempts
+}
+
+func (c *Cleanup) cleanLaunchTemplates(ctx context.Context, acct string, settledBefore time.Time) *launchTemplateRun {
+	run := &launchTemplateRun{kind: c.startKind(ec2v1.PlatformKindLaunchTemplate), did: newAttempts()}
+	if before, err := c.templates.ScanAccount(ctx, acct); err == nil {
+		run.before = before
+	}
+	cleaned, err := c.templates.CleanupAccountResidue(ctx, acct, settledBefore)
+	if err != nil {
+		slog.WarnContext(ctx, "Platform cleanup: launch-template residue cleanup failed", "err", err)
+		run.kind.incomplete(ec2v1.PlatformReasonStoreListFailed)
+		return run
+	}
+	run.cleaned = cleaned
+	keepAll(run.did.kept, cleaned.KeptTemplateIDs, ReasonReferenceUnknown)
+	return run
+}
+
+func (c *Cleanup) finishLaunchTemplates(ctx context.Context, acct string, run *launchTemplateRun, res *Result) ownedIDs {
+	defer c.finishKind(run.kind)
+	after, err := c.templates.ScanAccount(ctx, acct)
+	if err != nil {
+		slog.WarnContext(ctx, "Platform cleanup: launch-template re-scan failed", "err", err)
+		run.kind.incomplete(ec2v1.PlatformReasonStoreListFailed)
+		after = nil
+	}
+	present := map[string]bool{}
+	if after != nil {
+		markAll(present, after.LiveTemplateIDs)
+		markAll(present, after.UnreadableTemplateIDs)
+		if after.ReadFailures > 0 {
+			run.kind.incomplete(ec2v1.PlatformReasonRecordReadFailed)
+		}
+	}
+	var beforeLive, afterLive, kept []string
+	if run.before != nil {
+		beforeLive = run.before.LiveTemplateIDs
+	}
+	if after != nil {
+		afterLive = after.LiveTemplateIDs
+	}
+	if run.cleaned != nil {
+		kept = run.cleaned.KeptTemplateIDs
+	}
+	// Live templates are never deleted here, so every one still present is
+	// blocked on the unknown EKS nodegroup reference, whenever it appeared.
+	for _, id := range afterLive {
+		run.did.kept[id] = ReasonReferenceUnknown
+	}
+	for _, id := range sortedUnion(beforeLive, kept, afterLive) {
+		disposition, reason := itemDisposition(id, run.did, present[id], after != nil)
+		res.addItem(run.kind, id, disposition, reason, 0)
+	}
+	if after == nil {
+		return ownedIDs{}
+	}
+	res.addCount(run.kind, ec2v1.PlatformClassUnreadableRecord, DispositionResidue, ReasonUnreadableRecord,
+		len(after.UnreadableTemplateIDs)+after.UnreadableNameIndexEntries)
+	res.addCount(run.kind, ec2v1.PlatformClassUnrecognisedTemplateEntry, DispositionResidue, ReasonUnrecognisedRecord, after.UnrecognisedEntries)
+	var c0 ec2launchtemplate.AccountCleanup
+	if run.cleaned != nil {
+		c0 = *run.cleaned
+	}
+	var beforeBodies, beforeNames int
+	if run.before != nil {
+		beforeBodies, beforeNames = run.before.OrphanVersionBodies, run.before.OrphanNameIndexEntries
+	}
+	res.residueCounts(run.kind, ec2v1.PlatformClassVersionBodyWithoutHeader, beforeBodies, after.OrphanVersionBodies,
+		c0.OrphanBodiesRemoved, c0.OrphanBodiesFailed, c0.OrphanBodiesUnsettled, nil)
+	res.residueCounts(run.kind, ec2v1.PlatformClassNameIndexWithoutHeader, beforeNames, after.OrphanNameIndexEntries,
+		c0.OrphanNamesRemoved, c0.OrphanNamesFailed, c0.OrphanNamesUnsettled, nil)
+	return ownedIDs{ids: present, resolved: after.ReadFailures == 0 && len(after.UnreadableTemplateIDs) == 0 &&
+		after.UnreadableNameIndexEntries == 0 && after.UnrecognisedEntries == 0}
+}
+
+// cleanTagIndex clears entries whose owner re-scan resolved and holds no
+// record for the ID, then re-lists the index to prove each clear. Entries of
+// resources that still exist are theirs, not residue, and are left alone.
+func (c *Cleanup) cleanTagIndex(ctx context.Context, acct string, kind *kindRun, tagged []string, listed bool, owners map[string]ownedIDs, res *Result) KindReport {
+	if listed {
+		c.clearOrphanTags(ctx, acct, kind, tagged, owners, res)
+	}
+	c.finishKind(kind)
+	return kind.report
+}
+
+func (c *Cleanup) clearOrphanTags(ctx context.Context, acct string, kind *kindRun, tagged []string, owners map[string]ownedIDs, res *Result) {
+	// A resource this run reports as not deleted owns its tags even when its
+	// kind's re-scan failed.
+	reported := map[string]bool{}
+	for _, it := range res.Items {
+		if it.Disposition != DispositionDeleted {
+			reported[it.ResourceID] = true
+		}
+	}
+	var cleared []string
+	for _, id := range tagged {
+		owner := owners[idPrefix(id)]
+		switch {
+		case owner.ids[id], reported[id]:
+			kind.found = true
+		case !owner.resolved:
+			kind.incomplete(ec2v1.PlatformReasonOwnerRecordsUnresolved)
+			res.addItem(kind, id, DispositionResidue, ReasonOwnerRecordsUnresolved, 0)
+		default:
+			if err := c.tags.DeleteAllTags(ctx, acct, id); err != nil {
+				slog.WarnContext(ctx, "Platform cleanup: tag-index clear failed", "resourceId", id, "err", err)
+				res.addItem(kind, id, DispositionFailed, ReasonOwnerDeleteFailed, 0)
+				continue
+			}
+			cleared = append(cleared, id)
+		}
+	}
+	after, err := c.tags.ScanAccountResourceIDs(ctx, acct, keyPairIDPrefix, placementGroupIDPrefix, launchTemplateIDPrefix)
+	if err != nil {
+		slog.WarnContext(ctx, "Platform cleanup: tag index re-scan failed", "err", err)
+		kind.incomplete(ec2v1.PlatformReasonStoreListFailed)
+	}
+	for _, id := range cleared {
+		disposition, reason := itemDisposition(id, attempts{deleted: map[string]bool{id: true}}, slices.Contains(after, id), err == nil)
+		res.addItem(kind, id, disposition, reason, 0)
+	}
+	if len(after) > 0 {
+		kind.found = true
+	}
+}
+
+func keepAll(set map[string]string, ids []string, reason string) {
+	for _, id := range ids {
+		set[id] = reason
+	}
+}
+
+// referenceCheck is the outcome of reading what instance records, spot
+// requests and launch-template versions name. complete is false when any
+// reader failed or is absent.
+type referenceCheck struct {
+	complete                  bool
+	instances, spot, template *References
+}
+
+// readReferences runs every reference check. A check that cannot complete is
+// never read as "no references".
+func (c *Cleanup) readReferences(ctx context.Context, acct string) referenceCheck {
+	r := c.refs
+	if r.Instances == nil || r.Spot == nil || r.Templates == nil {
+		slog.WarnContext(ctx, "Platform cleanup: reference reader not configured")
+		return referenceCheck{}
+	}
+	instances, err := r.Instances.AccountInstanceReferences(ctx, acct)
+	if err != nil || instances == nil {
+		slog.WarnContext(ctx, "Platform cleanup: instance reference check failed", "err", err)
+		return referenceCheck{}
+	}
+	launch, err := r.Spot.AccountLaunchReferences(ctx, acct)
+	if err != nil || launch == nil {
+		slog.WarnContext(ctx, "Platform cleanup: spot reference check failed", "err", err)
+		return referenceCheck{}
+	}
+	tmpl, err := r.Templates.AccountTemplateReferences(ctx, acct)
+	if err != nil || tmpl == nil {
+		slog.WarnContext(ctx, "Platform cleanup: launch-template reference check failed", "err", err)
+		return referenceCheck{}
+	}
+	inst := newReferences()
+	inst.merge(instances)
+	spot := newReferences()
+	spot.addKeys(launch.KeyNames...)
+	spot.addGroups(launch.PlacementGroupNames...)
+	template := newReferences()
+	template.addKeys(tmpl.KeyNames...)
+	template.addGroups(tmpl.PlacementGroupNames...)
+	template.addGroupIDs(tmpl.PlacementGroupIDs...)
+	return referenceCheck{complete: true, instances: inst, spot: spot, template: template}
+}
+
+func (r referenceCheck) keyReason(name string) string {
+	switch {
+	case r.instances.KeyNames[name]:
+		return ReasonInstanceReference
+	case r.spot.KeyNames[name]:
+		return ReasonSpotRequestReference
+	case r.template.KeyNames[name]:
+		return ReasonLaunchTemplateReference
+	}
+	return ""
+}
+
+func (r referenceCheck) groupReason(id, name string) string {
+	switch {
+	case r.instances.hasGroup(id, name):
+		return ReasonInstanceReference
+	case r.spot.hasGroup(id, name):
+		return ReasonSpotRequestReference
+	case r.template.hasGroup(id, name):
+		return ReasonLaunchTemplateReference
+	}
+	return ""
+}
+
+func markAll(set map[string]bool, ids []string) {
+	for _, id := range ids {
+		set[id] = true
+	}
+}

spinifex/domains/ec2/accountcleanup/references.go

@@ -0,0 +1,93 @@
+package accountcleanup
+
+import (
+	"context"
+
+	ec2launchtemplate "github.com/mulgadc/spinifex/spinifex/domains/ec2/launchtemplate"
+	ec2spotinstance "github.com/mulgadc/spinifex/spinifex/domains/ec2/spotinstance"
+)
+
+// References are the key-pair and placement-group names, and placement-group
+// IDs, that other EC2 records still name. They select what to keep and are
+// never reported.
+type References struct {
+	KeyNames            map[string]bool
+	PlacementGroupNames map[string]bool
+	PlacementGroupIDs   map[string]bool
+}
+
+// InstanceReferenceReader reads the key-pair and placement-group names the
+// account's non-terminated instance records carry, stopped ones included. It
+// is complete or an error, never a partial answer. A later wiring supplies it.
+type InstanceReferenceReader interface {
+	AccountInstanceReferences(ctx context.Context, accountID string) (*References, error)
+}
+
+// SpotReferenceReader reads the names the account's open or active spot
+// requests would launch with, complete or an error.
+type SpotReferenceReader interface {
+	AccountLaunchReferences(ctx context.Context, accountID string) (*ec2spotinstance.LaunchReferences, error)
+}
+
+// TemplateReferenceReader reads the names and IDs every version of every
+// live launch template would launch with, complete or an error.
+type TemplateReferenceReader interface {
+	AccountTemplateReferences(ctx context.Context, accountID string) (*ec2launchtemplate.TemplateReferences, error)
+}
+
+var (
+	_ SpotReferenceReader     = (*ec2spotinstance.SpotInstanceServiceImpl)(nil)
+	_ TemplateReferenceReader = (*ec2launchtemplate.LaunchTemplateServiceImpl)(nil)
+)
+
+// ReferenceReaders are the checks run before key pairs and placement groups
+// are deleted. A nil reader counts as a failed check.
+type ReferenceReaders struct {
+	Instances InstanceReferenceReader
+	Spot      SpotReferenceReader
+	Templates TemplateReferenceReader
+}
+
+func newReferences() *References {
+	return &References{KeyNames: map[string]bool{}, PlacementGroupNames: map[string]bool{}, PlacementGroupIDs: map[string]bool{}}
+}
+
+func (r *References) addKeys(names ...string) {
+	for _, name := range names {
+		if name != "" {
+			r.KeyNames[name] = true
+		}
+	}
+}
+
+func (r *References) addGroups(names ...string) {
+	for _, name := range names {
+		if name != "" {
+			r.PlacementGroupNames[name] = true
+		}
+	}
+}
+
+func (r *References) addGroupIDs(ids ...string) {
+	for _, id := range ids {
+		if id != "" {
+			r.PlacementGroupIDs[id] = true
+		}
+	}
+}
+
+func (r *References) merge(o *References) {
+	for name := range o.KeyNames {
+		r.addKeys(name)
+	}
+	for name := range o.PlacementGroupNames {
+		r.addGroups(name)
+	}
+	for id := range o.PlacementGroupIDs {
+		r.addGroupIDs(id)
+	}
+}
+
+func (r *References) hasGroup(id, name string) bool {
+	return r.PlacementGroupIDs[id] || r.PlacementGroupNames[name]
+}

spinifex/domains/ec2/accountcleanup/types.go

@@ -0,0 +1,97 @@
+package accountcleanup
+
+import "time"
+
+// Request names the account to clean. CorrelationRef is an opaque caller
+// reference echoed for logs; it grants nothing and is not a fence.
+type Request struct {
+	AccountID      string `json:"account_id"`
+	CorrelationRef string `json:"correlation_ref,omitempty"`
+}
+
+// Result is what the owner stores hold after one run. Kinds reuse the ec2.v1
+// platform kind and status vocabulary; Items and Counts carry resource IDs,
+// fixed reasons and counts only, never names, tags, key material or data.
+type Result struct {
+	AccountID      string       `json:"account_id"`
+	CorrelationRef string       `json:"correlation_ref,omitempty"`
+	Kinds          []KindReport `json:"kinds"`
+	Items          []Item       `json:"items,omitempty"`
+	Counts         []Count      `json:"counts,omitempty"`
+}
+
+// KindReport is one kind's status after its re-scan. Clean means the re-scan
+// read every store completely and found nothing; writers are not fenced, so
+// it is a snapshot, not evidence of absence.
+type KindReport struct {
+	Kind              string    `json:"kind"`
+	Status            string    `json:"status"`
+	IncompleteReasons []string  `json:"incomplete_reasons,omitempty"`
+	StartedAt         time.Time `json:"started_at"`
+	EndedAt           time.Time `json:"ended_at"`
+}
+
+// Item is one resource by its AWS resource ID. MemberCount is set only for a
+// placement group blocked by members; instance and request IDs are never reported.
+type Item struct {
+	ResourceID  string `json:"resource_id"`
+	Kind        string `json:"kind"`
+	Disposition string `json:"disposition"`
+	Reason      string `json:"reason,omitempty"`
+	MemberCount int    `json:"member_count,omitempty"`
+}
+
+// Count is the number of records of one class with one disposition, for
+// records that have no public identity.
+type Count struct {
+	Kind        string `json:"kind"`
+	Class       string `json:"class"`
+	Disposition string `json:"disposition"`
+	Reason      string `json:"reason,omitempty"`
+	Count       int    `json:"count"`
+}
+
+// Dispositions. Only a complete owner re-scan that no longer finds a record
+// makes it deleted.
+const (
+	DispositionDeleted = "deleted"
+	DispositionBlocked = "blocked"
+	DispositionFailed  = "failed"
+	DispositionResidue = "residue"
+)
+
+// Reasons are fixed categories, never store error text.
+const (
+	// ReasonInstanceMembers: a placement group still names instances.
+	ReasonInstanceMembers = "instance-members"
+	// ReasonOwnerDeleteFailed: the owner delete hit a store failure.
+	ReasonOwnerDeleteFailed = "owner-delete-failed"
+	// ReasonStillPresent: the owner reported a delete but the re-scan found it.
+	ReasonStillPresent = "still-present"
+	// ReasonAbsenceUnproven: the re-scan failed, so absence is not shown.
+	ReasonAbsenceUnproven = "absence-unproven"
+	// ReasonUnreadableRecord: a record the owner cannot read is never deleted.
+	ReasonUnreadableRecord = "unreadable-record"
+	// ReasonUnrecognisedRecord: an entry matching no owner record shape.
+	ReasonUnrecognisedRecord = "unrecognised-record"
+	// ReasonOwnerRecordsUnresolved: the record may belong to an unreadable one.
+	ReasonOwnerRecordsUnresolved = "owner-records-unresolved"
+	// ReasonRecentlyWritten: residue younger than the settle window, which may
+	// be a create still in flight.
+	ReasonRecentlyWritten = "recently-written"
+	// ReasonAppearedDuringCleanup: a record the run did not see until its re-scan.
+	ReasonAppearedDuringCleanup = "appeared-during-cleanup"
+	// ReasonInstanceReference: a non-terminated instance record names it.
+	ReasonInstanceReference = "instance-reference"
+	// ReasonSpotRequestReference: an open or active spot request names it.
+	ReasonSpotRequestReference = "spot-request-reference"
+	// ReasonLaunchTemplateReference: a version of a live launch template names it.
+	ReasonLaunchTemplateReference = "launch-template-reference"
+	// ReasonReferenceUnknown: a record EC2 cannot read may name it, so it is
+	// never deleted by account cleanup.
+	ReasonReferenceUnknown = "reference-unknown"
+	// ReasonReferenceCheckFailed: an instance, spot or launch-template reference
+	// check did not complete, so nothing it guards was deleted. Also a kind
+	// incomplete reason.
+	ReasonReferenceCheckFailed = "reference-check-failed"
+)

spinifex/domains/ec2/awsapi/instance/DescribeInstanceTypes.go

@@ -3,7 +3,6 @@ package instance
 import (
 	"context"
 	"encoding/json"
-	"errors"
 	"fmt"
 	"github.com/mulgadc/spinifex/spinifex/foundation/messaging/nats"
 	"log/slog"
@@ -101,10 +100,10 @@ func DescribeInstanceTypes(ctx context.Context, input *ec2.DescribeInstanceTypes
 
 		if missing := missingInstanceTypes(finalInstanceTypes, input.InstanceTypes); len(missing) > 0 {
 			if !proveAbsence || !everyNodeAnswered(nodeIDs, answered, sum) {
-				return nil, errors.Join(
+				return nil, awserrors.WithRetryAfter(
 					awserrors.Errorf(awserrors.ErrorServiceUnavailable,
 						"cannot confirm instance type(s) %s: describe fan-out did not complete", strings.Join(missing, ", ")),
-					awserrors.RetryAfter(awserrors.ErrorServiceUnavailable, describeIncompleteRetryAfter),
+					describeIncompleteRetryAfter,
 				)
 			}
 			return nil, awserrors.Errorf(awserrors.ErrorInvalidInstanceType,

spinifex/domains/ec2/awsapi/instance/DescribeInstances.go

@@ -126,10 +126,10 @@ func DescribeInstancesChecked(ctx context.Context, input *ec2.DescribeInstancesI
 			// An incomplete sweep cannot prove absence, so a missing id is
 			// unresolved, not gone. NotFound would be a false negative and an
 			// empty success would look identical to a real one, so return 503.
-			return nil, errors.Join(
+			return nil, awserrors.WithRetryAfter(
 				awserrors.Errorf(awserrors.ErrorServiceUnavailable,
 					"cannot confirm status of instance(s) %s: describe fan-out did not complete", strings.Join(missing, ", ")),
-				awserrors.RetryAfter(awserrors.ErrorServiceUnavailable, describeIncompleteRetryAfter),
+				describeIncompleteRetryAfter,
 			)
 		}
 	}
@@ -198,10 +198,10 @@ func DescribeAccountInstancesComplete(ctx context.Context, natsConn *nats.Conn,
 }
 
 func incompleteListingError() error {
-	return errors.Join(
+	return awserrors.WithRetryAfter(
 		awserrors.Errorf(awserrors.ErrorServiceUnavailable,
 			"cannot list instances: not every node and instance store answered"),
-		awserrors.RetryAfter(awserrors.ErrorServiceUnavailable, describeIncompleteRetryAfter),
+		describeIncompleteRetryAfter,
 	)
 }
 

spinifex/domains/ec2/awsapi/instance/placement.go

@@ -227,7 +227,7 @@ func aggregateResults(ctx context.Context, results []nodeLaunchResult, minCount
 			return nil, clientErr
 		}
 		// A launch that was attempted and failed is not a capacity shortage; surface the real cause.
-		if failErr := launchFailure(results); failErr != nil {
+		if failErr := launchFailure(ctx, results); failErr != nil {
 			slog.ErrorContext(ctx, "distributeInstances: launch fell short of minCount", "minCount", minCount, "launched", totalLaunched, "err", failErr)
 			return nil, failErr
 		}
@@ -321,7 +321,7 @@ func distributeInstancesSpread(ctx context.Context, input *ec2.RunInstancesInput
 		if clientErr := extractClientError(results); clientErr != nil {
 			return nil, clientErr
 		}
-		if failErr := launchFailure(results); failErr != nil {
+		if failErr := launchFailure(ctx, results); failErr != nil {
 			slog.ErrorContext(ctx, "distributeInstancesSpread: launch fell short of minCount", "minCount", minCount, "launched", totalLaunched, "err", failErr)
 			return nil, failErr
 		}
@@ -414,7 +414,7 @@ func distributeInstancesCluster(ctx context.Context, input *ec2.RunInstancesInpu
 		if clientErr := extractClientError(results); clientErr != nil {
 			return nil, clientErr
 		}
-		return nil, results[0].Err
+		return nil, launchFailure(ctx, results)
 	}
 
 	reservation := results[0].Reservation
@@ -468,17 +468,26 @@ func extractClientError(results []nodeLaunchResult) error {
 	return nil
 }
 
-// launchFailure returns the first node launch error, or nil when no node errored.
-// A sub-minCount launch whose nodes actually errored is not a capacity shortage:
-// genuine capacity is caught pre-launch in distributeInstances, and a node-side
-// capacity race is surfaced by extractClientError. Surfacing the real error (RPC
-// timeout, network/OVN failure) stops it masquerading as
-// InsufficientInstanceCapacity — the mislabel that derails triage.
-func launchFailure(results []nodeLaunchResult) error {
+// launchFailure returns the first real node launch error so an RPC or OVN failure is
+// not mislabelled as capacity; a node with no launch subscriber for the type cannot
+// place it, so when only those failed the result is InsufficientInstanceCapacity.
+func launchFailure(ctx context.Context, results []nodeLaunchResult) error {
+	var noResponders error
 	for _, r := range results {
-		if r.Err != nil {
-			return r.Err
+		if r.Err == nil {
+			continue
 		}
+		if errors.Is(r.Err, nats.ErrNoResponders) {
+			if noResponders == nil {
+				noResponders = r.Err
+			}
+			continue
+		}
+		return r.Err
+	}
+	if noResponders != nil {
+		slog.WarnContext(ctx, "distributeInstances: no launch subscriber on node, reporting InsufficientInstanceCapacity", "err", noResponders)
+		return errors.New(awserrors.ErrorInsufficientInstanceCapacity)
 	}
 	return nil
 }

spinifex/domains/ec2/awsapi/instance/placement_test.go

@@ -335,6 +335,32 @@ func TestAggregateResults_TimeoutSurfacedNotMasked(t *testing.T) {
 	assert.ErrorIs(t, err, context.DeadlineExceeded)
 }
 
+// A node with no subscriber on its launch subject surfaces as capacity, never the raw NATS error.
+func TestAggregateResults_NoRespondersIsCapacity(t *testing.T) {
+	t.Parallel()
+	noResp := fmt.Errorf("launch on node-1: %w", fmt.Errorf("NATS request to ec2.RunInstances.t3.large.node-1: %w", nats.ErrNoResponders))
+	results := []nodeLaunchResult{
+		{NodeID: "node-1", Err: noResp},
+	}
+
+	_, err := aggregateResults(context.Background(), results, 1, nil, "")
+	require.EqualError(t, err, awserrors.ErrorInsufficientInstanceCapacity)
+}
+
+// A real launch failure on another node outranks a no-responders node.
+func TestAggregateResults_RealFailureOutranksNoResponders(t *testing.T) {
+	t.Parallel()
+	noResp := fmt.Errorf("launch on node-1: %w", nats.ErrNoResponders)
+	timeoutErr := fmt.Errorf("launch on node-2: %w", context.DeadlineExceeded)
+	results := []nodeLaunchResult{
+		{NodeID: "node-1", Err: noResp},
+		{NodeID: "node-2", Err: timeoutErr},
+	}
+
+	_, err := aggregateResults(context.Background(), results, 1, nil, "")
+	assert.ErrorIs(t, err, context.DeadlineExceeded)
+}
+
 // Short launch with no node errors (nodes returned fewer instances than assigned
 // without erroring) keeps the generic capacity code.
 func TestAggregateResults_ShortWithoutNodeErrors(t *testing.T) {
@@ -1120,6 +1146,40 @@ func TestDistributeInstancesCluster_InsufficientCapacityOnPinnedNode(t *testing.
 	assert.Equal(t, awserrors.ErrorInsufficientInstanceCapacity, err.Error())
 }
 
+// A pinned node with no launch subscriber reports capacity, not the raw NATS error.
+func TestDistributeInstancesCluster_NoLaunchSubscriberIsCapacity(t *testing.T) {
+	t.Parallel()
+	_, nc := startTestNATSServer(t)
+
+	statusSub, err := nc.Subscribe("spinifex.node.status", func(msg *nats.Msg) {
+		data, _ := json.Marshal(types.NodeStatusResponse{Node: "node-1", InstanceTypes: []types.InstanceTypeCap{{Name: "t3.micro", Available: 10}}})
+		_ = nc.Publish(msg.Reply, data)
+	})
+	require.NoError(t, err)
+	defer statusSub.Unsubscribe()
+
+	reserveSub, err := nc.QueueSubscribe("ec2.ReserveClusterNode", "spinifex-workers", func(msg *nats.Msg) {
+		data, _ := json.Marshal(struct {
+			TargetNode string `json:"target_node"`
+		}{TargetNode: "node-1"})
+		_ = msg.Respond(data)
+	})
+	require.NoError(t, err)
+	defer reserveSub.Unsubscribe()
+
+	require.NoError(t, nc.Flush())
+
+	input := &ec2.RunInstancesInput{
+		ImageId:      aws.String("ami-test"),
+		InstanceType: aws.String("t3.micro"),
+		MinCount:     aws.Int64(1),
+		MaxCount:     aws.Int64(1),
+	}
+
+	_, err = distributeInstancesCluster(context.Background(), input, nc, "test-account", "my-cluster-group", 1)
+	require.EqualError(t, err, awserrors.ErrorInsufficientInstanceCapacity)
+}
+
 func TestDistributeInstancesCluster_PinnedNodeNotInCapacityResults(t *testing.T) {
 	t.Parallel()
 	// Pinned node has no capacity at all (not in fan-out results) → InsufficientInstanceCapacity

spinifex/domains/ec2/awsapi/instance/status_synthesis_test.go

@@ -12,6 +12,8 @@ import (
 
 	"github.com/aws/aws-sdk-go/aws"
 	"github.com/aws/aws-sdk-go/service/ec2"
+	awsidentifiers "github.com/mulgadc/spinifex/spinifex/foundation/aws/identifiers"
+	"github.com/mulgadc/spinifex/spinifex/foundation/aws/tags"
 	"github.com/mulgadc/spinifex/spinifex/runtime/compute/cache"
 	"github.com/mulgadc/spinifex/spinifex/runtime/compute/vm"
 	"github.com/nats-io/nats.go"
@@ -237,6 +239,44 @@ func TestSynthesis_HonoursInstanceIDFilter(t *testing.T) {
 	require.Equal(t, "i-asked", aws.StringValue(out.InstanceStatuses[0].InstanceId))
 }
 
+// Synthesis shares the node path's selection, so a platform-managed VM on a
+// stale node is hidden from the customer account it is recorded under.
+func TestSynthesis_HidesManagedSystemVMFromCustomer(t *testing.T) {
+	t.Parallel()
+	_, nc := startTestNATSServer(t)
+	answerWith(t, nc, "node-active")
+
+	managed := cachedVM("i-ekscp", "node-dead", "az-a", vm.StateRunning)
+	managed.ManagedBy = tags.ManagedByEKS
+	synth := StatusSynthesis{
+		Records:  fakeRecords{ready: true, vms: []*vm.VM{managed}},
+		Liveness: fakeLiveness{states: map[string]instancecache.NodeState{"node-dead": instancecache.NodeStale}},
+	}
+
+	out, err := DescribeInstanceStatus(context.Background(), &ec2.DescribeInstanceStatusInput{}, nc, 1, synthAccount, "az-a", synth)
+	require.NoError(t, err)
+	require.Empty(t, out.InstanceStatuses)
+}
+
+func TestSynthesis_RootSeesManagedSystemVM(t *testing.T) {
+	t.Parallel()
+	_, nc := startTestNATSServer(t)
+	answerWith(t, nc, "node-active")
+
+	managed := cachedVM("i-lb", "node-dead", "az-a", vm.StateRunning)
+	managed.AccountID = awsidentifiers.GlobalAccountID
+	managed.ManagedBy = tags.ManagedByELBv2
+	synth := StatusSynthesis{
+		Records:  fakeRecords{ready: true, vms: []*vm.VM{managed}},
+		Liveness: fakeLiveness{states: map[string]instancecache.NodeState{"node-dead": instancecache.NodeStale}},
+	}
+
+	out, err := DescribeInstanceStatus(context.Background(), &ec2.DescribeInstanceStatusInput{}, nc, 1, awsidentifiers.GlobalAccountID, "az-a", synth)
+	require.NoError(t, err)
+	require.Len(t, out.InstanceStatuses, 1)
+	require.Equal(t, "i-lb", aws.StringValue(out.InstanceStatuses[0].InstanceId))
+}
+
 // A cache that has not finished its first whole-set sync cannot support a
 // claim about what is missing, so it contributes nothing.
 func TestSynthesis_ColdCacheContributesNothing(t *testing.T) {

spinifex/domains/ec2/image/account_termination_current_behaviour_test.go

@@ -0,0 +1,264 @@
+package image_test
+
+import (
+	"context"
+	"testing"
+
+	"github.com/aws/aws-sdk-go/aws"
+	"github.com/aws/aws-sdk-go/service/ec2"
+	testutil "github.com/mulgadc/spinifex/internal/testkit"
+	"github.com/mulgadc/spinifex/spinifex/bootstrap/config"
+	ebsmetadata "github.com/mulgadc/spinifex/spinifex/domains/ec2/ebs/metadata"
+	ec2image "github.com/mulgadc/spinifex/spinifex/domains/ec2/image"
+	ec2snapshot "github.com/mulgadc/spinifex/spinifex/domains/ec2/snapshot"
+	ec2volume "github.com/mulgadc/spinifex/spinifex/domains/ec2/volume"
+	awserrors "github.com/mulgadc/spinifex/spinifex/foundation/aws/errors"
+	ebsprovider "github.com/mulgadc/spinifex/spinifex/providers/ebs"
+	"github.com/mulgadc/spinifex/spinifex/providers/objectstore"
+	"github.com/nats-io/nats.go/jetstream"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+// The TestCurrentBehaviour_AccountTermination_ tests pin how volumes,
+// snapshots and images guard one another today, wired the way the daemon wires
+// them; a failure means a dependency rule changed.
+
+const (
+	ctAccountID = "123456789012"
+	ctBucket    = "ct-bucket"
+	ctAZ        = "ct-az"
+)
+
+type storageServices struct {
+	volumes   *ec2volume.VolumeServiceImpl
+	snapshots *ec2snapshot.SnapshotServiceImpl
+	images    *ec2image.ImageServiceImpl
+	provider  *ebsprovider.MemoryProvider
+	metadata  *ebsmetadata.Store
+}
+
+// newStorageServices shares one object store, one provider and one
+// volume-to-snapshot index between the three services. As in the daemon, the
+// image service is given no handle on that index.
+func newStorageServices(t *testing.T) storageServices {
+	t.Helper()
+	_, nc, js := testutil.StartTestJetStream(t)
+	refs, err := js.CreateKeyValue(context.Background(), jetstream.KeyValueConfig{Bucket: ec2snapshot.KVBucketVolumeSnapshots})
+	require.NoError(t, err)
+
+	store := objectstore.NewMemoryObjectStore()
+	provider := ebsprovider.NewMemoryProvider(ebsprovider.Capabilities{SnapshotEnumeration: true})
+	cfg := &config.Config{AZ: ctAZ, WalDir: t.TempDir(), Predastore: config.PredastoreConfig{Bucket: ctBucket}}
+
+	volumes := ec2volume.NewVolumeServiceImplWithStore(cfg, store, nc, refs)
+	volumes.SetEBSProvider(provider)
+	snapshots := ec2snapshot.NewSnapshotServiceImplWithStore(cfg, store, nc, refs)
+	snapshots.SetEBSProvider(provider)
+	images := ec2image.NewImageServiceImplWithConfig(cfg, store, nil)
+	images.SetEBSProvider(provider)
+
+	return storageServices{
+		volumes: volumes, snapshots: snapshots, images: images,
+		provider: provider, metadata: ebsmetadata.NewStore(store, ctBucket),
+	}
+}
+
+func (s storageServices) seedVolume(t *testing.T, volumeID string) {
+	t.Helper()
+	ctx := context.Background()
+	created, err := s.provider.CreateVolume(ctx, ebsprovider.CreateVolumeRequest{
+		Versioned:     ebsprovider.NewVersioned(),
+		VolumeID:      volumeID,
+		CapacityRange: ebsprovider.CapacityRange{RequiredBytes: 1 << 30},
+	})
+	require.NoError(t, err)
+	require.NoError(t, s.metadata.PutVolume(ctx, ebsmetadata.Volume{
+		VolumeID: volumeID, TenantID: ctAccountID, CapacityGiB: 1,
+		State: "available", ProviderHandle: created.Handle,
+	}))
+}
+
+func (s storageServices) createImage(t *testing.T, rootVolumeID, name string) string {
+	t.Helper()
+	out, err := s.images.CreateImageFromInstance(ec2image.CreateImageParams{
+		Input:        &ec2.CreateImageInput{InstanceId: aws.String("i-0123456789abcdef0"), Name: aws.String(name)},
+		RootVolumeID: rootVolumeID,
+	}, ctAccountID)
+	require.NoError(t, err)
+	return aws.StringValue(out.ImageId)
+}
+
+func (s storageServices) deleteVolume(volumeID string) error {
+	_, err := s.volumes.DeleteVolume(context.Background(), &ec2.DeleteVolumeInput{VolumeId: aws.String(volumeID)}, ctAccountID)
+	return err
+}
+
+// A CreateSnapshot snapshot pins its source volume through the index, but a
+// CreateImage snapshot is never added to it, so its source volume deletes and
+// the image keeps naming a volume that no longer exists.
+func TestCurrentBehaviour_AccountTermination_CreateImageSnapshotDoesNotPinSourceVolume(t *testing.T) {
+	s := newStorageServices(t)
+	ctx := context.Background()
+
+	s.seedVolume(t, "vol-snapshotted")
+	_, err := s.snapshots.CreateSnapshot(ctx, &ec2.CreateSnapshotInput{VolumeId: aws.String("vol-snapshotted")}, ctAccountID)
+	require.NoError(t, err)
+	assert.EqualError(t, s.deleteVolume("vol-snapshotted"), awserrors.ErrorVolumeInUse,
+		"a CreateSnapshot snapshot blocks deleting its source volume")
+
+	s.seedVolume(t, "vol-imaged")
+	imageID := s.createImage(t, "vol-imaged", "ct-image")
+	require.NoError(t, s.deleteVolume("vol-imaged"), "a CreateImage snapshot does not block deleting its source volume")
+
+	_, err = s.metadata.GetVolume(ctx, ctAccountID, "vol-imaged")
+	require.Error(t, err, "the source volume's document is gone")
+
+	images, err := s.images.DescribeImages(ctx, &ec2.DescribeImagesInput{ImageIds: []*string{aws.String(imageID)}}, ctAccountID)
+	require.NoError(t, err)
+	require.Len(t, images.Images, 1)
+	assert.Equal(t, ec2.ImageStateAvailable, aws.StringValue(images.Images[0].State), "the image still reports available")
+
+	source, err := s.images.GetAMISourceVolumeID(ctx, imageID)
+	require.NoError(t, err)
+	assert.Equal(t, "vol-imaged", source, "a launch from the image still resolves the deleted volume")
+}
+
+// DeleteSnapshot consults AMI documents as well as volumes cloned from the
+// snapshot, so the snapshot under a live image refuses to delete until the
+// image is deregistered.
+func TestAccountTermination_DeleteSnapshotRefusesARegisteredImage(t *testing.T) {
+	s := newStorageServices(t)
+	ctx := context.Background()
+
+	s.seedVolume(t, "vol-registered")
+	snap, err := s.snapshots.CreateSnapshot(ctx, &ec2.CreateSnapshotInput{VolumeId: aws.String("vol-registered")}, ctAccountID)
+	require.NoError(t, err)
+	snapshotID := aws.StringValue(snap.SnapshotId)
+
+	registered, err := s.images.RegisterImage(ctx, &ec2.RegisterImageInput{
+		Name:           aws.String("ct-registered"),
+		RootDeviceName: aws.String("/dev/sda1"),
+		BlockDeviceMappings: []*ec2.BlockDeviceMapping{{
+			DeviceName: aws.String("/dev/sda1"),
+			Ebs:        &ec2.EbsBlockDevice{SnapshotId: aws.String(snapshotID)},
+		}},
+	}, ctAccountID)
+	require.NoError(t, err)
+
+	_, err = s.snapshots.DeleteSnapshot(ctx, &ec2.DeleteSnapshotInput{SnapshotId: aws.String(snapshotID)}, ctAccountID)
+	assert.EqualError(t, err, awserrors.ErrorInvalidSnapshotInUse, "the snapshot refuses to delete while an image is registered on it")
+
+	_, err = s.metadata.GetSnapshot(ctx, ctAccountID, snapshotID)
+	assert.NoError(t, err, "the snapshot document is still present")
+	_, held := s.provider.Snapshot(snapshotID)
+	assert.True(t, held, "the provider still holds the snapshot")
+
+	_, err = s.images.DeregisterImage(ctx, &ec2.DeregisterImageInput{ImageId: registered.ImageId}, ctAccountID)
+	require.NoError(t, err)
+
+	_, err = s.snapshots.DeleteSnapshot(ctx, &ec2.DeleteSnapshotInput{SnapshotId: aws.String(snapshotID)}, ctAccountID)
+	require.NoError(t, err, "the snapshot deletes once the image is deregistered")
+}
+
+// deleteDuringClone runs a hook once, after the provider has accepted a clone
+// from a snapshot and before CreateVolume writes the clone's document.
+type deleteDuringClone struct {
+	ebsprovider.EBSProvider
+
+	hook  func()
+	fired bool
+}
+
+func (p *deleteDuringClone) CreateVolume(ctx context.Context, req ebsprovider.CreateVolumeRequest) (*ebsprovider.Volume, error) {
+	created, err := p.EBSProvider.CreateVolume(ctx, req)
+	if err == nil && req.SourceSnapshotID != "" && !p.fired {
+		p.fired = true
+		p.hook()
+	}
+	return created, err
+}
+
+// DeleteSnapshot's clone check reads volume documents, and CreateVolume writes
+// the clone's document only after the provider has verified and cloned the
+// snapshot, so a delete in that window passes the check and strands the clone.
+func TestCurrentBehaviour_AccountTermination_DeleteSnapshotDuringCloneCreationStrandsTheClone(t *testing.T) {
+	s := newStorageServices(t)
+	ctx := context.Background()
+	createFrom := func(snapshotID string) (*ec2.Volume, error) {
+		return s.volumes.CreateVolume(ctx, &ec2.CreateVolumeInput{
+			AvailabilityZone: aws.String(ctAZ), SnapshotId: aws.String(snapshotID),
+		}, ctAccountID)
+	}
+	deleteSnapshot := func(snapshotID string) error {
+		_, err := s.snapshots.DeleteSnapshot(ctx, &ec2.DeleteSnapshotInput{SnapshotId: aws.String(snapshotID)}, ctAccountID)
+		return err
+	}
+
+	s.seedVolume(t, "vol-settled")
+	settled, err := s.snapshots.CreateSnapshot(ctx, &ec2.CreateSnapshotInput{VolumeId: aws.String("vol-settled")}, ctAccountID)
+	require.NoError(t, err)
+	_, err = createFrom(aws.StringValue(settled.SnapshotId))
+	require.NoError(t, err)
+	assert.True(t, awserrors.IsErrorCode(deleteSnapshot(aws.StringValue(settled.SnapshotId)), awserrors.ErrorInvalidSnapshotInUse),
+		"once the clone's document exists, the delete is refused")
+
+	s.seedVolume(t, "vol-raced")
+	raced, err := s.snapshots.CreateSnapshot(ctx, &ec2.CreateSnapshotInput{VolumeId: aws.String("vol-raced")}, ctAccountID)
+	require.NoError(t, err)
+	snapshotID := aws.StringValue(raced.SnapshotId)
+
+	var raceErr error
+	s.volumes.SetEBSProvider(&deleteDuringClone{EBSProvider: s.provider, hook: func() { raceErr = deleteSnapshot(snapshotID) }})
+	clone, err := createFrom(snapshotID)
+	require.NoError(t, err, "CreateVolume reports success")
+	require.NoError(t, raceErr, "the delete inside the window passes the clone check")
+
+	doc, err := s.metadata.GetVolume(ctx, ctAccountID, aws.StringValue(clone.VolumeId))
+	require.NoError(t, err)
+	assert.Equal(t, snapshotID, doc.SnapshotID, "the clone's document names the deleted snapshot")
+	_, err = s.metadata.GetSnapshot(ctx, ctAccountID, snapshotID)
+	require.Error(t, err, "the snapshot's document is gone")
+	_, held := s.provider.Snapshot(snapshotID)
+	assert.False(t, held, "the provider no longer holds the snapshot the clone was made from")
+
+	_, err = createFrom(snapshotID)
+	assert.EqualError(t, err, awserrors.ErrorInvalidSnapshotNotFound, "a clone started after the delete is refused")
+}
+
+// CreateImage on an instance re-reads the image the instance launched from,
+// after the provider has taken the snapshot, so once that image is
+// deregistered the call fails and leaves a provider snapshot with no document.
+func TestCurrentBehaviour_AccountTermination_CreateImageOfRetainedInstanceNeedsItsLaunchImage(t *testing.T) {
+	s := newStorageServices(t)
+	ctx := context.Background()
+
+	s.seedVolume(t, "vol-launch-source")
+	launchImageID := s.createImage(t, "vol-launch-source", "ct-launch-image")
+	_, err := s.images.DeregisterImage(ctx, &ec2.DeregisterImageInput{ImageId: aws.String(launchImageID)}, ctAccountID)
+	require.NoError(t, err)
+
+	s.seedVolume(t, "vol-retained-root")
+	_, err = s.images.CreateImageFromInstance(ec2image.CreateImageParams{
+		Input:         &ec2.CreateImageInput{InstanceId: aws.String("i-0123456789abcdef0"), Name: aws.String("ct-after-deregister")},
+		RootVolumeID:  "vol-retained-root",
+		SourceImageID: launchImageID,
+	}, ctAccountID)
+	assert.EqualError(t, err, awserrors.ErrorServerInternal, "the instance's image can no longer be created")
+
+	listed, err := s.provider.ListSnapshots(ctx, ebsprovider.ListSnapshotsRequest{Versioned: ebsprovider.NewVersioned()})
+	require.NoError(t, err)
+	var orphans []string
+	for _, ref := range listed.Snapshots {
+		if ref.SourceVolumeID == "vol-retained-root" {
+			orphans = append(orphans, ref.ID)
+		}
+	}
+	require.Len(t, orphans, 1, "the provider kept the snapshot it took before the image read failed")
+
+	docs, err := s.metadata.ListSnapshotsStrict(ctx, ctAccountID)
+	require.NoError(t, err)
+	for _, doc := range docs {
+		assert.NotEqual(t, orphans[0], doc.SnapshotID, "no document names that snapshot")
+	}
+}

spinifex/domains/ec2/instance/service_impl_test.go

@@ -4507,6 +4507,30 @@ func TestDescribeInstanceStatus_AccountFilteringHidesOtherTenant(t *testing.T) {
 	assert.Empty(t, out.InstanceStatuses)
 }
 
+// A platform-managed VM recorded under a customer account stays out of that
+// customer's status listing, matching DescribeInstances.
+func TestDescribeInstanceStatus_HidesManagedSystemVMFromCustomer(t *testing.T) {
+	owner := "111122223333"
+	v := runningVM("i-ekscp", owner)
+	v.ManagedBy = tags.ManagedByEKS
+	svc := instanceStatusService(t, "az-a", map[string]*vm.VM{v.ID: v})
+
+	out, err := svc.DescribeInstanceStatus(context.Background(), &ec2.DescribeInstanceStatusInput{}, owner)
+	require.NoError(t, err)
+	assert.Empty(t, out.InstanceStatuses, "managed system VM must not appear in customer status listing")
+}
+
+func TestDescribeInstanceStatus_RootSeesManagedSystemVM(t *testing.T) {
+	v := runningVM("i-lb", awsidentifiers.GlobalAccountID)
+	v.ManagedBy = tags.ManagedByELBv2
+	svc := instanceStatusService(t, "az-a", map[string]*vm.VM{v.ID: v})
+
+	out, err := svc.DescribeInstanceStatus(context.Background(), &ec2.DescribeInstanceStatusInput{}, awsidentifiers.GlobalAccountID)
+	require.NoError(t, err)
+	require.Len(t, out.InstanceStatuses, 1)
+	assert.Equal(t, "i-lb", *out.InstanceStatuses[0].InstanceId)
+}
+
 func TestDescribeInstanceStatus_StoppedExcludedByDefault(t *testing.T) {
 	owner := "111122223333"
 	stopped := runningVM("i-stop", owner)

spinifex/domains/ec2/instance/status_selection.go

@@ -59,7 +59,7 @@ func ParseStatusSelection(input *ec2.DescribeInstanceStatusInput, accountID stri
 // status is built. Split out so a caller can test membership without paying
 // for the projection.
 func (s StatusSelection) Includes(v *vm.VM) bool {
-	if v == nil || !IsInstanceVisible(s.accountID, v.AccountID) {
+	if v == nil || !IsInstanceVisibleToCaller(s.accountID, v) {
 		return false
 	}
 	if len(s.instanceIDs) > 0 && !s.instanceIDs[v.ID] {

spinifex/domains/ec2/key/account_cleanup.go

@@ -0,0 +1,75 @@
+package key
+
+import (
+	"context"
+	"log/slog"
+	"time"
+)
+
+// AccountCleanup is what one CleanupAccount run did. It holds key-pair IDs and
+// counts only; it is the owner's account of its attempts, not proof of
+// absence, which only a later ScanAccount gives.
+type AccountCleanup struct {
+	// DeletedKeyPairIDs completed the same owner delete DeleteKeyPair runs.
+	DeletedKeyPairIDs []string
+	// FailedKeyPairIDs had a store failure part way; the metadata remains.
+	FailedKeyPairIDs []string
+	// Referenced maps a kept key-pair ID to the reason the caller's predicate
+	// gave for keeping it.
+	Referenced map[string]string
+	// Orphan public-key objects removed, that failed to delete, or left because
+	// they were written after settledBefore and may belong to an in-flight create.
+	OrphansRemoved, OrphansFailed, OrphansUnsettled int
+	// OrphansReferenced counts public-key objects kept, by predicate reason.
+	OrphansReferenced map[string]int
+}
+
+// CleanupAccount deletes every readable key pair of the account through the
+// owner delete, and removes public-key objects no metadata names once they
+// were last written before settledBefore. A key pair or public-key object
+// for whose name referenced returns a reason is kept: IMDS serves the public
+// key by name to any instance naming it. Unreadable metadata, and public keys that
+// may belong to it, are left for the owner scan to report. A listing failure
+// is returned as an error and nothing is deleted.
+func (s *KeyServiceImpl) CleanupAccount(ctx context.Context, accountID string, settledBefore time.Time, referenced func(keyName string) string) (*AccountCleanup, error) {
+	records, err := s.classifyAccount(ctx, accountID)
+	if err != nil {
+		return nil, err
+	}
+	if referenced == nil {
+		referenced = func(string) string { return "" }
+	}
+	out := &AccountCleanup{Referenced: map[string]string{}, OrphansReferenced: map[string]int{}}
+	for _, kp := range records.live {
+		if reason := referenced(kp.name); reason != "" {
+			out.Referenced[kp.id] = reason
+			continue
+		}
+		if err := s.deleteKeyPairRecords(ctx, accountID, kp.name, kp.id); err != nil {
+			out.FailedKeyPairIDs = append(out.FailedKeyPairIDs, kp.id)
+			continue
+		}
+		out.DeletedKeyPairIDs = append(out.DeletedKeyPairIDs, kp.id)
+	}
+	for _, orphan := range records.orphans {
+		if reason := referenced(orphan.name); reason != "" {
+			out.OrphansReferenced[reason]++
+			continue
+		}
+		if orphan.modified == nil || !orphan.modified.Before(settledBefore) {
+			out.OrphansUnsettled++
+			continue
+		}
+		if err := s.deleteObject(ctx, orphan.key); err != nil {
+			slog.WarnContext(ctx, "Key-pair account cleanup: orphan public key delete failed", "err", err)
+			out.OrphansFailed++
+			continue
+		}
+		out.OrphansRemoved++
+	}
+	slog.InfoContext(ctx, "Key-pair account cleanup completed", "accountID", accountID,
+		"deleted", len(out.DeletedKeyPairIDs), "failed", len(out.FailedKeyPairIDs), "referenced", len(out.Referenced),
+		"orphansRemoved", out.OrphansRemoved, "orphansFailed", out.OrphansFailed,
+		"orphansUnsettled", out.OrphansUnsettled)
+	return out, nil
+}

spinifex/domains/ec2/key/account_scan.go

@@ -6,6 +6,7 @@ import (
 	"io"
 	"log/slog"
 	"strings"
+	"time"
 
 	"github.com/aws/aws-sdk-go/aws"
 	"github.com/aws/aws-sdk-go/service/s3"
@@ -29,10 +30,49 @@ type AccountScan struct {
 	ReadFailures int
 }
 
+// liveKeyPair is a readable metadata record and the name it gives its key.
+type liveKeyPair struct {
+	id, name string
+}
+
+// orphanObject is a public-key object no metadata names, the key name it is
+// served under, and its last write time when the listing reported one.
+type orphanObject struct {
+	key, name string
+	modified  *time.Time
+}
+
+// accountRecords is the private classification ScanAccount and CleanupAccount
+// share. It holds object keys and names, so it never leaves this package.
+type accountRecords struct {
+	live         []liveKeyPair
+	unreadable   []string
+	orphans      []orphanObject
+	unresolved   int
+	readFailures int
+}
+
 // ScanAccount lists and classifies every object under keys/<accountID>/. A
 // listing failure is returned as an error; per-object read failures are
 // counted. It reads only objects named as metadata (key-*.json).
 func (s *KeyServiceImpl) ScanAccount(ctx context.Context, accountID string) (*AccountScan, error) {
+	records, err := s.classifyAccount(ctx, accountID)
+	if err != nil {
+		return nil, err
+	}
+	scan := &AccountScan{
+		UnreadableKeyPairIDs: records.unreadable,
+		OrphanPublicKeys:     len(records.orphans),
+		UnresolvedPublicKeys: records.unresolved,
+		ReadFailures:         records.readFailures,
+	}
+	for _, kp := range records.live {
+		scan.LiveKeyPairIDs = append(scan.LiveKeyPairIDs, kp.id)
+	}
+	return scan, nil
+}
+
+func (s *KeyServiceImpl) classifyAccount(ctx context.Context, accountID string) (*accountRecords, error) {
 	prefix := fmt.Sprintf("keys/%s/", accountID)
 	objects, _, err := objectstore.ListAll(ctx, s.store, &s3.ListObjectsV2Input{
 		Bucket: aws.String(s.bucketName),
@@ -42,9 +82,9 @@ func (s *KeyServiceImpl) ScanAccount(ctx context.Context, accountID string) (*Ac
 		return nil, fmt.Errorf("list key-pair objects: %w", err)
 	}
 
-	scan := &AccountScan{}
+	records := &accountRecords{}
 	liveNames := make(map[string]bool)
-	var rest []string
+	var rest []*s3.Object
 	unreadable := make(map[string]bool)
 	for _, obj := range objects {
 		if obj == nil || obj.Key == nil {
@@ -53,24 +93,24 @@ func (s *KeyServiceImpl) ScanAccount(ctx context.Context, accountID string) (*Ac
 		name := strings.TrimPrefix(*obj.Key, prefix)
 		id, isJSON := strings.CutSuffix(name, ".json")
 		if !isJSON || !strings.HasPrefix(id, "key-") {
-			rest = append(rest, name)
+			rest = append(rest, obj)
 			continue
 		}
 		metadata, found, readErr := s.readScanMetadata(ctx, *obj.Key)
 		switch {
 		case readErr != nil:
-			scan.ReadFailures++
+			records.readFailures++
 			unreadable[name] = true
 		case !found:
 			// Removed between the listing and the read.
 		case metadata == nil || aws.StringValue(metadata.KeyPairId) != id || aws.StringValue(metadata.KeyName) == "":
 			unreadable[name] = true
 		default:
-			scan.LiveKeyPairIDs = append(scan.LiveKeyPairIDs, id)
+			records.live = append(records.live, liveKeyPair{id: id, name: *metadata.KeyName})
 			liveNames[*metadata.KeyName] = true
 		}
 		if unreadable[name] {
-			rest = append(rest, name)
+			rest = append(rest, obj)
 		}
 	}
 
@@ -81,19 +121,20 @@ func (s *KeyServiceImpl) ScanAccount(ctx context.Context, accountID string) (*Ac
 			delete(unreadable, name)
 		}
 	}
-	for _, name := range rest {
+	for _, obj := range rest {
+		name := strings.TrimPrefix(*obj.Key, prefix)
 		switch {
 		case liveNames[name]:
 			// The public key of a live key pair.
 		case unreadable[name]:
-			scan.UnreadableKeyPairIDs = append(scan.UnreadableKeyPairIDs, strings.TrimSuffix(name, ".json"))
+			records.unreadable = append(records.unreadable, strings.TrimSuffix(name, ".json"))
 		case len(unreadable) > 0:
-			scan.UnresolvedPublicKeys++
+			records.unresolved++
 		default:
-			scan.OrphanPublicKeys++
+			records.orphans = append(records.orphans, orphanObject{key: *obj.Key, name: name, modified: obj.LastModified})
 		}
 	}
-	return scan, nil
+	return records, nil
 }
 
 // readScanMetadata reads and decodes one metadata object. A decode failure is

spinifex/domains/ec2/key/account_scan_test.go

@@ -33,15 +33,12 @@ func TestScanAccountClassifiesEveryObjectUnderTheAccountPrefix(t *testing.T) {
 	live := importKey(t, svc, ctAccount, "live")
 	dotted := importKey(t, svc, ctAccount, "web.json")
 
-	importKey(t, svc, ctAccount, "deleted")
-	store.failDelete = func(key string) bool { return key == prefix+"deleted" }
-	_, err := svc.DeleteKeyPair(context.Background(), &ec2.DeleteKeyPairInput{KeyName: aws.String("deleted")}, ctAccount)
-	require.NoError(t, err)
-	store.failDelete = nil
+	// A public key whose metadata is gone, as an interrupted delete leaves it.
+	putObject(t, store, prefix+"deleted", ctPubKey+"\n")
 
 	store.failPut = func(key string) bool { return strings.HasSuffix(key, ".json") }
 	store.failDelete = func(string) bool { return true }
-	_, err = svc.CreateKeyPair(context.Background(), &ec2.CreateKeyPairInput{KeyName: aws.String("half")}, ctAccount)
+	_, err := svc.CreateKeyPair(context.Background(), &ec2.CreateKeyPairInput{KeyName: aws.String("half")}, ctAccount)
 	require.Error(t, err)
 	store.failPut, store.failDelete = nil, nil
 
@@ -50,7 +47,7 @@ func TestScanAccountClassifiesEveryObjectUnderTheAccountPrefix(t *testing.T) {
 	require.NoError(t, err)
 
 	assert.ElementsMatch(t, []string{aws.StringValue(live.KeyPairId), aws.StringValue(dotted.KeyPairId)}, scan.LiveKeyPairIDs)
-	assert.Equal(t, 2, scan.OrphanPublicKeys, "the failed-delete and failed-create public keys")
+	assert.Equal(t, 2, scan.OrphanPublicKeys, "the interrupted-delete and failed-create public keys")
 	assert.Empty(t, scan.UnreadableKeyPairIDs)
 	assert.Zero(t, scan.UnresolvedPublicKeys)
 	assert.Zero(t, scan.ReadFailures)

spinifex/domains/ec2/key/account_termination_current_behaviour_test.go

@@ -7,6 +7,7 @@ import (
 	"slices"
 	"strings"
 	"testing"
+	"time"
 
 	"github.com/aws/aws-sdk-go/aws"
 	"github.com/aws/aws-sdk-go/service/ec2"
@@ -19,8 +20,8 @@ import (
 )
 
 // The TestCurrentBehaviour_AccountTermination_ tests pin how key pairs are
-// stored and what a delete or Describe leaves behind today. They describe
-// current behaviour, not the target design; a failure means it changed.
+// stored and what Describe leaves behind today; a failure means it changed.
+// The TestAccountTermination_ tests pin the corrected owner delete.
 
 const (
 	ctBucket  = "test-bucket"
@@ -152,35 +153,97 @@ func TestCurrentBehaviour_AccountTermination_KeyPairStorageLayout(t *testing.T)
 	}
 }
 
-// Pins current behaviour that differs from AWS: a failed public-key delete is
-// logged, the metadata delete proceeds and the call reports success. The
-// leftover object still blocks the name and passes launch validation.
-func TestCurrentBehaviour_AccountTermination_DeleteKeyPairLeavesPublicKeyWhenItsDeleteFails(t *testing.T) {
+// A key pair is deleted public key first and metadata last, so a failed step
+// returns an error and leaves the key pair described and found by name; a
+// retried delete then removes both objects and frees the name.
+func TestAccountTermination_DeleteKeyPairKeepsTheKeyPairWhenAStepFails(t *testing.T) {
+	prefix := "keys/" + ctAccount + "/"
+	tests := []struct {
+		name   string
+		failOn func(keyPairID string) string
+		left   func(keyPairID string) []string
+	}{
+		{"public key delete fails", func(string) string { return prefix + "deploy" },
+			func(id string) []string { return sortedStrings(prefix+"deploy", prefix+id+".json") }},
+		{"metadata delete fails", func(id string) string { return prefix + id + ".json" },
+			func(id string) []string { return []string{prefix + id + ".json"} }},
+	}
+	for _, tc := range tests {
+		t.Run(tc.name, func(t *testing.T) {
+			store := newFaultyStore()
+			svc := ec2key.NewKeyServiceImplWithStore(store, ctBucket)
+			id := aws.StringValue(importKey(t, svc, ctAccount, "deploy").KeyPairId)
+			failing := tc.failOn(id)
+			store.failDelete = func(key string) bool { return key == failing }
+
+			_, err := svc.DeleteKeyPair(context.Background(), &ec2.DeleteKeyPairInput{KeyName: aws.String("deploy")}, ctAccount)
+			require.Error(t, err)
+			assert.Equal(t, awserrors.ErrorServerInternal, err.Error())
+			assert.Equal(t, tc.left(id), accountObjects(t, store, ctAccount))
+			assert.Equal(t, []string{"deploy"}, describedKeyNames(t, svc, ctAccount))
+
+			store.failDelete = nil
+			out, err := svc.DeleteKeyPair(context.Background(), &ec2.DeleteKeyPairInput{KeyName: aws.String("deploy")}, ctAccount)
+			require.NoError(t, err)
+			assert.Equal(t, id, aws.StringValue(out.KeyPairId))
+			assert.Empty(t, accountObjects(t, store, ctAccount))
+			assert.ErrorContains(t, svc.ValidateKeyPairExists(context.Background(), ctAccount, "deploy"), awserrors.ErrorInvalidKeyPairNotFound)
+
+			_, err = svc.CreateKeyPair(context.Background(), &ec2.CreateKeyPairInput{KeyName: aws.String("deploy")}, ctAccount)
+			assert.NoError(t, err, "the name is free again")
+		})
+	}
+}
+
+// A public key left with no metadata is that name's key: a delete by name
+// removes it and reports success without a key pair ID, freeing the name.
+func TestAccountTermination_DeleteKeyPairByNameRemovesALeftoverPublicKey(t *testing.T) {
 	store := newFaultyStore()
 	svc := ec2key.NewKeyServiceImplWithStore(store, ctBucket)
-	imported := importKey(t, svc, ctAccount, "deploy")
-	pubKeyPath := "keys/" + ctAccount + "/deploy"
-	store.failDelete = func(key string) bool { return key == pubKeyPath }
+	putObject(t, store, "keys/"+ctAccount+"/orphan", ctPubKey+"\n")
 
-	out, err := svc.DeleteKeyPair(context.Background(), &ec2.DeleteKeyPairInput{KeyName: aws.String("deploy")}, ctAccount)
+	out, err := svc.DeleteKeyPair(context.Background(), &ec2.DeleteKeyPairInput{KeyName: aws.String("orphan")}, ctAccount)
 	require.NoError(t, err)
 	assert.True(t, aws.BoolValue(out.Return))
-	assert.Equal(t, aws.StringValue(imported.KeyPairId), aws.StringValue(out.KeyPairId))
+	assert.Nil(t, out.KeyPairId)
+	assert.Empty(t, accountObjects(t, store, ctAccount))
+	_, err = svc.CreateKeyPair(context.Background(), &ec2.CreateKeyPairInput{KeyName: aws.String("orphan")}, ctAccount)
+	assert.NoError(t, err)
+}
 
-	assert.Equal(t, []string{pubKeyPath}, accountObjects(t, store, ctAccount), "only the public key object remains")
-	assert.Empty(t, describedKeyNames(t, svc, ctAccount))
-	assert.NoError(t, svc.ValidateKeyPairExists(context.Background(), ctAccount, "deploy"))
+// A by-name delete that cannot read some metadata cannot prove the name is
+// absent, so it fails rather than reporting an idempotent success.
+func TestAccountTermination_DeleteKeyPairByNameFailsWhenMetadataIsUnreadable(t *testing.T) {
+	store := newFaultyStore()
+	svc := ec2key.NewKeyServiceImplWithStore(store, ctBucket)
+	id := aws.StringValue(importKey(t, svc, ctAccount, "deploy").KeyPairId)
+	store.failGet = func(key string) bool { return strings.HasSuffix(key, id+".json") }
 
-	// A retried delete finds no metadata, reports success and removes nothing.
-	store.failDelete = nil
-	_, err = svc.DeleteKeyPair(context.Background(), &ec2.DeleteKeyPairInput{KeyName: aws.String("deploy")}, ctAccount)
-	require.NoError(t, err)
-	assert.Equal(t, []string{pubKeyPath}, accountObjects(t, store, ctAccount))
-
-	_, err = svc.CreateKeyPair(context.Background(), &ec2.CreateKeyPairInput{KeyName: aws.String("deploy")}, ctAccount)
+	_, err := svc.DeleteKeyPair(context.Background(), &ec2.DeleteKeyPairInput{KeyName: aws.String("deploy")}, ctAccount)
 	require.Error(t, err)
-	code, _ := awserrors.ResolveErrorCode(err)
-	assert.Equal(t, awserrors.ErrorInvalidKeyPairDuplicate, code)
+	assert.Equal(t, awserrors.ErrorServerInternal, err.Error())
+	assert.Len(t, accountObjects(t, store, ctAccount), 2, "nothing was deleted")
+}
+
+// Undecodable metadata may be the named key pair, so a by-name delete fails
+// and leaves both the metadata and the same-name public key in place.
+func TestAccountTermination_DeleteKeyPairByNameFailsOnUndecodableMetadata(t *testing.T) {
+	store := newFaultyStore()
+	svc := ec2key.NewKeyServiceImplWithStore(store, ctBucket)
+	broken := "keys/" + ctAccount + "/key-0badbadbadbadbad0.json"
+	pub := "keys/" + ctAccount + "/deploy"
+	putObject(t, store, broken, "{not json")
+	putObject(t, store, pub, ctPubKey+"\n")
+
+	_, err := svc.DeleteKeyPair(context.Background(), &ec2.DeleteKeyPairInput{KeyName: aws.String("deploy")}, ctAccount)
+	require.Error(t, err)
+	assert.Equal(t, awserrors.ErrorServerInternal, err.Error())
+	assert.Equal(t, sortedStrings(broken, pub), accountObjects(t, store, ctAccount))
+}
+
+func sortedStrings(s ...string) []string {
+	slices.Sort(s)
+	return s
 }
 
 // A create whose metadata write fails and whose cleanup delete also fails
@@ -231,3 +294,79 @@ func TestCurrentBehaviour_AccountTermination_DescribeKeyPairsHidesUnreadableReco
 		})
 	}
 }
+
+// Account cleanup deletes each readable key pair through the owner delete and
+// removes a settled public key no metadata names. A younger orphan may be a
+// create in flight and is left; unreadable metadata is never deleted.
+func TestAccountCleanup_DeletesKeyPairsAndSettledOrphans(t *testing.T) {
+	store := newFaultyStore()
+	svc := ec2key.NewKeyServiceImplWithStore(store, ctBucket)
+	importKey(t, svc, ctAccount, "deploy")
+	importKey(t, svc, "999988887777", "theirs")
+	putObject(t, store, "keys/"+ctAccount+"/old-orphan", ctPubKey)
+	broken := "keys/" + ctAccount + "/key-0badbadbadbadbad0.json"
+	putObject(t, store, broken, "{not json")
+
+	out, err := svc.CleanupAccount(context.Background(), ctAccount, time.Now().Add(-time.Hour), nil)
+	require.NoError(t, err)
+	assert.Len(t, out.DeletedKeyPairIDs, 1)
+	assert.Zero(t, out.OrphansRemoved, "orphans are left while a public key may belong to unreadable metadata")
+
+	store.Clear()
+	importKey(t, svc, ctAccount, "deploy")
+	putObject(t, store, "keys/"+ctAccount+"/old-orphan", ctPubKey)
+	out, err = svc.CleanupAccount(context.Background(), ctAccount, time.Now().Add(-time.Hour), nil)
+	require.NoError(t, err)
+	assert.Equal(t, 1, out.OrphansUnsettled, "an orphan written inside the settle window is kept")
+	assert.Equal(t, []string{"keys/" + ctAccount + "/old-orphan"}, accountObjects(t, store, ctAccount))
+
+	out, err = svc.CleanupAccount(context.Background(), ctAccount, time.Now().Add(time.Second), nil)
+	require.NoError(t, err)
+	assert.Equal(t, 1, out.OrphansRemoved)
+	assert.Empty(t, accountObjects(t, store, ctAccount))
+}
+
+// A key pair, and an orphan public key, whose name the reference check names
+// are kept with the given reason: IMDS serves the public key by name.
+func TestAccountCleanup_KeepsReferencedKeyPairsAndPublicKeys(t *testing.T) {
+	store := newFaultyStore()
+	svc := ec2key.NewKeyServiceImplWithStore(store, ctBucket)
+	kp := importKey(t, svc, ctAccount, "in-use")
+	importKey(t, svc, ctAccount, "free")
+	putObject(t, store, "keys/"+ctAccount+"/served-orphan", ctPubKey)
+	referenced := func(name string) string {
+		if name == "in-use" || name == "served-orphan" {
+			return "instance-reference"
+		}
+		return ""
+	}
+
+	out, err := svc.CleanupAccount(context.Background(), ctAccount, time.Now().Add(time.Second), referenced)
+	require.NoError(t, err)
+	assert.Equal(t, map[string]string{aws.StringValue(kp.KeyPairId): "instance-reference"}, out.Referenced)
+	assert.Len(t, out.DeletedKeyPairIDs, 1)
+	assert.Equal(t, map[string]int{"instance-reference": 1}, out.OrphansReferenced)
+	assert.Zero(t, out.OrphansRemoved)
+	assert.Equal(t, []string{"in-use"}, describedKeyNames(t, svc, ctAccount))
+	assert.Contains(t, accountObjects(t, store, ctAccount), "keys/"+ctAccount+"/served-orphan")
+}
+
+// A key pair whose delete fails part way is reported failed with its metadata
+// kept; the next run completes it.
+func TestAccountCleanup_InterruptedKeyPairDeleteCompletesOnRerun(t *testing.T) {
+	store := newFaultyStore()
+	svc := ec2key.NewKeyServiceImplWithStore(store, ctBucket)
+	kp := importKey(t, svc, ctAccount, "deploy")
+	store.failDelete = func(key string) bool { return strings.HasSuffix(key, ".json") }
+
+	out, err := svc.CleanupAccount(context.Background(), ctAccount, time.Now(), nil)
+	require.NoError(t, err)
+	assert.Equal(t, []string{aws.StringValue(kp.KeyPairId)}, out.FailedKeyPairIDs)
+	assert.Equal(t, []string{"deploy"}, describedKeyNames(t, svc, ctAccount))
+
+	store.failDelete = nil
+	out, err = svc.CleanupAccount(context.Background(), ctAccount, time.Now(), nil)
+	require.NoError(t, err)
+	assert.Equal(t, []string{aws.StringValue(kp.KeyPairId)}, out.DeletedKeyPairIDs)
+	assert.Empty(t, accountObjects(t, store, ctAccount))
+}

spinifex/domains/ec2/key/service_impl.go

@@ -500,8 +500,12 @@ func (s *KeyServiceImpl) findKeyPairIdFromKeyName(ctx context.Context, accountID
 			Key:    obj.Key,
 		})
 		if err != nil {
-			slog.Debug("Failed to get metadata file", "key", *obj.Key, "err", err)
-			continue
+			if objectstore.IsNoSuchKeyError(err) {
+				continue
+			}
+			// Unread metadata may be the key pair asked for, so absence is unproven.
+			slog.ErrorContext(ctx, "Failed to get metadata file", "key", *obj.Key, "err", err)
+			return "", fmt.Errorf("read metadata %s: %w", *obj.Key, err)
 		}
 
 		body, err := io.ReadAll(getResult.Body)
@@ -509,14 +513,15 @@ func (s *KeyServiceImpl) findKeyPairIdFromKeyName(ctx context.Context, accountID
 			slog.Debug("Failed to close metadata body", "key", *obj.Key, "err", err)
 		}
 		if err != nil {
-			slog.Debug("Failed to read metadata body", "key", *obj.Key, "err", err)
-			continue
+			slog.ErrorContext(ctx, "Failed to read metadata body", "key", *obj.Key, "err", err)
+			return "", fmt.Errorf("read metadata %s: %w", *obj.Key, err)
 		}
 
 		metadata, err := decodeKeyPairMetadata(body)
 		if err != nil {
-			slog.Debug("Failed to unmarshal metadata", "key", *obj.Key, "err", err)
-			continue
+			// Undecodable metadata may be this name's key pair, so absence is unproven.
+			slog.ErrorContext(ctx, "Undecodable key-pair metadata", "key", *obj.Key, "err", err)
+			return "", fmt.Errorf("decode metadata %s: %w", *obj.Key, err)
 		}
 
 		// Check if this metadata matches the keyName
@@ -611,7 +616,7 @@ func (s *KeyServiceImpl) describedPublicKey(accountID, keyName string) (string,
 	return fmt.Sprintf("%s %s\n", bytes.TrimSpace(ssh.MarshalAuthorizedKey(publicKey)), keyName), nil
 }
 
-// DeleteKeyPair removes a key pair (both public key and metadata from S3).
+// DeleteKeyPair removes a key pair: its public key, metadata and tag-index entry.
 func (s *KeyServiceImpl) DeleteKeyPair(ctx context.Context, input *ec2.DeleteKeyPairInput, accountID string) (*ec2.DeleteKeyPairOutput, error) {
 	if input == nil {
 		return nil, errors.New(awserrors.ErrorMissingParameter)
@@ -641,7 +646,7 @@ func (s *KeyServiceImpl) DeleteKeyPair(ctx context.Context, input *ec2.DeleteKey
 				return &ec2.DeleteKeyPairOutput{Return: aws.Bool(true)}, nil
 			}
 			slog.ErrorContext(ctx, "Failed to get keyName from keyPairId", "keyPairId", keyPairID, "err", err)
-			return nil, err
+			return nil, errors.New(awserrors.ErrorServerInternal)
 		}
 	} else if input.KeyName != nil && *input.KeyName != "" {
 		// KeyName provided - validate and find the keyPairId
@@ -655,50 +660,70 @@ func (s *KeyServiceImpl) DeleteKeyPair(ctx context.Context, input *ec2.DeleteKey
 
 		keyPairID, err = s.findKeyPairIdFromKeyName(ctx, accountID, keyName)
 		if err != nil {
-			// AWS DeleteKeyPair is idempotent — return success for non-existent keys
+			// AWS DeleteKeyPair is idempotent. A public key left with no metadata
+			// is this name's key, so it goes too and the name becomes reusable.
 			if err.Error() == awserrors.ErrorInvalidKeyPairNotFound {
+				if rmErr := s.deleteObject(ctx, publicKeyPath(accountID, keyName)); rmErr != nil {
+					slog.ErrorContext(ctx, "DeleteKeyPair: failed to remove leftover public key", "keyName", keyName, "err", rmErr)
+					return nil, errors.New(awserrors.ErrorServerInternal)
+				}
 				slog.DebugContext(ctx, "DeleteKeyPair: key pair not found, returning success (idempotent)", "keyName", keyName)
 				return &ec2.DeleteKeyPairOutput{Return: aws.Bool(true)}, nil
 			}
 			slog.ErrorContext(ctx, "Failed to find keyPairId from keyName", "keyName", keyName, "err", err)
-			return nil, err
+			return nil, errors.New(awserrors.ErrorServerInternal)
 		}
 	} else {
 		return nil, errors.New(awserrors.ErrorMissingParameter)
 	}
 
 	slog.InfoContext(ctx, "Deleting key pair", "keyName", keyName, "keyPairId", keyPairID)
-
-	// Delete public key
-	publicKeyPath := fmt.Sprintf("keys/%s/%s", accountID, keyName)
-	_, err = s.store.DeleteObject(ctx, &s3.DeleteObjectInput{
-		Bucket: aws.String(s.bucketName),
-		Key:    aws.String(publicKeyPath),
-	})
-	if err != nil {
-		slog.ErrorContext(ctx, "Failed to delete public key", "path", publicKeyPath, "err", err)
-		// Continue to try deleting metadata even if public key deletion fails
-	}
-
-	// Delete metadata file (stored with keyPairID)
-	metadataPath := fmt.Sprintf("keys/%s/%s.json", accountID, keyPairID)
-	_, err = s.store.DeleteObject(ctx, &s3.DeleteObjectInput{
-		Bucket: aws.String(s.bucketName),
-		Key:    aws.String(metadataPath),
-	})
-	if err != nil {
-		slog.ErrorContext(ctx, "Failed to delete metadata", "path", metadataPath, "err", err)
+	if err := s.deleteKeyPairRecords(ctx, accountID, keyName, keyPairID); err != nil {
 		return nil, errors.New(awserrors.ErrorServerInternal)
 	}
 
-	s.clearRecordTags(ctx, accountID, keyPairID)
-
 	slog.InfoContext(ctx, "Key pair deleted successfully", "keyName", keyName, "keyPairId", keyPairID)
 
 	// AWS names the key only when one was actually deleted.
 	return &ec2.DeleteKeyPairOutput{Return: aws.Bool(true), KeyPairId: aws.String(keyPairID)}, nil
 }
 
+// deleteKeyPairRecords removes the public key, then the metadata, then the
+// tag-index entry. Metadata goes last: while it exists the key pair is still
+// described and found by name or ID, so a retried delete finishes the job.
+func (s *KeyServiceImpl) deleteKeyPairRecords(ctx context.Context, accountID, keyName, keyPairID string) error {
+	if err := s.deleteObject(ctx, publicKeyPath(accountID, keyName)); err != nil {
+		slog.ErrorContext(ctx, "Failed to delete public key", "keyName", keyName, "keyPairId", keyPairID, "err", err)
+		return fmt.Errorf("delete public key: %w", err)
+	}
+	if err := s.deleteObject(ctx, metadataPath(accountID, keyPairID)); err != nil {
+		slog.ErrorContext(ctx, "Failed to delete key pair metadata", "keyPairId", keyPairID, "err", err)
+		return fmt.Errorf("delete metadata: %w", err)
+	}
+	s.clearRecordTags(ctx, accountID, keyPairID)
+	return nil
+}
+
+// deleteObject deletes one object; an absent object is already deleted.
+func (s *KeyServiceImpl) deleteObject(ctx context.Context, key string) error {
+	_, err := s.store.DeleteObject(ctx, &s3.DeleteObjectInput{
+		Bucket: aws.String(s.bucketName),
+		Key:    aws.String(key),
+	})
+	if err != nil && !objectstore.IsNoSuchKeyError(err) {
+		return err
+	}
+	return nil
+}
+
+func publicKeyPath(accountID, keyName string) string {
+	return fmt.Sprintf("keys/%s/%s", accountID, keyName)
+}
+
+func metadataPath(accountID, keyPairID string) string {
+	return fmt.Sprintf("keys/%s/%s.json", accountID, keyPairID)
+}
+
 // DescribeKeyPairs lists available key pairs by reading metadata files from S3
 // describeKeyPairsValidFilters defines the set of filter names accepted by DescribeKeyPairs.
 var describeKeyPairsValidFilters = map[string]bool{
@@ -769,7 +794,7 @@ func (s *KeyServiceImpl) DescribeKeyPairs(ctx context.Context, input *ec2.Descri
 			Key:    obj.Key,
 		})
 		if err != nil {
-			slog.DebugContext(ctx, "Failed to get metadata file", "key", *obj.Key, "err", err)
+			slog.WarnContext(ctx, "DescribeKeyPairs: skipping metadata the store cannot read", "key", *obj.Key, "err", err)
 			continue
 		}
 
@@ -778,13 +803,14 @@ func (s *KeyServiceImpl) DescribeKeyPairs(ctx context.Context, input *ec2.Descri
 			slog.DebugContext(ctx, "Failed to close metadata body", "key", *obj.Key, "err", err)
 		}
 		if err != nil {
-			slog.DebugContext(ctx, "Failed to read metadata body", "key", *obj.Key, "err", err)
+			slog.WarnContext(ctx, "DescribeKeyPairs: skipping metadata whose body cannot be read", "key", *obj.Key, "err", err)
 			continue
 		}
 
+		// The decoder's message is not logged: it can quote the stored bytes.
 		metadata, err := decodeKeyPairMetadata(body)
 		if err != nil {
-			slog.DebugContext(ctx, "Failed to unmarshal metadata", "key", *obj.Key, "err", err)
+			slog.WarnContext(ctx, "DescribeKeyPairs: skipping undecodable metadata", "key", *obj.Key)
 			continue
 		}
 

spinifex/domains/ec2/launchtemplate/account_cleanup.go

@@ -0,0 +1,132 @@
+package launchtemplate
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"log/slog"
+	"strings"
+	"time"
+
+	"github.com/mulgadc/spinifex/spinifex/foundation/state/kvutil"
+	"github.com/nats-io/nats.go/jetstream"
+)
+
+// AccountCleanup is what one CleanupAccountResidue run did. It holds template
+// IDs and counts only; it is the owner's account of its attempts, not proof of
+// absence, which only a later ScanAccount gives.
+type AccountCleanup struct {
+	// KeptTemplateIDs are live templates; account cleanup never deletes them.
+	KeptTemplateIDs []string
+	// Version bodies and name-index entries with no header: removed, failed,
+	// or left because they were written after settledBefore and may belong to
+	// a create still in flight.
+	OrphanBodiesRemoved, OrphanBodiesFailed, OrphanBodiesUnsettled int
+	OrphanNamesRemoved, OrphanNamesFailed, OrphanNamesUnsettled    int
+}
+
+// CleanupAccountResidue removes the account's version bodies and name-index
+// entries whose header is absent, once written before settledBefore: without
+// a header no ID, name or launch can resolve them. Live templates are listed,
+// not deleted, because not every record that may name one is visible to EC2.
+// Unreadable headers, their bodies and unrecognised entries are left for the
+// owner scan to report. A listing failure is returned and nothing is removed.
+func (s *LaunchTemplateServiceImpl) CleanupAccountResidue(ctx context.Context, accountID string, settledBefore time.Time) (*AccountCleanup, error) {
+	keys, err := kvutil.Keys(ctx, s.kv)
+	if err != nil && !errors.Is(err, jetstream.ErrNoKeysFound) {
+		return nil, fmt.Errorf("list launch-template keys: %w", err)
+	}
+
+	prefix := accountID + "."
+	present := make(map[string]bool)
+	out := &AccountCleanup{}
+	var bodies, names []string
+	for _, key := range keys {
+		rest, ok := strings.CutPrefix(key, prefix)
+		if !ok {
+			continue
+		}
+		switch {
+		case isHeaderKey(key, prefix):
+			entry, err := s.kv.Get(ctx, key)
+			if errors.Is(err, jetstream.ErrKeyNotFound) {
+				continue
+			}
+			// An unreadable header still owns its bodies and name.
+			present[rest] = true
+			if err != nil {
+				continue
+			}
+			var header LaunchTemplateHeader
+			if json.Unmarshal(entry.Value(), &header) == nil && header.LaunchTemplateId == rest {
+				out.KeptTemplateIDs = append(out.KeptTemplateIDs, rest)
+			}
+		case strings.HasPrefix(rest, "name."):
+			names = append(names, key)
+		case versionBodyTemplateID(rest) != "":
+			bodies = append(bodies, key)
+		}
+	}
+
+	for _, key := range bodies {
+		if present[versionBodyTemplateID(strings.TrimPrefix(key, prefix))] {
+			continue
+		}
+		tallyOrphan(s.removeSettled(ctx, key, settledBefore, nil), &out.OrphanBodiesRemoved, &out.OrphanBodiesFailed, &out.OrphanBodiesUnsettled)
+	}
+	for _, key := range names {
+		// A name still naming a present header is that template's, not residue.
+		owned := func(value []byte) bool { return present[string(value)] }
+		tallyOrphan(s.removeSettled(ctx, key, settledBefore, owned), &out.OrphanNamesRemoved, &out.OrphanNamesFailed, &out.OrphanNamesUnsettled)
+	}
+	slog.InfoContext(ctx, "Launch-template account residue cleanup completed", "accountID", accountID,
+		"kept", len(out.KeptTemplateIDs), "orphanBodiesRemoved", out.OrphanBodiesRemoved, "orphanNamesRemoved", out.OrphanNamesRemoved)
+	return out, nil
+}
+
+// orphanOutcome is what removeSettled did with one headerless record.
+type orphanOutcome int
+
+const (
+	orphanGone orphanOutcome = iota
+	orphanRemoved
+	orphanFailed
+	orphanUnsettled
+	orphanOwned
+)
+
+// removeSettled deletes key at the revision read when it was written before
+// settledBefore and owned does not claim its value. The value is never decoded
+// or logged: a version body holds template data and user data.
+func (s *LaunchTemplateServiceImpl) removeSettled(ctx context.Context, key string, settledBefore time.Time, owned func([]byte) bool) orphanOutcome {
+	entry, err := s.kv.Get(ctx, key)
+	switch {
+	case errors.Is(err, jetstream.ErrKeyNotFound):
+		return orphanGone
+	case err != nil:
+		slog.WarnContext(ctx, "Launch-template account cleanup: orphan read failed", "err", err)
+		return orphanFailed
+	case owned != nil && owned(entry.Value()):
+		return orphanOwned
+	case !entry.Created().Before(settledBefore):
+		return orphanUnsettled
+	}
+	if err := s.kv.Delete(ctx, key, jetstream.LastRevision(entry.Revision())); err != nil {
+		slog.WarnContext(ctx, "Launch-template account cleanup: orphan delete failed", "err", err)
+		return orphanFailed
+	}
+	return orphanRemoved
+}
+
+func tallyOrphan(outcome orphanOutcome, removed, failed, unsettled *int) {
+	switch outcome {
+	case orphanRemoved:
+		*removed++
+	case orphanFailed:
+		*failed++
+	case orphanUnsettled:
+		*unsettled++
+	case orphanGone, orphanOwned:
+	}
+}

spinifex/domains/ec2/launchtemplate/account_scan_test.go

@@ -6,7 +6,6 @@ package launchtemplate
 import (
 	"context"
 	"encoding/base64"
-	"strings"
 	"testing"
 
 	"github.com/aws/aws-sdk-go/aws"
@@ -43,17 +42,17 @@ func TestScanAccountClassifiesLaunchTemplateRecords(t *testing.T) {
 	svc := setupTestService(t)
 	live := createWithUserData(t, svc, "live")
 
+	// A header removed alone, as a delete interrupted after it leaves the rest.
 	deleted := createWithUserData(t, svc, "deleted")
 	inner := svc.kv
-	deletedHeader, deletedName := kvutil.AccountKey(testAccountID, deleted), nameKey(testAccountID, "deleted")
-	svc.kv = &faultyKV{KeyValue: inner, failDelete: func(key string) bool {
-		return strings.HasPrefix(key, deletedHeader+".v") || key == deletedName
-	}}
-	_, err := svc.DeleteLaunchTemplate(context.Background(), &ec2.DeleteLaunchTemplateInput{LaunchTemplateId: aws.String(deleted)}, testAccountID)
-	require.NoError(t, err)
+	require.NoError(t, inner.Delete(t.Context(), kvutil.AccountKey(testAccountID, deleted)))
 
-	svc.kv = &faultyKV{KeyValue: inner, failPut: func(key string) bool { return isHeaderKey(key, testAccountID+".") }}
-	_, err = svc.CreateLaunchTemplate(context.Background(), &ec2.CreateLaunchTemplateInput{
+	// A create whose header write and clean-up both fail.
+	svc.kv = &faultyKV{KeyValue: inner,
+		failPut:    func(key string) bool { return isHeaderKey(key, testAccountID+".") },
+		failDelete: func(string) bool { return true },
+	}
+	_, err := svc.CreateLaunchTemplate(context.Background(), &ec2.CreateLaunchTemplateInput{
 		LaunchTemplateName: aws.String("interrupted"),
 		LaunchTemplateData: &ec2.RequestLaunchTemplateData{UserData: aws.String(base64.StdEncoding.EncodeToString([]byte(ctUserData)))},
 	}, testAccountID)

spinifex/domains/ec2/launchtemplate/account_termination_current_behaviour_test.go

@@ -9,7 +9,9 @@ import (
 	"encoding/json"
 	"errors"
 	"slices"
+	"strings"
 	"testing"
+	"time"
 
 	"github.com/aws/aws-sdk-go/aws"
 	"github.com/aws/aws-sdk-go/service/ec2"
@@ -21,8 +23,8 @@ import (
 )
 
 // The TestCurrentBehaviour_AccountTermination_ tests pin how launch templates
-// are stored and what Describe and Delete leave behind today. They describe
-// current behaviour, not the target design; a failure means it changed.
+// are stored and what Describe leaves behind today; a failure means it
+// changed. The TestAccountTermination_ tests pin the corrected owner delete.
 
 var errInjected = errors.New("injected store failure")
 
@@ -31,10 +33,31 @@ var errInjected = errors.New("injected store failure")
 type faultyKV struct {
 	jetstream.KeyValue
 
-	failGet    func(key string) bool
-	failPut    func(key string) bool
-	failDelete func(key string) bool
-	failKeys   func() bool
+	failGet      func(key string) bool
+	failPut      func(key string) bool
+	failDelete   func(key string) bool
+	failKeys     func() bool
+	failUpdate   func(key string) bool
+	beforeDelete func(key string)
+	beforeCreate func(key string)
+	beforeUpdate func(key string)
+}
+
+func (k *faultyKV) Create(ctx context.Context, key string, value []byte, opts ...jetstream.KVCreateOpt) (uint64, error) {
+	if k.beforeCreate != nil {
+		k.beforeCreate(key)
+	}
+	return k.KeyValue.Create(ctx, key, value, opts...)
+}
+
+func (k *faultyKV) Update(ctx context.Context, key string, value []byte, revision uint64) (uint64, error) {
+	if k.beforeUpdate != nil {
+		k.beforeUpdate(key)
+	}
+	if k.failUpdate != nil && k.failUpdate(key) {
+		return 0, errInjected
+	}
+	return k.KeyValue.Update(ctx, key, value, revision)
 }
 
 func (k *faultyKV) Get(ctx context.Context, key string) (jetstream.KeyValueEntry, error) {
@@ -52,6 +75,9 @@ func (k *faultyKV) Put(ctx context.Context, key string, value []byte) (uint64, e
 }
 
 func (k *faultyKV) Delete(ctx context.Context, key string, opts ...jetstream.KVDeleteOpt) error {
+	if k.beforeDelete != nil {
+		k.beforeDelete(key)
+	}
 	if k.failDelete != nil && k.failDelete(key) {
 		return errInjected
 	}
@@ -138,40 +164,44 @@ func TestCurrentBehaviour_AccountTermination_LaunchTemplateStorageLayout(t *test
 	assert.Equal(t, ctUserData, storedUserData(t, svc.kv, header+".v2"), "a version copied from v1 keeps its user data")
 }
 
-// Pins current behaviour that differs from AWS: the header delete alone
-// decides success, and a failed name-index or version-body cleanup leaves
-// those records, user data included, with no header to find them by.
-func TestCurrentBehaviour_AccountTermination_DeleteLaunchTemplateLeavesRecordsWhenCleanupFails(t *testing.T) {
+// The header is deleted last, so a failed body, listing, name-index or header
+// delete returns an error with the template still described, and a retried
+// delete by ID removes every record, user data included.
+func TestAccountTermination_DeleteLaunchTemplateKeepsTheTemplateWhenAStepFails(t *testing.T) {
 	tests := []struct {
-		name     string
-		inject   func(kv *faultyKV, header, nameIdx string)
-		residual func(header, nameIdx string) []string
+		name   string
+		inject func(kv *faultyKV, header, nameIdx string)
+		left   func(header, nameIdx string) []string
 	}{
 		{
 			name: "version body deletes fail",
 			inject: func(kv *faultyKV, header, _ string) {
 				kv.failDelete = func(key string) bool { return key == header+".v1" || key == header+".v2" }
 			},
-			residual: func(header, _ string) []string { return []string{header + ".v1", header + ".v2"} },
+			left: func(header, nameIdx string) []string {
+				return []string{header, header + ".v1", header + ".v2", nameIdx}
+			},
+		},
+		{
+			name:   "version listing fails",
+			inject: func(kv *faultyKV, _, _ string) { kv.failKeys = func() bool { return true } },
+			left: func(header, nameIdx string) []string {
+				return []string{header, header + ".v1", header + ".v2", nameIdx}
+			},
 		},
 		{
 			name: "name index delete fails",
 			inject: func(kv *faultyKV, _, nameIdx string) {
 				kv.failDelete = func(key string) bool { return key == nameIdx }
 			},
-			residual: func(_, nameIdx string) []string { return []string{nameIdx} },
+			left: func(header, nameIdx string) []string { return []string{header, nameIdx} },
 		},
 		{
-			name: "version listing fails after the header delete",
+			name: "header delete fails",
 			inject: func(kv *faultyKV, header, _ string) {
-				headerGone := false
-				kv.failDelete = func(key string) bool {
-					headerGone = headerGone || key == header
-					return false
-				}
-				kv.failKeys = func() bool { return headerGone }
+				kv.failDelete = func(key string) bool { return key == header }
 			},
-			residual: func(header, _ string) []string { return []string{header + ".v1", header + ".v2"} },
+			left: func(header, _ string) []string { return []string{header} },
 		},
 	}
 	for _, tc := range tests {
@@ -184,56 +214,85 @@ func TestCurrentBehaviour_AccountTermination_DeleteLaunchTemplateLeavesRecordsWh
 			tc.inject(faulty, header, nameIdx)
 			svc.kv = faulty
 
+			_, err := svc.DeleteLaunchTemplate(context.Background(), &ec2.DeleteLaunchTemplateInput{LaunchTemplateId: aws.String(id)}, testAccountID)
+			require.Error(t, err)
+			assert.Equal(t, awserrors.ErrorServerInternal, err.Error())
+
+			svc.kv = inner
+			assert.Equal(t, sortedKeys(append(tc.left(header, nameIdx), kvutil.VersionKey)...), storedKeys(t, inner))
+			assert.Equal(t, []string{id}, describedTemplateIDs(t, svc, testAccountID))
+
 			out, err := svc.DeleteLaunchTemplate(context.Background(), &ec2.DeleteLaunchTemplateInput{LaunchTemplateId: aws.String(id)}, testAccountID)
 			require.NoError(t, err)
 			assert.Equal(t, id, aws.StringValue(out.LaunchTemplate.LaunchTemplateId))
-
-			svc.kv = inner
-			want := append(tc.residual(header, nameIdx), kvutil.VersionKey)
-			assert.Equal(t, sortedKeys(want...), storedKeys(t, inner))
+			assert.Equal(t, []string{kvutil.VersionKey}, storedKeys(t, inner))
 			assert.Empty(t, describedTemplateIDs(t, svc, testAccountID))
-			if slices.Contains(want, header+".v1") {
-				assert.Equal(t, ctUserData, storedUserData(t, inner, header+".v1"))
-			}
 		})
 	}
 }
 
-// A create interrupted before its header write leaves the name claim and
-// version 1 behind. A retry under the same name reclaims the name, and the
-// first attempt's body is never reachable again.
-func TestCurrentBehaviour_AccountTermination_InterruptedCreateLeavesAnOrphanBody(t *testing.T) {
+// A header rewritten and a version added while the delete runs send it back
+// through the bodies at the new revision, and the late body is removed too.
+func TestAccountTermination_DeleteLaunchTemplateRestartsWhenTheHeaderChanges(t *testing.T) {
 	svc := setupTestService(t)
+	id := createWithUserData(t, svc, "web")
 	inner := svc.kv
-	svc.kv = &faultyKV{KeyValue: inner, failPut: func(key string) bool { return isHeaderKey(key, testAccountID+".") }}
-
-	_, err := svc.CreateLaunchTemplate(context.Background(), &ec2.CreateLaunchTemplateInput{
-		LaunchTemplateName: aws.String("web"),
-		LaunchTemplateData: &ec2.RequestLaunchTemplateData{
-			UserData: aws.String(base64.StdEncoding.EncodeToString([]byte(ctUserData))),
-		},
-	}, testAccountID)
-	require.Error(t, err)
-	assert.Equal(t, awserrors.ErrorServerInternal, err.Error())
-
-	svc.kv = inner
-	keys := storedKeys(t, inner)
-	require.Len(t, keys, 3, "name index, version 1 and the bucket version key")
-	nameEntry, err := inner.Get(t.Context(), nameKey(testAccountID, "web"))
+	header := testAccountID + "." + id
+	headerEntry, err := inner.Get(t.Context(), header)
 	require.NoError(t, err)
-	orphanBody := testAccountID + "." + string(nameEntry.Value()) + ".v1"
-	assert.Contains(t, keys, orphanBody)
-	assert.Equal(t, ctUserData, storedUserData(t, inner, orphanBody))
-	assert.Empty(t, describedTemplateIDs(t, svc, testAccountID))
 
-	retried := createTemplate(t, svc, "web", "t3.micro")
-	assert.NotEqual(t, string(nameEntry.Value()), aws.StringValue(retried.LaunchTemplateId))
-	assert.Contains(t, storedKeys(t, inner), orphanBody)
-	assert.Equal(t, []string{aws.StringValue(retried.LaunchTemplateId)}, describedTemplateIDs(t, svc, testAccountID))
+	raced := false
+	svc.kv = &faultyKV{KeyValue: inner, beforeDelete: func(key string) {
+		if key != header || raced {
+			return
+		}
+		raced = true
+		_, err := inner.Put(t.Context(), header, headerEntry.Value())
+		require.NoError(t, err)
+		_, err = inner.Put(t.Context(), header+".v3", []byte(`{"version_number":3}`))
+		require.NoError(t, err)
+	}}
+
+	_, err = svc.DeleteLaunchTemplate(context.Background(), &ec2.DeleteLaunchTemplateInput{LaunchTemplateId: aws.String(id)}, testAccountID)
+	require.NoError(t, err)
+	assert.True(t, raced)
+	assert.Equal(t, []string{kvutil.VersionKey}, storedKeys(t, inner))
 }
 
-// Describe skips a header it cannot read or decode, and still succeeds. An
-// undecodable header also cannot be removed through DeleteLaunchTemplate.
+// A create that fails after claiming the name removes what it wrote, so no
+// body is orphaned and a retry under the same name succeeds.
+func TestAccountTermination_FailedCreateRemovesWhatItWrote(t *testing.T) {
+	for _, failing := range []string{"version", "header"} {
+		t.Run(failing+" write fails", func(t *testing.T) {
+			svc := setupTestService(t)
+			inner := svc.kv
+			svc.kv = &faultyKV{KeyValue: inner, failPut: func(key string) bool {
+				if failing == "header" {
+					return isHeaderKey(key, testAccountID+".")
+				}
+				return versionBodyTemplateID(strings.TrimPrefix(key, testAccountID+".")) != ""
+			}}
+
+			_, err := svc.CreateLaunchTemplate(context.Background(), &ec2.CreateLaunchTemplateInput{
+				LaunchTemplateName: aws.String("web"),
+				LaunchTemplateData: &ec2.RequestLaunchTemplateData{
+					UserData: aws.String(base64.StdEncoding.EncodeToString([]byte(ctUserData))),
+				},
+			}, testAccountID)
+			require.Error(t, err)
+			assert.Equal(t, awserrors.ErrorServerInternal, err.Error())
+
+			svc.kv = inner
+			assert.Equal(t, []string{kvutil.VersionKey}, storedKeys(t, inner))
+			retried := createTemplate(t, svc, "web", "t3.micro")
+			assert.Equal(t, []string{aws.StringValue(retried.LaunchTemplateId)}, describedTemplateIDs(t, svc, testAccountID))
+		})
+	}
+}
+
+// Describe skips a header it cannot read or decode, and still succeeds. Such
+// a header cannot be removed through DeleteLaunchTemplate either, and a read
+// failure is an internal error, never a missing template.
 func TestCurrentBehaviour_AccountTermination_DescribeLaunchTemplatesHidesUnreadableHeaders(t *testing.T) {
 	const brokenID = "lt-0badbadbadbadbad0"
 	brokenHeader := kvutil.AccountKey(testAccountID, brokenID)
@@ -252,7 +311,7 @@ func TestCurrentBehaviour_AccountTermination_DescribeLaunchTemplatesHidesUnreada
 			_, err = svc.kv.Put(t.Context(), brokenHeader, data)
 			require.NoError(t, err)
 			svc.kv = &faultyKV{KeyValue: svc.kv, failGet: func(key string) bool { return key == brokenHeader }}
-		}, awserrors.ErrorInvalidLaunchTemplateIdNotFound},
+		}, awserrors.ErrorServerInternal},
 	}
 	for _, tc := range tests {
 		t.Run(tc.name, func(t *testing.T) {
@@ -271,3 +330,124 @@ func TestCurrentBehaviour_AccountTermination_DescribeLaunchTemplatesHidesUnreada
 		})
 	}
 }
+
+// Account cleanup never deletes a live template: every key of a readable
+// template stays and its ID is reported kept. Headerless bodies and names are
+// removed once settled; a record younger than the settle window, and an
+// unreadable header with its bodies and name, are kept.
+func TestAccountCleanup_KeepsTemplatesAndRemovesSettledResidue(t *testing.T) {
+	svc := setupTestService(t)
+	live := createWithUserData(t, svc, "web")
+	liveKeys := accountKeys(t, svc.kv, testAccountID)
+	other := "999988887777"
+	_, err := svc.CreateLaunchTemplate(context.Background(), &ec2.CreateLaunchTemplateInput{
+		LaunchTemplateName: aws.String("theirs"), LaunchTemplateData: &ec2.RequestLaunchTemplateData{ImageId: aws.String("ami-123")},
+	}, other)
+	require.NoError(t, err)
+	orphanBody := versionKey(testAccountID, "lt-0orphan000000000", 1)
+	orphanName := nameKey(testAccountID, "gone")
+	brokenHeader := kvutil.AccountKey(testAccountID, "lt-0broken000000000")
+	brokenBody := versionKey(testAccountID, "lt-0broken000000000", 1)
+	brokenName := nameKey(testAccountID, "broken")
+	for key, value := range map[string]string{
+		orphanBody: `{"version_number":1}`, orphanName: "lt-0orphan000000000",
+		brokenHeader: "{not json", brokenBody: `{"version_number":1}`, brokenName: "lt-0broken000000000",
+	} {
+		_, err := svc.kv.Put(t.Context(), key, []byte(value))
+		require.NoError(t, err)
+	}
+	keep := sortedKeys(append(liveKeys, brokenHeader, brokenBody, brokenName)...)
+
+	out, err := svc.CleanupAccountResidue(t.Context(), testAccountID, time.Now().Add(-time.Hour))
+	require.NoError(t, err)
+	assert.Equal(t, []string{live}, out.KeptTemplateIDs)
+	assert.Equal(t, 1, out.OrphanBodiesUnsettled)
+	assert.Equal(t, 1, out.OrphanNamesUnsettled)
+	assert.Equal(t, sortedKeys(append(keep, orphanBody, orphanName)...), accountKeys(t, svc.kv, testAccountID))
+
+	out, err = svc.CleanupAccountResidue(t.Context(), testAccountID, time.Now().Add(time.Second))
+	require.NoError(t, err)
+	assert.Equal(t, []string{live}, out.KeptTemplateIDs)
+	assert.Equal(t, 1, out.OrphanBodiesRemoved)
+	assert.Equal(t, 1, out.OrphanNamesRemoved)
+	assert.Equal(t, keep, accountKeys(t, svc.kv, testAccountID))
+	assert.Equal(t, []string{live}, describedTemplateIDs(t, svc, testAccountID))
+	assert.Len(t, describedTemplateIDs(t, svc, other), 1)
+}
+
+// A public delete interrupted at its header leaves the template live, so the
+// account cleanup keeps it rather than finishing it; the public retry does.
+func TestAccountCleanup_InterruptedTemplateDeleteIsLeftToThePublicRetry(t *testing.T) {
+	svc := setupTestService(t)
+	id := createWithUserData(t, svc, "web")
+	header := kvutil.AccountKey(testAccountID, id)
+	kv := svc.kv
+	svc.kv = &faultyKV{KeyValue: kv, failDelete: func(key string) bool { return key == header }}
+	_, err := svc.DeleteLaunchTemplate(context.Background(), &ec2.DeleteLaunchTemplateInput{LaunchTemplateId: aws.String(id)}, testAccountID)
+	require.Error(t, err)
+	svc.kv = kv
+	require.Equal(t, []string{header}, accountKeys(t, kv, testAccountID), "only the header, the delete's anchor, remains")
+
+	out, err := svc.CleanupAccountResidue(t.Context(), testAccountID, time.Now().Add(time.Second))
+	require.NoError(t, err)
+	assert.Equal(t, []string{id}, out.KeptTemplateIDs)
+	assert.Equal(t, []string{header}, accountKeys(t, kv, testAccountID))
+
+	_, err = svc.DeleteLaunchTemplate(context.Background(), &ec2.DeleteLaunchTemplateInput{LaunchTemplateId: aws.String(id)}, testAccountID)
+	require.NoError(t, err)
+	assert.Empty(t, accountKeys(t, kv, testAccountID))
+}
+
+// A listing failure is returned and nothing is deleted.
+func TestAccountCleanup_TemplateListingFailureDeletesNothing(t *testing.T) {
+	svc := setupTestService(t)
+	id := createWithUserData(t, svc, "web")
+	kv := svc.kv
+	svc.kv = &faultyKV{KeyValue: kv, failKeys: func() bool { return true }}
+
+	_, err := svc.CleanupAccountResidue(t.Context(), testAccountID, time.Now())
+	require.Error(t, err)
+	svc.kv = kv
+	assert.Equal(t, []string{id}, describedTemplateIDs(t, svc, testAccountID))
+}
+
+func accountKeys(t *testing.T, kv jetstream.KeyValue, account string) []string {
+	t.Helper()
+	var mine []string
+	for _, key := range storedKeys(t, kv) {
+		if strings.HasPrefix(key, account+".") {
+			mine = append(mine, key)
+		}
+	}
+	return mine
+}
+
+// A settled orphan the store cannot read or delete is counted failed, never
+// removed, and is still there for the next run.
+func TestAccountCleanup_OrphanReadOrDeleteFailureIsCountedFailed(t *testing.T) {
+	for name, fault := range map[string]func(kv *faultyKV, key string){
+		"read":   func(kv *faultyKV, key string) { kv.failGet = func(k string) bool { return k == key } },
+		"delete": func(kv *faultyKV, key string) { kv.failDelete = func(k string) bool { return k == key } },
+	} {
+		t.Run(name, func(t *testing.T) {
+			svc := setupTestService(t)
+			body := versionKey(testAccountID, "lt-0orphan000000000", 1)
+			nameIdx := nameKey(testAccountID, "gone")
+			for key, value := range map[string]string{body: `{"version_number":1}`, nameIdx: "lt-0orphan000000000"} {
+				_, err := svc.kv.Put(t.Context(), key, []byte(value))
+				require.NoError(t, err)
+			}
+			inner := svc.kv
+			faulty := &faultyKV{KeyValue: inner}
+			fault(faulty, body)
+			svc.kv = faulty
+
+			out, err := svc.CleanupAccountResidue(t.Context(), testAccountID, time.Now().Add(time.Second))
+			require.NoError(t, err)
+			assert.Equal(t, 1, out.OrphanBodiesFailed)
+			assert.Zero(t, out.OrphanBodiesRemoved)
+			assert.Equal(t, 1, out.OrphanNamesRemoved, "a fault on one record does not stop the rest")
+			assert.Equal(t, []string{body}, accountKeys(t, inner, testAccountID))
+		})
+	}
+}

spinifex/domains/ec2/launchtemplate/service_impl.go

@@ -52,6 +52,9 @@ type LaunchTemplateServiceImpl struct {
 	config   *config.Config
 	natsConn *nats.Conn
 	kv       jetstream.KeyValue
+
+	// Optional: nil leaves the central tag index untouched by delete.
+	centralTags CentralTagStore
 }
 
 // NewLaunchTemplateServiceImplWithNATS creates a launch template service with NATS JetStream.
@@ -103,7 +106,12 @@ func versionPrefix(accountID, ltID string) string {
 func (s *LaunchTemplateServiceImpl) getHeaderByID(ctx context.Context, accountID, ltID string) (*LaunchTemplateHeader, jetstream.KeyValueEntry, error) {
 	entry, err := s.kv.Get(ctx, kvutil.AccountKey(accountID, ltID))
 	if err != nil {
-		return nil, nil, errors.New(awserrors.ErrorInvalidLaunchTemplateIdNotFound)
+		if errors.Is(err, jetstream.ErrKeyNotFound) {
+			return nil, nil, errors.New(awserrors.ErrorInvalidLaunchTemplateIdNotFound)
+		}
+		// A header the store cannot return now is not proven absent.
+		slog.ErrorContext(ctx, "Failed to read launch template header", "launchTemplateId", ltID, "err", err)
+		return nil, nil, errors.New(awserrors.ErrorServerInternal)
 	}
 	var h LaunchTemplateHeader
 	if err := json.Unmarshal(entry.Value(), &h); err != nil {
@@ -117,12 +125,19 @@ func (s *LaunchTemplateServiceImpl) getHeaderByID(ctx context.Context, accountID
 func (s *LaunchTemplateServiceImpl) getHeaderByName(ctx context.Context, accountID, name string) (*LaunchTemplateHeader, jetstream.KeyValueEntry, error) {
 	entry, err := s.kv.Get(ctx, nameKey(accountID, name))
 	if err != nil {
-		return nil, nil, errors.New(awserrors.ErrorInvalidLaunchTemplateNameNotFoundException)
+		if errors.Is(err, jetstream.ErrKeyNotFound) {
+			return nil, nil, errors.New(awserrors.ErrorInvalidLaunchTemplateNameNotFoundException)
+		}
+		slog.ErrorContext(ctx, "Failed to read launch template name index", "err", err)
+		return nil, nil, errors.New(awserrors.ErrorServerInternal)
 	}
 	ltID := string(entry.Value())
 	h, hentry, err := s.getHeaderByID(ctx, accountID, ltID)
 	if err != nil {
-		return nil, nil, errors.New(awserrors.ErrorInvalidLaunchTemplateNameNotFoundException)
+		if err.Error() == awserrors.ErrorInvalidLaunchTemplateIdNotFound {
+			return nil, nil, errors.New(awserrors.ErrorInvalidLaunchTemplateNameNotFoundException)
+		}
+		return nil, nil, err
 	}
 	return h, hentry, nil
 }
@@ -291,6 +306,7 @@ func (s *LaunchTemplateServiceImpl) CreateLaunchTemplate(ctx context.Context, in
 		Data:               data,
 	}
 	if err := s.putVersion(ctx, accountID, ltID, &rec); err != nil {
+		s.abandonCreate(ctx, accountID, name, ltID, false)
 		return nil, err
 	}
 
@@ -304,6 +320,7 @@ func (s *LaunchTemplateServiceImpl) CreateLaunchTemplate(ctx context.Context, in
 		Tags:                 awstags.Extract(input.TagSpecifications, launchTemplateTagResourceType),
 	}
 	if err := s.putHeader(ctx, accountID, &header); err != nil {
+		s.abandonCreate(ctx, accountID, name, ltID, true)
 		return nil, err
 	}
 
@@ -314,6 +331,39 @@ func (s *LaunchTemplateServiceImpl) CreateLaunchTemplate(ctx context.Context, in
 	}, nil
 }
 
+// abandonCreate removes what a failed create wrote, newest first. A failure
+// is logged; what remains has no header, and the account cleanup removes it
+// once it is older than an in-flight create could be.
+func (s *LaunchTemplateServiceImpl) abandonCreate(ctx context.Context, accountID, name, ltID string, wroteVersion bool) {
+	if wroteVersion {
+		if err := s.kv.Delete(ctx, versionKey(accountID, ltID, 1)); err != nil {
+			slog.ErrorContext(ctx, "CreateLaunchTemplate: failed to remove version 1 after a failed create", "launchTemplateId", ltID, "err", err)
+		}
+	}
+	if err := s.releaseName(ctx, accountID, name, ltID); err != nil {
+		slog.ErrorContext(ctx, "CreateLaunchTemplate: failed to release the name after a failed create", "launchTemplateId", ltID, "err", err)
+	}
+}
+
+// releaseName removes the name index only while it still names ltID, at the
+// revision read, so a name another template now holds is never released.
+func (s *LaunchTemplateServiceImpl) releaseName(ctx context.Context, accountID, name, ltID string) error {
+	key := nameKey(accountID, name)
+	entry, err := s.kv.Get(ctx, key)
+	switch {
+	case errors.Is(err, jetstream.ErrKeyNotFound):
+		return nil
+	case err != nil:
+		return fmt.Errorf("read name index: %w", err)
+	case string(entry.Value()) != ltID:
+		return nil
+	}
+	if err := s.kv.Delete(ctx, key, jetstream.LastRevision(entry.Revision())); err != nil {
+		return fmt.Errorf("delete name index: %w", err)
+	}
+	return nil
+}
+
 // errNameInUse is AWS's error for a launch template name already taken.
 func errNameInUse() error {
 	return awserrors.Errorf(awserrors.ErrorInvalidLaunchTemplateNameAlreadyExistsException, "Launch template name already in use.")
@@ -377,7 +427,7 @@ func (s *LaunchTemplateServiceImpl) CreateLaunchTemplateVersion(ctx context.Cont
 	if input.LaunchTemplateData == nil {
 		return nil, errors.New(awserrors.ErrorMissingParameter)
 	}
-	header, _, err := s.resolveHeader(ctx, accountID, input.LaunchTemplateId, input.LaunchTemplateName)
+	header, entry, err := s.resolveHeader(ctx, accountID, input.LaunchTemplateId, input.LaunchTemplateName)
 	if err != nil {
 		return nil, err
 	}
@@ -429,6 +479,13 @@ func (s *LaunchTemplateServiceImpl) CreateLaunchTemplateVersion(ctx context.Cont
 			slog.ErrorContext(ctx, "CreateLaunchTemplateVersion: version-key create failed", "next", next, "err", err)
 			return nil, errors.New(awserrors.ErrorServerInternal)
 		}
+		if err := s.commitVersion(ctx, accountID, header.LaunchTemplateId, entry); err != nil {
+			s.abandonVersion(ctx, accountID, header.LaunchTemplateId, next)
+			if errors.Is(err, errTemplateGone) {
+				return nil, templateNotFound(input.LaunchTemplateId, input.LaunchTemplateName)
+			}
+			return nil, errors.New(awserrors.ErrorServerInternal)
+		}
 		slog.InfoContext(ctx, "CreateLaunchTemplateVersion completed", "launchTemplateId", header.LaunchTemplateId, "version", next, "accountID", accountID)
 		return &ec2.CreateLaunchTemplateVersionOutput{
 			LaunchTemplateVersion: versionToEC2(header, &rec),
@@ -439,6 +496,53 @@ func (s *LaunchTemplateServiceImpl) CreateLaunchTemplateVersion(ctx context.Cont
 	return nil, errors.New(awserrors.ErrorServerInternal)
 }
 
+// errTemplateGone: the header was deleted while a version was being added.
+var errTemplateGone = errors.New("launch template deleted")
+
+// commitVersion rewrites the header's current value under a revision guard
+// after a new body is written. A delete that removed the header at any earlier
+// revision makes it fail, so the body is never left without its template.
+func (s *LaunchTemplateServiceImpl) commitVersion(ctx context.Context, accountID, ltID string, entry jetstream.KeyValueEntry) error {
+	key := kvutil.AccountKey(accountID, ltID)
+	for range maxVersionRetries {
+		_, err := s.kv.Update(ctx, key, entry.Value(), entry.Revision())
+		if err == nil {
+			return nil
+		}
+		if !errors.Is(err, jetstream.ErrKeyRevisionMismatch) && !errors.Is(err, jetstream.ErrKeyExists) {
+			slog.ErrorContext(ctx, "CreateLaunchTemplateVersion: header commit failed", "launchTemplateId", ltID, "err", err)
+			return err
+		}
+		next, gerr := s.kv.Get(ctx, key)
+		if errors.Is(gerr, jetstream.ErrKeyNotFound) {
+			return errTemplateGone
+		}
+		if gerr != nil {
+			return gerr
+		}
+		entry = next
+	}
+	return errors.New("header commit retries exhausted")
+}
+
+// abandonVersion removes a body whose header commit failed. A failure is
+// logged and leaves a headerless body, which no ID, name or launch resolves.
+func (s *LaunchTemplateServiceImpl) abandonVersion(ctx context.Context, accountID, ltID string, n int64) {
+	if err := s.kv.Delete(ctx, versionKey(accountID, ltID, n)); err != nil {
+		slog.ErrorContext(ctx, "CreateLaunchTemplateVersion: failed to remove an uncommitted version", "launchTemplateId", ltID, "version", n, "err", err)
+	}
+}
+
+// templateNotFound is the NotFound error for the template the request named.
+func templateNotFound(ltID, name *string) error {
+	if id := aws.StringValue(ltID); id != "" {
+		return awserrors.Errorf(awserrors.ErrorInvalidLaunchTemplateIdNotFound,
+			"The specified launch template, with template ID %s, does not exist.", id)
+	}
+	return awserrors.Errorf(awserrors.ErrorInvalidLaunchTemplateNameNotFoundException,
+		"The specified launch template, with template name %s, does not exist.", aws.StringValue(name))
+}
+
 // --- ModifyLaunchTemplate ---
 
 // ModifyLaunchTemplate implements the EC2 action. ClientToken is accepted but not stored: this only
@@ -481,8 +585,11 @@ func (s *LaunchTemplateServiceImpl) ModifyLaunchTemplate(ctx context.Context, in
 
 // --- DeleteLaunchTemplate ---
 
+// DeleteLaunchTemplate deletes a template and every version. A store failure
+// returns an error with the header still in place, so the template stays
+// visible and a retried delete finishes it.
 func (s *LaunchTemplateServiceImpl) DeleteLaunchTemplate(ctx context.Context, input *ec2.DeleteLaunchTemplateInput, accountID string) (*ec2.DeleteLaunchTemplateOutput, error) {
-	header, _, err := s.resolveHeader(ctx, accountID, input.LaunchTemplateId, input.LaunchTemplateName)
+	header, entry, err := s.resolveHeader(ctx, accountID, input.LaunchTemplateId, input.LaunchTemplateName)
 	if err != nil {
 		return nil, err
 	}
@@ -493,28 +600,92 @@ func (s *LaunchTemplateServiceImpl) DeleteLaunchTemplate(ctx context.Context, in
 	}
 	out := headerToEC2(header, latest)
 
-	// Delete the header first: the template immediately vanishes from every
-	// describe. Version bodies and the name index are best-effort cleanup.
-	if err := s.kv.Delete(ctx, kvutil.AccountKey(accountID, header.LaunchTemplateId)); err != nil {
+	if err := s.deleteTemplate(ctx, accountID, header, entry); err != nil {
+		slog.ErrorContext(ctx, "DeleteLaunchTemplate failed", "launchTemplateId", header.LaunchTemplateId, "accountID", accountID, "err", err)
 		return nil, errors.New(awserrors.ErrorServerInternal)
 	}
-	if err := s.kv.Delete(ctx, nameKey(accountID, header.LaunchTemplateName)); err != nil {
-		slog.WarnContext(ctx, "DeleteLaunchTemplate: name index cleanup failed", "name", header.LaunchTemplateName, "err", err)
-	}
-	nums, err := s.listVersionNumbers(ctx, accountID, header.LaunchTemplateId)
-	if err != nil {
-		slog.WarnContext(ctx, "DeleteLaunchTemplate: version enumeration failed, bodies left in place", "launchTemplateId", header.LaunchTemplateId, "err", err)
-	}
-	for _, n := range nums {
-		if err := s.kv.Delete(ctx, versionKey(accountID, header.LaunchTemplateId, n)); err != nil {
-			slog.WarnContext(ctx, "DeleteLaunchTemplate: version cleanup failed", "launchTemplateId", header.LaunchTemplateId, "version", n, "err", err)
-		}
-	}
 
 	slog.InfoContext(ctx, "DeleteLaunchTemplate completed", "launchTemplateId", header.LaunchTemplateId, "accountID", accountID)
 	return &ec2.DeleteLaunchTemplateOutput{LaunchTemplate: out}, nil
 }
 
+// maxHeaderDeleteAttempts bounds restarts when the header changes mid-delete.
+const maxHeaderDeleteAttempts = 3
+
+// deleteTemplate removes every version body, then the name index, then the
+// header at the revision read, and last any body a concurrent version create
+// added meanwhile. The header goes last so every earlier failure is retryable
+// through the public delete; a header changed in between restarts the sweep.
+func (s *LaunchTemplateServiceImpl) deleteTemplate(ctx context.Context, accountID string, header *LaunchTemplateHeader, entry jetstream.KeyValueEntry) error {
+	ltID := header.LaunchTemplateId
+	headerKey := kvutil.AccountKey(accountID, ltID)
+	for attempt := 1; ; attempt++ {
+		if err := s.deleteVersionBodies(ctx, accountID, ltID); err != nil {
+			return err
+		}
+		if err := s.releaseName(ctx, accountID, header.LaunchTemplateName, ltID); err != nil {
+			return err
+		}
+		err := s.kv.Delete(ctx, headerKey, jetstream.LastRevision(entry.Revision()))
+		if err == nil {
+			break
+		}
+		if (!errors.Is(err, jetstream.ErrKeyRevisionMismatch) && !errors.Is(err, jetstream.ErrKeyExists)) || attempt >= maxHeaderDeleteAttempts {
+			return fmt.Errorf("delete header: %w", err)
+		}
+		next, gerr := s.kv.Get(ctx, headerKey)
+		if errors.Is(gerr, jetstream.ErrKeyNotFound) {
+			break
+		}
+		if gerr != nil {
+			return fmt.Errorf("re-read header: %w", gerr)
+		}
+		entry = next
+	}
+	s.clearRecordTags(ctx, accountID, ltID)
+	// A version create in flight fails its header commit and removes its own
+	// body; this sweep catches one that committed nothing before stopping.
+	return s.deleteVersionBodies(ctx, accountID, ltID)
+}
+
+// deleteVersionBodies deletes every version body of ltID, stopping at the
+// first failure so the caller can keep the header for a retry.
+func (s *LaunchTemplateServiceImpl) deleteVersionBodies(ctx context.Context, accountID, ltID string) error {
+	nums, err := s.listVersionNumbers(ctx, accountID, ltID)
+	if err != nil {
+		return fmt.Errorf("list versions: %w", err)
+	}
+	for _, n := range nums {
+		if err := s.kv.Delete(ctx, versionKey(accountID, ltID, n)); err != nil {
+			return fmt.Errorf("delete version %d: %w", n, err)
+		}
+	}
+	return nil
+}
+
+// CentralTagStore clears a deleted template's entry from the central EC2 tag
+// index. Implemented by domains/ec2/tags.TagsServiceImpl.
+type CentralTagStore interface {
+	DeleteAllTags(ctx context.Context, accountID, resourceID string) error
+}
+
+// SetCentralTagStore injects the central tag index that delete clears.
+func (s *LaunchTemplateServiceImpl) SetCentralTagStore(st CentralTagStore) {
+	s.centralTags = st
+}
+
+// clearRecordTags drops a deleted template's tag-index entry. It runs only
+// after the header delete committed; a failure is logged, as a retried delete
+// would find nothing, and the entry is left as tag residue.
+func (s *LaunchTemplateServiceImpl) clearRecordTags(ctx context.Context, accountID, ltID string) {
+	if s.centralTags == nil {
+		return
+	}
+	if err := s.centralTags.DeleteAllTags(ctx, accountID, ltID); err != nil {
+		slog.ErrorContext(ctx, "central tag store clear failed", "launchTemplateId", ltID, "err", err)
+	}
+}
+
 // --- DeleteLaunchTemplateVersions ---
 
 func (s *LaunchTemplateServiceImpl) DeleteLaunchTemplateVersions(ctx context.Context, input *ec2.DeleteLaunchTemplateVersionsInput, accountID string) (*ec2.DeleteLaunchTemplateVersionsOutput, error) {

spinifex/domains/ec2/launchtemplate/template_lifecycle_test.go

@@ -0,0 +1,235 @@
+package launchtemplate
+
+//test:in-package — the interleavings are forced by wrapping the service's
+// unexported KV handle.
+
+import (
+	"context"
+	"testing"
+
+	"github.com/aws/aws-sdk-go/aws"
+	"github.com/aws/aws-sdk-go/service/ec2"
+	"github.com/mulgadc/spinifex/spinifex/foundation/aws/errors"
+	"github.com/mulgadc/spinifex/spinifex/foundation/state/kvutil"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+// raceDelete runs a whole DeleteLaunchTemplate on the unwrapped bucket, so the
+// hook that calls it is not re-entered.
+func raceDelete(t *testing.T, svc *LaunchTemplateServiceImpl, id string) {
+	t.Helper()
+	other := &LaunchTemplateServiceImpl{kv: svc.kv.(*faultyKV).KeyValue}
+	_, err := other.DeleteLaunchTemplate(context.Background(), &ec2.DeleteLaunchTemplateInput{LaunchTemplateId: aws.String(id)}, testAccountID)
+	require.NoError(t, err)
+}
+
+func newVersionInput(id string) *ec2.CreateLaunchTemplateVersionInput {
+	return &ec2.CreateLaunchTemplateVersionInput{
+		LaunchTemplateId:   aws.String(id),
+		LaunchTemplateData: &ec2.RequestLaunchTemplateData{InstanceType: aws.String("t3.small")},
+	}
+}
+
+// A version create paused before it writes its body or before it commits the
+// header, while a whole delete runs, fails NotFound and leaves nothing behind.
+func TestAccountTermination_VersionCreateRacingADeleteLeavesNoBody(t *testing.T) {
+	for _, pause := range []string{"before-body", "before-commit"} {
+		t.Run(pause, func(t *testing.T) {
+			svc := setupTestService(t)
+			id := createWithUserData(t, svc, "web")
+			header := kvutil.AccountKey(testAccountID, id)
+			raced := false
+			race := func(key string, at string) {
+				if raced || pause != at {
+					return
+				}
+				if (at == "before-body" && key == versionKey(testAccountID, id, 3)) || (at == "before-commit" && key == header) {
+					raced = true
+					raceDelete(t, svc, id)
+				}
+			}
+			svc.kv = &faultyKV{
+				KeyValue:     svc.kv,
+				beforeCreate: func(key string) { race(key, "before-body") },
+				beforeUpdate: func(key string) { race(key, "before-commit") },
+			}
+
+			_, err := svc.CreateLaunchTemplateVersion(context.Background(), newVersionInput(id), testAccountID)
+			require.True(t, raced)
+			require.Error(t, err)
+			code, _ := awserrors.ResolveErrorCode(err)
+			assert.Equal(t, awserrors.ErrorInvalidLaunchTemplateIdNotFound, code)
+			assert.Empty(t, accountKeys(t, svc.kv, testAccountID), "no headerless body is left")
+		})
+	}
+}
+
+// A version that commits before the delete removes the header moves the
+// header revision, so the delete restarts and removes the new body too.
+func TestAccountTermination_DeleteRacingACommittedVersionRemovesIt(t *testing.T) {
+	svc := setupTestService(t)
+	id := createWithUserData(t, svc, "web")
+	inner := svc.kv
+	header := kvutil.AccountKey(testAccountID, id)
+	raced := false
+	svc.kv = &faultyKV{KeyValue: inner, beforeDelete: func(key string) {
+		if key != header || raced {
+			return
+		}
+		raced = true
+		other := &LaunchTemplateServiceImpl{kv: inner}
+		_, err := other.CreateLaunchTemplateVersion(context.Background(), newVersionInput(id), testAccountID)
+		require.NoError(t, err)
+	}}
+
+	_, err := svc.DeleteLaunchTemplate(context.Background(), &ec2.DeleteLaunchTemplateInput{LaunchTemplateId: aws.String(id)}, testAccountID)
+	require.NoError(t, err)
+	assert.True(t, raced)
+	assert.Empty(t, accountKeys(t, inner, testAccountID))
+}
+
+// A header commit that fails removes the body it guarded and returns an
+// error; the template keeps exactly the versions it had.
+func TestAccountTermination_FailedVersionCommitRemovesTheBody(t *testing.T) {
+	svc := setupTestService(t)
+	id := createWithUserData(t, svc, "web")
+	before := accountKeys(t, svc.kv, testAccountID)
+	header := kvutil.AccountKey(testAccountID, id)
+	svc.kv = &faultyKV{KeyValue: svc.kv, failUpdate: func(key string) bool { return key == header }}
+
+	_, err := svc.CreateLaunchTemplateVersion(context.Background(), newVersionInput(id), testAccountID)
+	require.Error(t, err)
+	assert.Equal(t, awserrors.ErrorServerInternal, err.Error())
+	assert.Equal(t, before, accountKeys(t, svc.kv, testAccountID))
+}
+
+// A version create racing a default-version change commits at the new header
+// revision and keeps the change.
+func TestAccountTermination_VersionCommitKeepsAConcurrentHeaderChange(t *testing.T) {
+	svc := setupTestService(t)
+	id := createWithUserData(t, svc, "web")
+	inner := svc.kv
+	header := kvutil.AccountKey(testAccountID, id)
+	raced := false
+	svc.kv = &faultyKV{KeyValue: inner, beforeUpdate: func(key string) {
+		if key != header || raced {
+			return
+		}
+		raced = true
+		other := &LaunchTemplateServiceImpl{kv: inner}
+		_, err := other.ModifyLaunchTemplate(context.Background(), &ec2.ModifyLaunchTemplateInput{
+			LaunchTemplateId: aws.String(id), DefaultVersion: aws.String("2"),
+		}, testAccountID)
+		require.NoError(t, err)
+	}}
+
+	out, err := svc.CreateLaunchTemplateVersion(context.Background(), newVersionInput(id), testAccountID)
+	require.NoError(t, err)
+	assert.True(t, raced)
+	assert.Equal(t, int64(3), aws.Int64Value(out.LaunchTemplateVersion.VersionNumber))
+	h, _, err := svc.getHeaderByID(context.Background(), testAccountID, id)
+	require.NoError(t, err)
+	assert.Equal(t, int64(2), h.DefaultVersionNumber)
+}
+
+// Template data carrying a key name and a placement group reaches the
+// RunInstances input for $Default, $Latest and an explicit version.
+func TestCurrentBehaviour_AccountTermination_TemplateKeyAndGroupReachRunInstances(t *testing.T) {
+	svc := setupTestService(t)
+	out, err := svc.CreateLaunchTemplate(context.Background(), &ec2.CreateLaunchTemplateInput{
+		LaunchTemplateName: aws.String("web"),
+		LaunchTemplateData: &ec2.RequestLaunchTemplateData{
+			ImageId: aws.String("ami-123"), KeyName: aws.String("v1-key"),
+			Placement: &ec2.LaunchTemplatePlacementRequest{GroupName: aws.String("v1-group")},
+		},
+	}, testAccountID)
+	require.NoError(t, err)
+	id := aws.StringValue(out.LaunchTemplate.LaunchTemplateId)
+	_, err = svc.CreateLaunchTemplateVersion(context.Background(), &ec2.CreateLaunchTemplateVersionInput{
+		LaunchTemplateId: aws.String(id),
+		LaunchTemplateData: &ec2.RequestLaunchTemplateData{
+			ImageId: aws.String("ami-123"), KeyName: aws.String("v2-key"),
+			Placement: &ec2.LaunchTemplatePlacementRequest{GroupName: aws.String("v2-group")},
+		},
+	}, testAccountID)
+	require.NoError(t, err)
+
+	for version, want := range map[string]string{"$Default": "v1", "$Latest": "v2", "2": "v2", "1": "v1"} {
+		in := &ec2.RunInstancesInput{LaunchTemplate: &ec2.LaunchTemplateSpecification{LaunchTemplateId: aws.String(id), Version: aws.String(version)}}
+		require.NoError(t, ExpandRunInstances(context.Background(), svc, in, testAccountID), version)
+		assert.Equal(t, want+"-key", aws.StringValue(in.KeyName), version)
+		require.NotNil(t, in.Placement, version)
+		assert.Equal(t, want+"-group", aws.StringValue(in.Placement.GroupName), version)
+	}
+}
+
+// The reference reader returns what every version of every live template
+// names, and nothing from headerless bodies or other accounts.
+func TestAccountTemplateReferences_ReadsEveryVersionOfLiveTemplates(t *testing.T) {
+	svc := setupTestService(t)
+	out, err := svc.CreateLaunchTemplate(context.Background(), &ec2.CreateLaunchTemplateInput{
+		LaunchTemplateName: aws.String("web"),
+		LaunchTemplateData: &ec2.RequestLaunchTemplateData{KeyName: aws.String("default-key")},
+	}, testAccountID)
+	require.NoError(t, err)
+	id := aws.StringValue(out.LaunchTemplate.LaunchTemplateId)
+	_, err = svc.CreateLaunchTemplateVersion(context.Background(), &ec2.CreateLaunchTemplateVersionInput{
+		LaunchTemplateId: aws.String(id),
+		LaunchTemplateData: &ec2.RequestLaunchTemplateData{
+			KeyName:   aws.String("later-key"),
+			Placement: &ec2.LaunchTemplatePlacementRequest{GroupName: aws.String("later-group"), GroupId: aws.String("pg-0123456789abcdef0")},
+		},
+	}, testAccountID)
+	require.NoError(t, err)
+	_, err = svc.CreateLaunchTemplate(context.Background(), &ec2.CreateLaunchTemplateInput{
+		LaunchTemplateName: aws.String("theirs"), LaunchTemplateData: &ec2.RequestLaunchTemplateData{KeyName: aws.String("their-key")},
+	}, "999988887777")
+	require.NoError(t, err)
+	_, err = svc.kv.Put(t.Context(), versionKey(testAccountID, "lt-0orphan000000000", 1), []byte(`{"version_number":1,"data":{"KeyName":"orphan-key"}}`))
+	require.NoError(t, err)
+
+	refs, err := svc.AccountTemplateReferences(t.Context(), testAccountID)
+	require.NoError(t, err)
+	assert.ElementsMatch(t, []string{"default-key", "later-key"}, refs.KeyNames)
+	assert.Equal(t, []string{"later-group"}, refs.PlacementGroupNames)
+	assert.Equal(t, []string{"pg-0123456789abcdef0"}, refs.PlacementGroupIDs)
+}
+
+// A header or version the reader cannot read or decode, or a listing
+// failure, is an error: the unread record may name the resource.
+func TestAccountTemplateReferences_FailsClosed(t *testing.T) {
+	tests := map[string]func(t *testing.T, svc *LaunchTemplateServiceImpl, id string){
+		"listing fails": func(t *testing.T, svc *LaunchTemplateServiceImpl, _ string) {
+			svc.kv = &faultyKV{KeyValue: svc.kv, failKeys: func() bool { return true }}
+		},
+		"version unreadable": func(t *testing.T, svc *LaunchTemplateServiceImpl, id string) {
+			svc.kv = &faultyKV{KeyValue: svc.kv, failGet: func(key string) bool { return key == versionKey(testAccountID, id, 2) }}
+		},
+		"version undecodable": func(t *testing.T, svc *LaunchTemplateServiceImpl, id string) {
+			_, err := svc.kv.Put(t.Context(), versionKey(testAccountID, id, 2), []byte("{not json"))
+			require.NoError(t, err)
+		},
+		"version without data": func(t *testing.T, svc *LaunchTemplateServiceImpl, id string) {
+			_, err := svc.kv.Put(t.Context(), versionKey(testAccountID, id, 2), []byte(`{"version_number":2}`))
+			require.NoError(t, err)
+		},
+		"header unreadable": func(t *testing.T, svc *LaunchTemplateServiceImpl, id string) {
+			svc.kv = &faultyKV{KeyValue: svc.kv, failGet: func(key string) bool { return key == kvutil.AccountKey(testAccountID, id) }}
+		},
+		"header undecodable": func(t *testing.T, svc *LaunchTemplateServiceImpl, _ string) {
+			_, err := svc.kv.Put(t.Context(), kvutil.AccountKey(testAccountID, "lt-0broken000000000"), []byte("{not json"))
+			require.NoError(t, err)
+		},
+	}
+	for name, seed := range tests {
+		t.Run(name, func(t *testing.T) {
+			svc := setupTestService(t)
+			id := createWithUserData(t, svc, "web")
+			seed(t, svc, id)
+			refs, err := svc.AccountTemplateReferences(t.Context(), testAccountID)
+			require.Error(t, err)
+			assert.Nil(t, refs)
+		})
+	}
+}

spinifex/domains/ec2/launchtemplate/template_references.go

@@ -0,0 +1,95 @@
+package launchtemplate
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"strings"
+
+	"github.com/aws/aws-sdk-go/aws"
+	"github.com/mulgadc/spinifex/spinifex/foundation/state/kvutil"
+	"github.com/nats-io/nats.go/jetstream"
+)
+
+// TemplateReferences are the key-pair names, placement-group names and
+// placement-group IDs that the account's live launch-template versions carry.
+// They stay inside EC2; callers must not report them.
+type TemplateReferences struct {
+	KeyNames            []string
+	PlacementGroupNames []string
+	PlacementGroupIDs   []string
+}
+
+// AccountTemplateReferences reads every version body of every live template
+// in the account, since RunInstances may resolve any of them by $Default,
+// $Latest or number. It is complete or an error: an unlistable bucket, or a
+// header or body it cannot read or decode, is an error, because the unread
+// record may name the resource. Headerless bodies resolve from no launch.
+func (s *LaunchTemplateServiceImpl) AccountTemplateReferences(ctx context.Context, accountID string) (*TemplateReferences, error) {
+	keys, err := kvutil.Keys(ctx, s.kv)
+	if err != nil && !errors.Is(err, jetstream.ErrNoKeysFound) {
+		return nil, fmt.Errorf("list launch-template keys: %w", err)
+	}
+	prefix := accountID + "."
+	var live []string
+	for _, key := range keys {
+		if !isHeaderKey(key, prefix) {
+			continue
+		}
+		entry, err := s.kv.Get(ctx, key)
+		if errors.Is(err, jetstream.ErrKeyNotFound) {
+			continue
+		}
+		if err != nil {
+			return nil, errors.New("launch-template header unreadable")
+		}
+		var header LaunchTemplateHeader
+		if json.Unmarshal(entry.Value(), &header) != nil || prefix+header.LaunchTemplateId != key {
+			return nil, errors.New("launch-template header undecodable")
+		}
+		live = append(live, header.LaunchTemplateId)
+	}
+
+	out := &TemplateReferences{}
+	for _, ltID := range live {
+		bodyPrefix := versionPrefix(accountID, ltID)
+		for _, key := range keys {
+			if !strings.HasPrefix(key, bodyPrefix) {
+				continue
+			}
+			if err := s.addVersionReferences(ctx, key, out); err != nil {
+				return nil, err
+			}
+		}
+	}
+	return out, nil
+}
+
+// addVersionReferences adds what one version body would launch with. The
+// body is never logged: it holds template data and user data.
+func (s *LaunchTemplateServiceImpl) addVersionReferences(ctx context.Context, key string, out *TemplateReferences) error {
+	entry, err := s.kv.Get(ctx, key)
+	if errors.Is(err, jetstream.ErrKeyNotFound) {
+		return nil
+	}
+	if err != nil {
+		return errors.New("launch-template version unreadable")
+	}
+	var rec LaunchTemplateVersionRec
+	if json.Unmarshal(entry.Value(), &rec) != nil || rec.Data == nil {
+		return errors.New("launch-template version undecodable")
+	}
+	if name := aws.StringValue(rec.Data.KeyName); name != "" {
+		out.KeyNames = append(out.KeyNames, name)
+	}
+	if p := rec.Data.Placement; p != nil {
+		if name := aws.StringValue(p.GroupName); name != "" {
+			out.PlacementGroupNames = append(out.PlacementGroupNames, name)
+		}
+		if id := aws.StringValue(p.GroupId); id != "" {
+			out.PlacementGroupIDs = append(out.PlacementGroupIDs, id)
+		}
+	}
+	return nil
+}

spinifex/domains/ec2/placementgroup/account_cleanup.go

@@ -0,0 +1,85 @@
+package placementgroup
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"log/slog"
+	"strings"
+
+	"github.com/mulgadc/spinifex/spinifex/foundation/state/kvutil"
+	"github.com/nats-io/nats.go/jetstream"
+)
+
+// BlockedGroup is a group the owner refused to delete because its member map
+// still names instances. Only the count is kept; member IDs never leave.
+type BlockedGroup struct {
+	GroupID     string
+	MemberCount int
+}
+
+// AccountCleanup is what one CleanupAccount run did, by group ID. It is the
+// owner's account of its attempts, not proof of absence, which only a later
+// ScanAccount gives.
+type AccountCleanup struct {
+	// DeletedGroupIDs were removed by the owner delete or were already gone.
+	DeletedGroupIDs []string
+	// FailedGroupIDs hit a store failure; the record remains.
+	FailedGroupIDs []string
+	// Blocked groups still hold instances and are never forced.
+	Blocked []BlockedGroup
+	// Referenced maps a kept group ID to the reason the caller's predicate gave.
+	Referenced map[string]string
+}
+
+// CleanupAccount deletes every readable, empty placement group of the account
+// through the same revision-guarded delete DeletePlacementGroup runs. A group
+// holding instances is reported blocked, and one for whose ID or name
+// referenced returns a reason is kept; unreadable records are left for the owner scan to report. A
+// listing failure is returned and nothing is deleted.
+func (s *PlacementGroupServiceImpl) CleanupAccount(ctx context.Context, accountID string, referenced func(groupID, groupName string) string) (*AccountCleanup, error) {
+	keys, err := kvutil.Keys(ctx, s.kv)
+	if err != nil && !errors.Is(err, jetstream.ErrNoKeysFound) {
+		return nil, fmt.Errorf("list placement-group keys: %w", err)
+	}
+
+	prefix := accountID + "."
+	out := &AccountCleanup{Referenced: map[string]string{}}
+	for _, key := range keys {
+		if key == kvutil.VersionKey || !strings.HasPrefix(key, prefix) {
+			continue
+		}
+		entry, err := s.kv.Get(ctx, key)
+		if err != nil {
+			continue
+		}
+		var record PlacementGroupRecord
+		if err := json.Unmarshal(entry.Value(), &record); err != nil || record.GroupId == "" {
+			continue
+		}
+		if referenced != nil {
+			reason := referenced(record.GroupId, record.GroupName)
+			if reason == "" {
+				reason = referenced(record.GroupId, strings.TrimPrefix(key, prefix))
+			}
+			if reason != "" {
+				out.Referenced[record.GroupId] = reason
+				continue
+			}
+		}
+		members, err := s.deleteGroup(ctx, accountID, key)
+		switch {
+		case err == nil, errors.Is(err, errPGAbsent):
+			out.DeletedGroupIDs = append(out.DeletedGroupIDs, record.GroupId)
+		case errors.Is(err, errPGInUse):
+			out.Blocked = append(out.Blocked, BlockedGroup{GroupID: record.GroupId, MemberCount: members})
+		default:
+			slog.WarnContext(ctx, "Placement-group account cleanup: delete failed", "groupId", record.GroupId, "err", err)
+			out.FailedGroupIDs = append(out.FailedGroupIDs, record.GroupId)
+		}
+	}
+	slog.InfoContext(ctx, "Placement-group account cleanup completed", "accountID", accountID,
+		"deleted", len(out.DeletedGroupIDs), "failed", len(out.FailedGroupIDs), "blocked", len(out.Blocked), "referenced", len(out.Referenced))
+	return out, nil
+}

spinifex/domains/ec2/placementgroup/account_termination_current_behaviour_test.go

@@ -21,16 +21,18 @@ import (
 )
 
 // The TestCurrentBehaviour_AccountTermination_ tests pin how placement groups
-// are stored and what Describe and Delete do with them today. They describe
-// current behaviour, not the target design; a failure means it changed.
+// are stored and what Describe and Delete do with them today; a failure means
+// it changed. The TestAccountTermination_ tests pin the revision-guarded delete.
 
-// hookedKV fails reads of failGet and runs afterGet once, after the first read
-// of key returns, standing in for a writer on another node.
+// hookedKV fails reads of failGet and runs afterGet once, after read number
+// skip+1 of key returns, standing in for a writer on another node.
 type hookedKV struct {
 	jetstream.KeyValue
 
 	failGet  string
 	key      string
+	skip     int
+	reads    int
 	once     sync.Once
 	afterGet func()
 }
@@ -41,7 +43,9 @@ func (k *hookedKV) Get(ctx context.Context, key string) (jetstream.KeyValueEntry
 	}
 	entry, err := k.KeyValue.Get(ctx, key)
 	if key == k.key && k.afterGet != nil {
-		k.once.Do(k.afterGet)
+		if k.reads++; k.reads > k.skip {
+			k.once.Do(k.afterGet)
+		}
 	}
 	return entry, err
 }
@@ -92,8 +96,9 @@ func TestCurrentBehaviour_AccountTermination_PlacementGroupStorageLayout(t *test
 	assert.Empty(t, record.NodeInstances)
 }
 
-// Describe skips a record it cannot read or decode, and still succeeds. An
-// undecodable record also cannot be removed through DeletePlacementGroup.
+// Describe skips a record it cannot read or decode, and still succeeds. Such
+// a record cannot be removed through DeletePlacementGroup either, and a read
+// failure is an internal error, never an unknown group.
 func TestCurrentBehaviour_AccountTermination_DescribePlacementGroupsHidesUnreadableRecords(t *testing.T) {
 	brokenKey := kvutil.AccountKey(testAccountID, "broken")
 	tests := []struct {
@@ -108,7 +113,7 @@ func TestCurrentBehaviour_AccountTermination_DescribePlacementGroupsHidesUnreada
 		{"record the store cannot read", func(t *testing.T, svc *PlacementGroupServiceImpl) {
 			createTestGroup(t, svc, "broken", ec2.PlacementStrategySpread)
 			svc.kv = &hookedKV{KeyValue: svc.kv, failGet: brokenKey}
-		}, awserrors.ErrorInvalidPlacementGroupUnknown},
+		}, awserrors.ErrorServerInternal},
 	}
 	for _, tc := range tests {
 		t.Run(tc.name, func(t *testing.T) {
@@ -128,10 +133,10 @@ func TestCurrentBehaviour_AccountTermination_DescribePlacementGroupsHidesUnreada
 	}
 }
 
-// Pins current behaviour that differs from AWS: the delete reads the record,
-// checks membership, then deletes without a revision guard. A launch that
-// records an instance between the two is deleted with the group.
-func TestCurrentBehaviour_AccountTermination_DeletePlacementGroupDropsAConcurrentMember(t *testing.T) {
+// The delete commits only at the revision whose membership it checked. A
+// launch that records an instance between the check and the delete wins: the
+// delete re-reads, refuses with InUse, and the member stays.
+func TestAccountTermination_DeletePlacementGroupLosesToAConcurrentMember(t *testing.T) {
 	svc := setupTestService(t)
 	createTestGroup(t, svc, "cluster-a", ec2.PlacementStrategyCluster)
 	key := kvutil.AccountKey(testAccountID, "cluster-a")
@@ -151,17 +156,162 @@ func TestCurrentBehaviour_AccountTermination_DeletePlacementGroupDropsAConcurren
 	}}
 
 	_, err := svc.DeletePlacementGroup(context.Background(), &ec2.DeletePlacementGroupInput{GroupName: aws.String("cluster-a")}, testAccountID)
-	require.NoError(t, err, "the delete succeeds although the group gained a member")
+	require.Error(t, err)
+	assert.Equal(t, awserrors.ErrorInvalidPlacementGroupInUse, err.Error())
 	require.NoError(t, reserveErr)
 	require.NoError(t, finalizeErr, "the interleaved launch committed its member")
 
-	_, err = launcher.kv.Get(t.Context(), key)
-	assert.ErrorIs(t, err, jetstream.ErrKeyNotFound)
-	assert.Empty(t, describedGroupNames(t, launcher, testAccountID))
+	entry, err := launcher.kv.Get(t.Context(), key)
+	require.NoError(t, err)
+	var record PlacementGroupRecord
+	require.NoError(t, json.Unmarshal(entry.Value(), &record))
+	assert.Equal(t, map[string][]string{"node-1": {"i-0123456789abcdef0"}}, record.NodeInstances)
 
-	// The member's later shutdown finds no group and reports success.
+	// Once the member leaves, the same delete succeeds.
 	_, err = launcher.RemoveInstance(context.Background(), &RemoveInstanceInput{
 		GroupName: "cluster-a", NodeName: "node-1", InstanceID: "i-0123456789abcdef0",
 	}, testAccountID)
-	assert.NoError(t, err)
+	require.NoError(t, err)
+	_, err = launcher.DeletePlacementGroup(context.Background(), &ec2.DeletePlacementGroupInput{GroupName: aws.String("cluster-a")}, testAccountID)
+	require.NoError(t, err)
+	assert.Empty(t, describedGroupNames(t, launcher, testAccountID))
+}
+
+// A reservation placeholder is a launch in flight, not a member, so the
+// delete may win; the launch's finalize then finds no group and fails, which
+// makes the launcher roll its instances back rather than lose them.
+func TestAccountTermination_DeletePlacementGroupBeforeFinalizeFailsTheLaunch(t *testing.T) {
+	svc := setupTestService(t)
+	createTestGroup(t, svc, "cluster-a", ec2.PlacementStrategyCluster)
+	reserved, err := svc.ReserveClusterNode(context.Background(), &ReserveClusterNodeInput{
+		GroupName: "cluster-a", EligibleNodes: []string{"node-1"},
+	}, testAccountID)
+	require.NoError(t, err)
+
+	_, err = svc.DeletePlacementGroup(context.Background(), &ec2.DeletePlacementGroupInput{GroupName: aws.String("cluster-a")}, testAccountID)
+	require.NoError(t, err)
+
+	_, err = svc.FinalizeClusterInstances(context.Background(), &FinalizeClusterInstancesInput{
+		GroupName: "cluster-a", NodeInstances: map[string][]string{reserved.TargetNode: {"i-0123456789abcdef0"}},
+	}, testAccountID)
+	require.Error(t, err)
+	assert.Equal(t, awserrors.ErrorInvalidPlacementGroupUnknown, err.Error())
+	assert.Empty(t, describedGroupNames(t, svc, testAccountID))
+}
+
+// addMember reserves a node in a cluster group and records one instance on it.
+func addMember(t *testing.T, svc *PlacementGroupServiceImpl, group, instanceID string) error {
+	t.Helper()
+	reserved, err := svc.ReserveClusterNode(context.Background(), &ReserveClusterNodeInput{
+		GroupName: group, EligibleNodes: []string{"node-1"},
+	}, testAccountID)
+	if err != nil {
+		return err
+	}
+	_, err = svc.FinalizeClusterInstances(context.Background(), &FinalizeClusterInstancesInput{
+		GroupName: group, NodeInstances: map[string][]string{reserved.TargetNode: {instanceID}},
+	}, testAccountID)
+	return err
+}
+
+// Account cleanup deletes empty groups through the revision-guarded delete,
+// reports a group holding instances as blocked with its member count only,
+// and leaves records it cannot read for the owner scan to report.
+func TestAccountCleanup_DeletesEmptyGroupsAndBlocksOnMembers(t *testing.T) {
+	svc := setupTestService(t)
+	empty := createTestGroup(t, svc, "empty", ec2.PlacementStrategySpread)
+	held := createTestGroup(t, svc, "held", ec2.PlacementStrategyCluster)
+	require.NoError(t, addMember(t, svc, "held", "i-0123456789abcdef0"))
+	_, err := svc.kv.Put(t.Context(), kvutil.AccountKey(testAccountID, "broken"), []byte("{not json"))
+	require.NoError(t, err)
+	other := "999988887777"
+	_, err = svc.CreatePlacementGroup(context.Background(), &ec2.CreatePlacementGroupInput{
+		GroupName: aws.String("theirs"), Strategy: aws.String(ec2.PlacementStrategySpread),
+	}, other)
+	require.NoError(t, err)
+
+	out, err := svc.CleanupAccount(t.Context(), testAccountID, nil)
+	require.NoError(t, err)
+	assert.Equal(t, []string{aws.StringValue(empty.GroupId)}, out.DeletedGroupIDs)
+	assert.Equal(t, []BlockedGroup{{GroupID: aws.StringValue(held.GroupId), MemberCount: 1}}, out.Blocked)
+	assert.Empty(t, out.FailedGroupIDs)
+
+	assert.Equal(t, []string{"held"}, describedGroupNames(t, svc, testAccountID))
+	assert.Equal(t, []string{"theirs"}, describedGroupNames(t, svc, other))
+	_, err = svc.kv.Get(t.Context(), kvutil.AccountKey(testAccountID, "broken"))
+	require.NoError(t, err, "an unreadable record is never deleted")
+}
+
+// A launch that records a member between the cleanup's membership check and
+// its delete wins: the cleanup reports the group blocked and the member stays.
+func TestAccountCleanup_ConcurrentMemberBlocksTheDelete(t *testing.T) {
+	svc := setupTestService(t)
+	group := createTestGroup(t, svc, "cluster-a", ec2.PlacementStrategyCluster)
+	key := kvutil.AccountKey(testAccountID, "cluster-a")
+	launcher := &PlacementGroupServiceImpl{kv: svc.kv, region: svc.region}
+	var memberErr error
+	// Read 1 is the cleanup's listing read; read 2 is the delete's check.
+	svc.kv = &hookedKV{KeyValue: svc.kv, key: key, skip: 1, afterGet: func() {
+		memberErr = addMember(t, launcher, "cluster-a", "i-0123456789abcdef0")
+	}}
+
+	out, err := svc.CleanupAccount(t.Context(), testAccountID, nil)
+	require.NoError(t, err)
+	require.NoError(t, memberErr, "the interleaved launch committed its member")
+	assert.Equal(t, []BlockedGroup{{GroupID: aws.StringValue(group.GroupId), MemberCount: 1}}, out.Blocked)
+	assert.Empty(t, out.DeletedGroupIDs)
+	assert.Equal(t, map[string][]string{"node-1": {"i-0123456789abcdef0"}}, readRecord(t, launcher, "cluster-a").NodeInstances)
+}
+
+// A group the caller's reference check names is kept with the given reason,
+// whether it is matched by its name or its group ID.
+func TestAccountCleanup_KeepsReferencedGroups(t *testing.T) {
+	svc := setupTestService(t)
+	named := createTestGroup(t, svc, "named", ec2.PlacementStrategySpread)
+	byID := createTestGroup(t, svc, "by-id", ec2.PlacementStrategySpread)
+	free := createTestGroup(t, svc, "free", ec2.PlacementStrategySpread)
+	referenced := func(id, name string) string {
+		switch {
+		case name == "named":
+			return "spot-request-reference"
+		case id == aws.StringValue(byID.GroupId):
+			return "launch-template-reference"
+		}
+		return ""
+	}
+
+	out, err := svc.CleanupAccount(t.Context(), testAccountID, referenced)
+	require.NoError(t, err)
+	assert.Equal(t, map[string]string{
+		aws.StringValue(named.GroupId): "spot-request-reference",
+		aws.StringValue(byID.GroupId):  "launch-template-reference",
+	}, out.Referenced)
+	assert.Equal(t, []string{aws.StringValue(free.GroupId)}, out.DeletedGroupIDs)
+	assert.ElementsMatch(t, []string{"named", "by-id"}, describedGroupNames(t, svc, testAccountID))
+}
+
+// A listing failure is returned and nothing is deleted.
+func TestAccountCleanup_ListingFailureDeletesNothing(t *testing.T) {
+	svc := setupTestService(t)
+	createTestGroup(t, svc, "empty", ec2.PlacementStrategySpread)
+	kv := svc.kv
+	svc.kv = &listFailingKV{KeyValue: kv}
+
+	_, err := svc.CleanupAccount(t.Context(), testAccountID, nil)
+	require.Error(t, err)
+	svc.kv = kv
+	assert.Equal(t, []string{"empty"}, describedGroupNames(t, svc, testAccountID))
+}
+
+// listFailingKV fails every key listing.
+type listFailingKV struct {
+	jetstream.KeyValue
+}
+
+func (k *listFailingKV) ListKeys(context.Context, ...jetstream.WatchOpt) (jetstream.KeyLister, error) {
+	return nil, errors.New("injected listing failure")
+}
+
+func (k *listFailingKV) Keys(context.Context, ...jetstream.WatchOpt) ([]string, error) {
+	return nil, errors.New("injected listing failure")
 }

spinifex/domains/ec2/placementgroup/service_impl.go

@@ -52,6 +52,32 @@ type PlacementGroupServiceImpl struct {
 	natsConn *nats.Conn
 	kv       jetstream.KeyValue
 	region   string
+
+	// Optional: nil leaves the central tag index untouched by delete.
+	centralTags CentralTagStore
+}
+
+// CentralTagStore clears a deleted group's entry from the central EC2 tag
+// index. Implemented by domains/ec2/tags.TagsServiceImpl.
+type CentralTagStore interface {
+	DeleteAllTags(ctx context.Context, accountID, resourceID string) error
+}
+
+// SetCentralTagStore injects the central tag index that delete clears.
+func (s *PlacementGroupServiceImpl) SetCentralTagStore(st CentralTagStore) {
+	s.centralTags = st
+}
+
+// clearRecordTags drops a deleted group's tag-index entry. It runs only after
+// the record delete committed; a failure is logged, as the group is gone and a
+// retried delete would find nothing, and the entry is left as tag residue.
+func (s *PlacementGroupServiceImpl) clearRecordTags(ctx context.Context, accountID, groupID string) {
+	if s.centralTags == nil || groupID == "" {
+		return
+	}
+	if err := s.centralTags.DeleteAllTags(ctx, accountID, groupID); err != nil {
+		slog.ErrorContext(ctx, "central tag store clear failed", "groupId", groupID, "err", err)
+	}
 }
 
 // NewPlacementGroupServiceImplWithNATS creates a placement group service with NATS JetStream.
@@ -142,37 +168,19 @@ func (s *PlacementGroupServiceImpl) CreatePlacementGroup(ctx context.Context, in
 	}, nil
 }
 
-// DeletePlacementGroup deletes a placement group.
+// DeletePlacementGroup deletes a placement group that holds no instances.
 func (s *PlacementGroupServiceImpl) DeletePlacementGroup(ctx context.Context, input *ec2.DeletePlacementGroupInput, accountID string) (*ec2.DeletePlacementGroupOutput, error) {
 	if input.GroupName == nil || *input.GroupName == "" {
 		return nil, errors.New(awserrors.ErrorMissingParameter)
 	}
 
 	groupName := *input.GroupName
-	key := kvutil.AccountKey(accountID, groupName)
-
-	entry, err := s.kv.Get(ctx, key)
-	if err != nil {
-		// Lower case, unlike DescribePlacementGroups' message for the same code.
-		return nil, awserrors.Errorf(awserrors.ErrorInvalidPlacementGroupUnknown, "The placement group '%s' is unknown.", groupName)
-	}
-
-	var record PlacementGroupRecord
-	if err := json.Unmarshal(entry.Value(), &record); err != nil {
-		return nil, errors.New(awserrors.ErrorServerInternal)
-	}
-
-	// Check for running instances
-	instanceCount := 0
-	for _, ids := range record.NodeInstances {
-		instanceCount += len(ids)
-	}
-	if instanceCount > 0 {
-		return nil, errors.New(awserrors.ErrorInvalidPlacementGroupInUse)
-	}
-
-	if err := s.kv.Delete(ctx, key); err != nil {
-		return nil, errors.New(awserrors.ErrorServerInternal)
+	if _, err := s.deleteGroup(ctx, accountID, kvutil.AccountKey(accountID, groupName)); err != nil {
+		if errors.Is(err, errPGAbsent) {
+			// Lower case, unlike DescribePlacementGroups' message for the same code.
+			return nil, awserrors.Errorf(awserrors.ErrorInvalidPlacementGroupUnknown, "The placement group '%s' is unknown.", groupName)
+		}
+		return nil, mapCASError(ctx, "DeletePlacementGroup", groupName, accountID, err)
 	}
 
 	slog.InfoContext(ctx, "DeletePlacementGroup completed", "groupName", groupName, "accountID", accountID)
@@ -180,6 +188,26 @@ func (s *PlacementGroupServiceImpl) DeletePlacementGroup(ctx context.Context, in
 	return &ec2.DeletePlacementGroupOutput{}, nil
 }
 
+// deleteGroup removes the record at key only while it names no instance, at
+// the revision that check read, so a reservation committed in between sends
+// the delete back to re-check. Its tag-index entry is cleared afterwards.
+// The returned count is the members seen by the last check.
+func (s *PlacementGroupServiceImpl) deleteGroup(ctx context.Context, accountID, key string) (int, error) {
+	members := 0
+	record, err := kvutil.DeleteIf(ctx, s.kv, key, casConfig(), func(record *PlacementGroupRecord) error {
+		members = memberCount(record.NodeInstances)
+		if members > 0 {
+			return errPGInUse
+		}
+		return nil
+	})
+	if err != nil {
+		return members, err
+	}
+	s.clearRecordTags(ctx, accountID, record.GroupId)
+	return 0, nil
+}
+
 var describePlacementGroupsValidFilters = map[string]bool{
 	"group-id":     true,
 	"strategy":     true,
@@ -324,11 +352,12 @@ func pgMatchesFilters(record *PlacementGroupRecord, filters map[string][]string)
 const maxCASRetries = 5
 
 // errPGAbsent and errPGContended mark the two kvutil outcomes that map to
-// specific AWS errors rather than a generic internal failure. Neither
-// escapes this file.
+// specific AWS errors rather than a generic internal failure, and neither
+// escapes this file. errPGInUse is the AWS error itself.
 var (
 	errPGAbsent    = errors.New("placementgroup: record absent")
 	errPGContended = errors.New("placementgroup: CAS retries exhausted")
+	errPGInUse     = errors.New(awserrors.ErrorInvalidPlacementGroupInUse)
 )
 
 // mapCASError turns a kvutil failure into the AWS error the caller returns.

spinifex/domains/ec2/snapshot/service_impl.go

@@ -662,10 +662,10 @@ func snapshotMatchesFilters(cfg ebsmetadata.Snapshot, filters map[string][]strin
 	return awsfilters.MatchesTags(filters, cfg.Tags)
 }
 
-// snapshotInUseByVolumes checks if any volume was created from the given
-// snapshot. A clone records its source in its ebsmetadata document; missing one
-// here would let a delete strip the chunks a live clone still reads.
-func (s *SnapshotServiceImpl) snapshotInUseByVolumes(ctx context.Context, snapshotID string) (bool, error) {
+// snapshotInUseByVolumes reports whether any volume was created from the given
+// snapshot, and the ID of one owned by accountID if any. A clone records its
+// source in its ebsmetadata document; missing one would strip chunks it reads.
+func (s *SnapshotServiceImpl) snapshotInUseByVolumes(ctx context.Context, snapshotID, accountID string) (bool, string, error) {
 	// Whole-cluster, and not scoping this to the snapshot's own account is the
 	// point: launching from a system AMI writes a root volume owned by the
 	// launching tenant whose SnapshotID is the system snapshot, so a clone
@@ -675,10 +675,34 @@ func (s *SnapshotServiceImpl) snapshotInUseByVolumes(ctx context.Context, snapsh
 	// the snapshot's chunks, and this answer gates deleting them.
 	volumes, err := s.metadata.ListAllVolumesStrict(ctx)
 	if err != nil {
-		return false, fmt.Errorf("snapshotInUseByVolumes: failed to list volumes: %w", err)
+		return false, "", fmt.Errorf("snapshotInUseByVolumes: failed to list volumes: %w", err)
 	}
+	inUse := false
 	for _, volume := range volumes {
-		if volume.SnapshotID == snapshotID {
+		if volume.SnapshotID != snapshotID {
+			continue
+		}
+		if volume.TenantID == accountID {
+			return true, volume.VolumeID, nil
+		}
+		inUse = true
+	}
+	return inUse, "", nil
+}
+
+// amiInUseBySnapshot checks if any registered AMI names the given snapshot as
+// its root device. AWS refuses DeleteSnapshot while a registered AMI
+// references the snapshot, the same way it refuses one backing a volume.
+func (s *SnapshotServiceImpl) amiInUseBySnapshot(ctx context.Context, snapshotID string) (bool, error) {
+	// Whole-cluster and strict for the same reason as the volume check above:
+	// a system AMI can be registered by an account other than the snapshot's
+	// owner, and an AMI this cannot read is not evidence that none exists.
+	amis, err := s.metadata.ListAMIsStrict(ctx)
+	if err != nil {
+		return false, fmt.Errorf("amiInUseBySnapshot: failed to list AMIs: %w", err)
+	}
+	for _, ami := range amis {
+		if ami.SnapshotID == snapshotID {
 			return true, nil
 		}
 	}
@@ -715,13 +739,29 @@ func (s *SnapshotServiceImpl) DeleteSnapshot(ctx context.Context, input *ec2.Del
 	}
 
 	// Check if any volumes were created from this snapshot
-	inUse, err := s.snapshotInUseByVolumes(ctx, snapshotID)
+	inUse, volumeID, err := s.snapshotInUseByVolumes(ctx, snapshotID, accountID)
 	if err != nil {
 		slog.ErrorContext(ctx, "DeleteSnapshot failed to check snapshot usage", "snapshotId", snapshotID, "err", err)
 		return nil, errors.New(awserrors.ErrorServerInternal)
 	}
 	if inUse {
-		slog.InfoContext(ctx, "DeleteSnapshot blocked: snapshot in use by volume", "snapshotId", snapshotID)
+		slog.InfoContext(ctx, "DeleteSnapshot blocked: snapshot in use by volume", "snapshotId", snapshotID, "volumeId", volumeID)
+		// The registered message blames AMIs. Name the volume only when the
+		// caller owns it: the scan is whole-cluster.
+		if volumeID == "" {
+			return nil, awserrors.Errorf(awserrors.ErrorInvalidSnapshotInUse, "The snapshot '%s' is in use by a volume.", snapshotID)
+		}
+		return nil, awserrors.Errorf(awserrors.ErrorInvalidSnapshotInUse, "The snapshot '%s' is in use by volume '%s'.", snapshotID, volumeID)
+	}
+
+	// Check if any registered AMI names this snapshot as its root device
+	amiInUse, err := s.amiInUseBySnapshot(ctx, snapshotID)
+	if err != nil {
+		slog.ErrorContext(ctx, "DeleteSnapshot failed to check AMI usage", "snapshotId", snapshotID, "err", err)
+		return nil, errors.New(awserrors.ErrorServerInternal)
+	}
+	if amiInUse {
+		slog.InfoContext(ctx, "DeleteSnapshot blocked: snapshot in use by registered AMI", "snapshotId", snapshotID)
 		return nil, errors.New(awserrors.ErrorInvalidSnapshotInUse)
 	}
 	if s.provider == nil {

spinifex/domains/ec2/snapshot/service_impl_test.go

@@ -205,9 +205,10 @@ func TestSnapshotInUseByVolumes_EncryptedVolume(t *testing.T) {
 		VolumeID: "vol-fromsnap", CapacityGiB: 50, SnapshotID: "snap-src", Encrypted: true,
 	})
 
-	inUse, err := svc.snapshotInUseByVolumes(context.Background(), "snap-src")
+	inUse, volumeID, err := svc.snapshotInUseByVolumes(context.Background(), "snap-src", testAccountID)
 	require.NoError(t, err)
-	assert.True(t, inUse, "encrypted volume's SnapshotID must be visible through the envelope")
+	assert.True(t, inUse)
+	assert.Equal(t, "vol-fromsnap", volumeID, "encrypted volume's SnapshotID must be visible through the envelope")
 }
 
 // TestCreateSnapshot_VolumeNotFound tests creating a snapshot from non-existent volume.
@@ -459,6 +460,18 @@ func TestDeleteSnapshot_WrongAccount(t *testing.T) {
 	assert.Len(t, result.Snapshots, 1)
 }
 
+// clientErrorDetail returns the code and message the gateway renders for err:
+// a message attached at the call site, else the code's registered EC2 wording.
+func clientErrorDetail(t *testing.T, err error) (string, string) {
+	t.Helper()
+	code, message, ok := awserrors.ResolveErrorDetail(err)
+	require.True(t, ok, "unregistered error: %v", err)
+	if message == "" {
+		message = awserrors.LookupErrorMessage("ec2", code).Message
+	}
+	return code, message
+}
+
 // TestDeleteSnapshot_InUseByVolume tests that deleting a snapshot fails when a volume was created from it.
 func TestDeleteSnapshot_InUseByVolume(t *testing.T) {
 	svc, store := setupTestSnapshotService(t)
@@ -480,7 +493,9 @@ func TestDeleteSnapshot_InUseByVolume(t *testing.T) {
 		SnapshotId: snap.SnapshotId,
 	}, testAccountID)
 	require.Error(t, err)
-	assert.Contains(t, err.Error(), awserrors.ErrorInvalidSnapshotInUse)
+	code, message := clientErrorDetail(t, err)
+	assert.Equal(t, awserrors.ErrorInvalidSnapshotInUse, code)
+	assert.Equal(t, "The snapshot '"+*snap.SnapshotId+"' is in use by volume 'vol-cloned'.", message)
 
 	// Verify snapshot still exists
 	result, err := svc.DescribeSnapshots(context.Background(), &ec2.DescribeSnapshotsInput{
@@ -490,6 +505,139 @@ func TestDeleteSnapshot_InUseByVolume(t *testing.T) {
 	assert.Len(t, result.Snapshots, 1)
 }
 
+// TestDeleteSnapshot_InUseByAnotherAccountsVolume tests that the refusal
+// never names a volume the caller does not own, since the dependency scan is
+// whole-cluster and an endpoint never confirms another account's IDs.
+func TestDeleteSnapshot_InUseByAnotherAccountsVolume(t *testing.T) {
+	svc, store := setupTestSnapshotService(t)
+
+	createTestVolume(t, svc, store, "vol-source", 50)
+	snap, err := svc.CreateSnapshot(context.Background(), &ec2.CreateSnapshotInput{
+		VolumeId: aws.String("vol-source"),
+	}, testAccountID)
+	require.NoError(t, err)
+
+	seedVolumeDocument(t, store, ebsmetadata.Volume{
+		VolumeID: "vol-foreign", TenantID: otherAccountID, CapacityGiB: 50, SnapshotID: *snap.SnapshotId,
+	})
+
+	_, err = svc.DeleteSnapshot(context.Background(), &ec2.DeleteSnapshotInput{
+		SnapshotId: snap.SnapshotId,
+	}, testAccountID)
+	require.Error(t, err)
+	code, message := clientErrorDetail(t, err)
+	assert.Equal(t, awserrors.ErrorInvalidSnapshotInUse, code)
+	assert.NotContains(t, message, "vol-foreign")
+	assert.Equal(t, "The snapshot '"+*snap.SnapshotId+"' is in use by a volume.", message)
+
+	// A volume the caller owns is named even when a foreign one also matches.
+	seedVolumeDocument(t, store, ebsmetadata.Volume{
+		VolumeID: "vol-own", CapacityGiB: 50, SnapshotID: *snap.SnapshotId,
+	})
+	_, err = svc.DeleteSnapshot(context.Background(), &ec2.DeleteSnapshotInput{
+		SnapshotId: snap.SnapshotId,
+	}, testAccountID)
+	_, message = clientErrorDetail(t, err)
+	assert.Equal(t, "The snapshot '"+*snap.SnapshotId+"' is in use by volume 'vol-own'.", message)
+}
+
+// TestDeleteSnapshot_InUseByRegisteredAMI tests that deleting a snapshot
+// fails while an AMI's own CreateImage snapshot still names it, matching AWS.
+func TestDeleteSnapshot_InUseByRegisteredAMI(t *testing.T) {
+	svc, store := setupTestSnapshotService(t)
+
+	createTestVolume(t, svc, store, "vol-source", 50)
+	snap, err := svc.CreateSnapshot(context.Background(), &ec2.CreateSnapshotInput{
+		VolumeId: aws.String("vol-source"),
+	}, testAccountID)
+	require.NoError(t, err)
+
+	require.NoError(t, svc.metadata.PutAMI(context.Background(), ebsmetadata.AMI{
+		ImageID: "ami-registered", Name: "ct-ami", SnapshotID: *snap.SnapshotId,
+	}))
+
+	_, err = svc.DeleteSnapshot(context.Background(), &ec2.DeleteSnapshotInput{
+		SnapshotId: snap.SnapshotId,
+	}, testAccountID)
+	require.Error(t, err)
+	code, message := clientErrorDetail(t, err)
+	assert.Equal(t, awserrors.ErrorInvalidSnapshotInUse, code)
+	assert.Equal(t, "The snapshot that you are trying to delete is in use by one or more AMIs.", message)
+
+	// The provider and control-plane document both survive the refusal.
+	result, err := svc.DescribeSnapshots(context.Background(), &ec2.DescribeSnapshotsInput{
+		SnapshotIds: []*string{snap.SnapshotId},
+	}, testAccountID)
+	require.NoError(t, err)
+	assert.Len(t, result.Snapshots, 1)
+
+	require.NoError(t, svc.metadata.DeleteAMI(context.Background(), "ami-registered"))
+	_, err = svc.DeleteSnapshot(context.Background(), &ec2.DeleteSnapshotInput{
+		SnapshotId: snap.SnapshotId,
+	}, testAccountID)
+	require.NoError(t, err, "once the AMI is gone, the delete succeeds")
+}
+
+// TestDeleteSnapshot_InUseByRegisteredAMI_AnotherAccount tests that the AMI
+// check is whole-cluster: a system AMI registered by a different account than
+// the snapshot's owner still blocks the delete.
+func TestDeleteSnapshot_InUseByRegisteredAMI_AnotherAccount(t *testing.T) {
+	svc, store := setupTestSnapshotService(t)
+
+	createTestVolume(t, svc, store, "vol-source", 50)
+	snap, err := svc.CreateSnapshot(context.Background(), &ec2.CreateSnapshotInput{
+		VolumeId: aws.String("vol-source"),
+	}, testAccountID)
+	require.NoError(t, err)
+
+	require.NoError(t, svc.metadata.PutAMI(context.Background(), ebsmetadata.AMI{
+		ImageID: "ami-other-account", Name: "ct-system-ami",
+		ImageOwnerAlias: otherAccountID, SnapshotID: *snap.SnapshotId,
+	}))
+
+	_, err = svc.DeleteSnapshot(context.Background(), &ec2.DeleteSnapshotInput{
+		SnapshotId: snap.SnapshotId,
+	}, testAccountID)
+	require.Error(t, err)
+	assert.Contains(t, err.Error(), awserrors.ErrorInvalidSnapshotInUse)
+}
+
+// TestDeleteSnapshot_AMIReadFailureFailsClosed tests that an AMI document
+// this cannot decode aborts the delete rather than being treated as absent:
+// a partial AMI listing is not evidence that no AMI names the snapshot.
+func TestDeleteSnapshot_AMIReadFailureFailsClosed(t *testing.T) {
+	svc, store := setupTestSnapshotService(t)
+
+	createTestVolume(t, svc, store, "vol-source", 50)
+	snap, err := svc.CreateSnapshot(context.Background(), &ec2.CreateSnapshotInput{
+		VolumeId: aws.String("vol-source"),
+	}, testAccountID)
+	require.NoError(t, err)
+
+	key, err := ebsmetadata.AMIKey("ami-corrupt")
+	require.NoError(t, err)
+	_, err = store.PutObject(context.Background(), &s3.PutObjectInput{
+		Bucket: aws.String("test-bucket"),
+		Key:    aws.String(key),
+		Body:   strings.NewReader("not-json"),
+	})
+	require.NoError(t, err)
+
+	_, err = svc.DeleteSnapshot(context.Background(), &ec2.DeleteSnapshotInput{
+		SnapshotId: snap.SnapshotId,
+	}, testAccountID)
+	require.Error(t, err)
+	assert.Contains(t, err.Error(), awserrors.ErrorServerInternal)
+
+	// Nothing was deleted: the snapshot survives in both the provider and the
+	// control-plane document.
+	result, err := svc.DescribeSnapshots(context.Background(), &ec2.DescribeSnapshotsInput{
+		SnapshotIds: []*string{snap.SnapshotId},
+	}, testAccountID)
+	require.NoError(t, err)
+	assert.Len(t, result.Snapshots, 1)
+}
+
 // TestDeleteSnapshot_NotFound tests deleting a non-existent snapshot.
 func TestDeleteSnapshot_NotFound(t *testing.T) {
 	svc, _ := setupTestSnapshotService(t)
@@ -1074,7 +1222,7 @@ func TestDeleteSnapshot_Provider_BlockedByCloneRecordedInMetadata(t *testing.T)
 	}))
 
 	_, err = svc.DeleteSnapshot(ctx, &ec2.DeleteSnapshotInput{SnapshotId: aws.String(snapshotID)}, testAccountID)
-	assert.EqualError(t, err, awserrors.ErrorInvalidSnapshotInUse)
+	assert.True(t, awserrors.IsErrorCode(err, awserrors.ErrorInvalidSnapshotInUse), "got %v", err)
 
 	// Once the clone is gone the snapshot is deletable again.
 	require.NoError(t, svc.metadata.DeleteVolume(ctx, testAccountID, "vol-clone"))

spinifex/domains/ec2/snapshot/snapshot_partition_test.go

@@ -270,7 +270,7 @@ func TestDeleteSnapshot_BlockedByACloneInAnotherAccount(t *testing.T) {
 
 	_, err = svc.DeleteSnapshot(ctx, &ec2.DeleteSnapshotInput{SnapshotId: snap.SnapshotId}, awsidentifiers.GlobalAccountID)
 	require.Error(t, err)
-	assert.Equal(t, awserrors.ErrorInvalidSnapshotInUse, err.Error())
+	assert.True(t, awserrors.IsErrorCode(err, awserrors.ErrorInvalidSnapshotInUse), "got %v", err)
 }
 
 // The irreversible case. The teardown reaper lists with OwnerIds set to the

spinifex/domains/ec2/spotinstance/launch_references.go

@@ -0,0 +1,59 @@
+package spotinstance
+
+import (
+	"context"
+	"errors"
+	"fmt"
+	"strings"
+
+	"github.com/aws/aws-sdk-go/aws"
+	"github.com/aws/aws-sdk-go/service/ec2"
+	"github.com/mulgadc/spinifex/spinifex/foundation/state/kvutil"
+	"github.com/nats-io/nats.go/jetstream"
+)
+
+// LaunchReferences are the key-pair and placement-group names that an
+// account's open or active spot requests would launch with. They stay inside
+// EC2; callers must not report them.
+type LaunchReferences struct {
+	KeyNames            []string
+	PlacementGroupNames []string
+}
+
+// AccountLaunchReferences reads every open or active spot request of the
+// account. It is complete or an error: an unlistable bucket or an unreadable
+// request is an error, because the unread request may name the resource.
+func (s *SpotInstanceServiceImpl) AccountLaunchReferences(ctx context.Context, accountID string) (*LaunchReferences, error) {
+	keys, err := kvutil.Keys(ctx, s.activeKV)
+	if err != nil && !errors.Is(err, jetstream.ErrNoKeysFound) {
+		return nil, fmt.Errorf("list spot requests: %w", err)
+	}
+	prefix := accountID + "."
+	out := &LaunchReferences{}
+	for _, key := range keys {
+		if key == kvutil.VersionKey || !strings.HasPrefix(key, prefix) {
+			continue
+		}
+		record, err := s.readRecord(ctx, s.activeKV, key)
+		if errors.Is(err, jetstream.ErrKeyNotFound) {
+			continue
+		}
+		if err != nil || record.Request == nil {
+			return nil, errors.New("spot request unreadable")
+		}
+		switch aws.StringValue(record.Request.State) {
+		case ec2.SpotInstanceStateClosed, ec2.SpotInstanceStateCancelled, ec2.SpotInstanceStateFailed:
+			continue
+		}
+		spec := launchSpec(record.Request)
+		if name := aws.StringValue(spec.KeyName); name != "" {
+			out.KeyNames = append(out.KeyNames, name)
+		}
+		if spec.Placement != nil {
+			if name := aws.StringValue(spec.Placement.GroupName); name != "" {
+				out.PlacementGroupNames = append(out.PlacementGroupNames, name)
+			}
+		}
+	}
+	return out, nil
+}

spinifex/domains/ec2/spotinstance/launch_references_test.go

@@ -0,0 +1,49 @@
+package spotinstance
+
+import (
+	"testing"
+
+	"github.com/aws/aws-sdk-go/aws"
+	"github.com/aws/aws-sdk-go/service/ec2"
+	"github.com/mulgadc/spinifex/spinifex/foundation/state/kvutil"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+// Open and active requests contribute their key and group names; cancelled
+// requests and other accounts' requests do not.
+func TestAccountLaunchReferences_ReadsOpenAndActiveRequests(t *testing.T) {
+	svc := setupTestService(t)
+	active := makeSIR("sir-active01", "i-0000000000000001")
+	active.LaunchSpecification.Placement = &ec2.SpotPlacement{GroupName: aws.String("pg-spread")}
+	open := makeSIR("sir-open0001", "")
+	open.State = aws.String(ec2.SpotInstanceStateOpen)
+	open.LaunchSpecification.KeyName = aws.String("open-key")
+	putSIRs(t, svc, active, open)
+	cancelled := makeSIR("sir-cancel01", "i-0000000000000002")
+	cancelled.LaunchSpecification.KeyName = aws.String("cancelled-key")
+	putSIRs(t, svc, cancelled)
+	_, err := svc.CancelSpotInstanceRequests(t.Context(), &ec2.CancelSpotInstanceRequestsInput{
+		SpotInstanceRequestIds: []*string{aws.String("sir-cancel01")},
+	}, testAccountID)
+	require.NoError(t, err)
+	_, err = svc.PutSpotInstanceRequests(t.Context(), &PutSpotRequestsInput{Requests: []*ec2.SpotInstanceRequest{makeSIR("sir-other001", "")}}, "999988887777")
+	require.NoError(t, err)
+
+	refs, err := svc.AccountLaunchReferences(t.Context(), testAccountID)
+	require.NoError(t, err)
+	assert.ElementsMatch(t, []string{"my-key", "open-key"}, refs.KeyNames)
+	assert.Equal(t, []string{"pg-spread"}, refs.PlacementGroupNames)
+}
+
+// An unreadable request may name anything, so the check fails rather than
+// answer without it.
+func TestAccountLaunchReferences_UnreadableRequestFailsTheCheck(t *testing.T) {
+	svc := setupTestService(t)
+	putSIRs(t, svc, makeSIR("sir-active01", "i-0000000000000001"))
+	_, err := svc.activeKV.Put(t.Context(), kvutil.AccountKey(testAccountID, "sir-broken01"), []byte("{not json"))
+	require.NoError(t, err)
+
+	_, err = svc.AccountLaunchReferences(t.Context(), testAccountID)
+	require.Error(t, err)
+}

spinifex/domains/ec2/volume/account_termination_current_behaviour_test.go

@@ -0,0 +1,79 @@
+package volume_test
+
+import (
+	"context"
+	"testing"
+
+	"github.com/aws/aws-sdk-go/aws"
+	"github.com/aws/aws-sdk-go/service/ec2"
+	testutil "github.com/mulgadc/spinifex/internal/testkit"
+	"github.com/mulgadc/spinifex/spinifex/bootstrap/config"
+	ebsmetadata "github.com/mulgadc/spinifex/spinifex/domains/ec2/ebs/metadata"
+	ec2snapshot "github.com/mulgadc/spinifex/spinifex/domains/ec2/snapshot"
+	ec2volume "github.com/mulgadc/spinifex/spinifex/domains/ec2/volume"
+	awserrors "github.com/mulgadc/spinifex/spinifex/foundation/aws/errors"
+	ebsprovider "github.com/mulgadc/spinifex/spinifex/providers/ebs"
+	"github.com/mulgadc/spinifex/spinifex/providers/objectstore"
+	"github.com/nats-io/nats.go/jetstream"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+// The TestCurrentBehaviour_AccountTermination_ tests pin what terminate does
+// to a DeleteOnTermination volume today; a failure means it changed.
+
+// A snapshotted DeleteOnTermination volume loses its attachment before the
+// delete refuses, so it survives as an unattached volume that the leak reaper,
+// which keys on the attachment, never marks.
+func TestCurrentBehaviour_AccountTermination_SnapshottedDeleteOnTerminationVolumeIsDetachedAndUnmarked(t *testing.T) {
+	const (
+		accountID  = "123456789012"
+		bucket     = "ct-bucket"
+		volumeID   = "vol-dot-snapshotted"
+		instanceID = "i-0123456789abcdef0"
+	)
+	ctx := context.Background()
+	_, nc, js := testutil.StartTestJetStream(t)
+	refs, err := js.CreateKeyValue(ctx, jetstream.KeyValueConfig{Bucket: ec2snapshot.KVBucketVolumeSnapshots})
+	require.NoError(t, err)
+
+	store := objectstore.NewMemoryObjectStore()
+	provider := ebsprovider.NewMemoryProvider(ebsprovider.Capabilities{})
+	cfg := &config.Config{WalDir: t.TempDir(), Predastore: config.PredastoreConfig{Bucket: bucket}}
+	volumes := ec2volume.NewVolumeServiceImplWithStore(cfg, store, nc, refs)
+	volumes.SetEBSProvider(provider)
+	snapshots := ec2snapshot.NewSnapshotServiceImplWithStore(cfg, store, nc, refs)
+	snapshots.SetEBSProvider(provider)
+	metadata := ebsmetadata.NewStore(store, bucket)
+
+	created, err := provider.CreateVolume(ctx, ebsprovider.CreateVolumeRequest{
+		Versioned: ebsprovider.NewVersioned(), VolumeID: volumeID,
+		CapacityRange: ebsprovider.CapacityRange{RequiredBytes: 1 << 30},
+	})
+	require.NoError(t, err)
+	require.NoError(t, metadata.PutVolume(ctx, ebsmetadata.Volume{
+		VolumeID: volumeID, TenantID: accountID, CapacityGiB: 1,
+		State: "available", ProviderHandle: created.Handle, DeleteOnTermination: true,
+	}))
+	_, err = snapshots.CreateSnapshot(ctx, &ec2.CreateSnapshotInput{VolumeId: aws.String(volumeID)}, accountID)
+	require.NoError(t, err)
+	require.NoError(t, volumes.UpdateVolumeState(accountID, volumeID, "in-use", instanceID, "/dev/sda1"))
+
+	err = volumes.DeleteVolumeOnTerminate(ctx, volumeID, accountID)
+	assert.EqualError(t, err, awserrors.ErrorVolumeInUse, "the snapshot index refuses the terminate delete")
+
+	survivor, err := metadata.GetVolume(ctx, accountID, volumeID)
+	require.NoError(t, err, "the volume survives its instance")
+	assert.Empty(t, survivor.AttachedInstance, "the attachment was cleared before the refusal")
+	assert.Equal(t, "available", survivor.State)
+
+	reaper := volumes.NewVolumeLeakReaper(func() (map[string]bool, error) {
+		return map[string]bool{instanceID: true}, nil
+	})
+	marked, err := reaper.Sweep(ctx)
+	require.NoError(t, err)
+	assert.Zero(t, marked, "with no attachment left, the leak reaper cannot attribute the volume to its instance")
+
+	_, err = volumes.DeleteVolume(ctx, &ec2.DeleteVolumeInput{VolumeId: aws.String(volumeID)}, accountID)
+	assert.EqualError(t, err, awserrors.ErrorVolumeInUse, "an ordinary delete is refused for the same reason")
+}

spinifex/domains/ec2/volume/delete_volume_unreadable_test.go

@@ -0,0 +1,211 @@
+package volume
+
+import (
+	"context"
+	"errors"
+	"testing"
+
+	"github.com/aws/aws-sdk-go/aws"
+	"github.com/aws/aws-sdk-go/service/ec2"
+	"github.com/aws/aws-sdk-go/service/s3"
+	"github.com/mulgadc/spinifex/spinifex/domains/ec2/ebs/metadata"
+	awserrors "github.com/mulgadc/spinifex/spinifex/foundation/aws/errors"
+	ebsprovider "github.com/mulgadc/spinifex/spinifex/providers/ebs"
+	"github.com/mulgadc/spinifex/spinifex/providers/objectstore"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+// getFailingStore fails the read of one key with an error that is not a
+// missing key, standing in for a document that is present but unreadable
+// (e.g. a storage-layer reconstruction failure).
+type getFailingStore struct {
+	*objectstore.MemoryObjectStore
+
+	failKey string
+}
+
+func (s *getFailingStore) GetObject(ctx context.Context, input *s3.GetObjectInput) (*s3.GetObjectOutput, error) {
+	if aws.StringValue(input.Key) == s.failKey {
+		return nil, errors.New("reconstruction failed: shard sizes do not match")
+	}
+	return s.MemoryObjectStore.GetObject(ctx, input)
+}
+
+// countingFailAfterStore fails every read of one key from its (succeedCalls+1)th
+// call onward, deterministically reproducing a document that answers one read
+// and then goes unreadable on the next, without timing.
+type countingFailAfterStore struct {
+	*objectstore.MemoryObjectStore
+
+	failKey      string
+	succeedCalls int
+	calls        int
+}
+
+func (s *countingFailAfterStore) GetObject(ctx context.Context, input *s3.GetObjectInput) (*s3.GetObjectOutput, error) {
+	if aws.StringValue(input.Key) == s.failKey {
+		s.calls++
+		if s.calls > s.succeedCalls {
+			return nil, errors.New("reconstruction failed: shard sizes do not match")
+		}
+	}
+	return s.MemoryObjectStore.GetObject(ctx, input)
+}
+
+// TestDeleteVolume_UnreadableDocumentIsInternalError pins the fix: a volume
+// document that fails to read for a reason other than absence must answer
+// InternalError, never InvalidVolume.NotFound. Reporting NotFound here is
+// what let terminate cleanup and the legacy reaper count a volume whose
+// state is unknown as already deleted.
+func TestDeleteVolume_UnreadableDocumentIsInternalError(t *testing.T) {
+	memory := objectstore.NewMemoryObjectStore()
+	svc := newTestVolumeServiceWithStore("ap-southeast-2a", memory)
+
+	volumeID := "vol-unreadable0000"
+	createVolumeInStoreWithMeta(t, svc, memory, volumeID, ebsmetadata.Volume{
+		VolumeID: volumeID, CapacityGiB: 10, State: "available",
+	})
+
+	failing := &getFailingStore{MemoryObjectStore: memory, failKey: metadataKey(t, volumeID)}
+	svc.store = failing
+	svc.metadata = ebsmetadata.NewStore(failing, "test-bucket")
+
+	_, err := svc.DeleteVolume(context.Background(), &ec2.DeleteVolumeInput{
+		VolumeId: aws.String(volumeID),
+	}, testVolAccountID)
+
+	require.Error(t, err)
+	assert.Equal(t, awserrors.ErrorServerInternal, awserrors.ValidErrorCodeFromError(err),
+		"an unreadable document must answer InternalError, not NotFound")
+	assert.NotContains(t, err.Error(), awserrors.ErrorInvalidVolumeNotFound)
+
+	// Nothing was deleted: the document is still there (read through the
+	// underlying store, bypassing the fault), and so is the provider volume.
+	_, getErr := ebsmetadata.NewStore(memory, "test-bucket").GetVolume(context.Background(), testVolAccountID, volumeID)
+	require.NoError(t, getErr, "the control-plane document must survive a failed delete")
+
+	provVol, provErr := svc.provider.GetVolume(context.Background(), ebsprovider.GetVolumeRequest{
+		Versioned: ebsprovider.NewVersioned(), VolumeID: volumeID,
+	})
+	require.NoError(t, provErr)
+	require.NotNil(t, provVol, "the provider's data must survive a failed delete")
+}
+
+// TestDeleteVolume_AbsentDocumentIsStillNotFound guards the other side of the
+// fix: absence (NoSuchKey) must keep returning InvalidVolume.NotFound, not be
+// swept up into the new InternalError branch.
+func TestDeleteVolume_AbsentDocumentIsStillNotFound(t *testing.T) {
+	svc := newTestVolumeService("ap-southeast-2a")
+
+	_, err := svc.DeleteVolume(context.Background(), &ec2.DeleteVolumeInput{
+		VolumeId: aws.String("vol-absent00000001"),
+	}, testVolAccountID)
+
+	require.Error(t, err)
+	assert.Equal(t, awserrors.ErrorInvalidVolumeNotFound, err.Error())
+}
+
+// TestDeleteVolume_CrossAccountDocumentIsStillNotFound guards the cross-tenant
+// branch: a document that exists but belongs to another account must keep
+// returning InvalidVolume.NotFound, exactly as before the fix.
+func TestDeleteVolume_CrossAccountDocumentIsStillNotFound(t *testing.T) {
+	store := objectstore.NewMemoryObjectStore()
+	svc := newTestVolumeServiceWithStore("ap-southeast-2a", store)
+
+	createVolumeInStoreWithMeta(t, svc, store, "vol-othertenant000", ebsmetadata.Volume{
+		VolumeID: "vol-othertenant000", CapacityGiB: 10, State: "available", TenantID: "222222222222",
+	})
+
+	_, err := svc.DeleteVolume(context.Background(), &ec2.DeleteVolumeInput{
+		VolumeId: aws.String("vol-othertenant000"),
+	}, testVolAccountID)
+
+	require.Error(t, err)
+	assert.Equal(t, awserrors.ErrorInvalidVolumeNotFound, err.Error())
+}
+
+// TestDeleteVolumeOnTerminate_DeleteReadFailureIsNotSuccess reproduces the
+// production shape from mulga-salhr: the attachment-clear read inside
+// UpdateVolumeState succeeds, but the document goes unreadable before
+// DeleteVolume's own read of the same key. That read failure must come back
+// as a non-NotFound error, which is exactly the signal
+// daemon/vm_adapters.go's DeleteVolumes and instance/service_impl.go's
+// deleteInstanceVolumes use to tell a real failure from an already-gone
+// volume (!awserrors.IsNotFound(err) && !objectstore.IsNoSuchKeyError(err)).
+func TestDeleteVolumeOnTerminate_DeleteReadFailureIsNotSuccess(t *testing.T) {
+	memory := objectstore.NewMemoryObjectStore()
+	svc := newTestVolumeServiceWithStore("ap-southeast-2a", memory)
+
+	volumeID := "vol-stopped-root2"
+	createVolumeInStoreWithMeta(t, svc, memory, volumeID, ebsmetadata.Volume{
+		VolumeID:         volumeID,
+		CapacityGiB:      10,
+		State:            "available",
+		AttachedInstance: "i-stopped",
+	})
+
+	// The first read of the document (UpdateVolumeState's attachment clear)
+	// succeeds; every read from the second onward (DeleteVolume's own read)
+	// fails, deterministically, with no reliance on timing.
+	failing := &countingFailAfterStore{MemoryObjectStore: memory, failKey: metadataKey(t, volumeID), succeedCalls: 1}
+	svc.store = failing
+	svc.metadata = ebsmetadata.NewStore(failing, "test-bucket")
+
+	err := svc.DeleteVolumeOnTerminate(context.Background(), volumeID, testVolAccountID)
+
+	require.Error(t, err, "a delete-read failure after a successful attachment clear must not be reported as success")
+	assert.False(t, awserrors.IsNotFound(err),
+		"terminate cleanup treats IsNotFound as already-gone; this must not qualify")
+	assert.False(t, objectstore.IsNoSuchKeyError(err),
+		"terminate cleanup treats IsNoSuchKeyError as already-gone; this must not qualify")
+
+	// The attachment was cleared (terminate implies detach), but the document
+	// and the provider volume must both still exist: nothing was deleted.
+	_, getErr := ebsmetadata.NewStore(memory, "test-bucket").GetVolume(context.Background(), testVolAccountID, volumeID)
+	require.NoError(t, getErr, "the control-plane document must survive a failed delete")
+
+	provVol, provErr := svc.provider.GetVolume(context.Background(), ebsprovider.GetVolumeRequest{
+		Versioned: ebsprovider.NewVersioned(), VolumeID: volumeID,
+	})
+	require.NoError(t, provErr)
+	require.NotNil(t, provVol, "the provider's data must survive a failed delete")
+}
+
+// TestForceDetachVolume_UnreadableDocumentIsInternalError: an attached volume
+// whose document is present but unreadable must answer InternalError and keep
+// its attachment; absence and another account's volume stay NotFound.
+func TestForceDetachVolume_UnreadableDocumentIsInternalError(t *testing.T) {
+	memory := objectstore.NewMemoryObjectStore()
+	svc := newTestVolumeServiceWithStore("ap-southeast-2a", memory)
+
+	volumeID := "vol-forcedetach000"
+	createVolumeInStoreWithMeta(t, svc, memory, volumeID, ebsmetadata.Volume{
+		VolumeID: volumeID, CapacityGiB: 10, State: "in-use", AttachedInstance: "i-0123456789abcdef0",
+	})
+
+	failing := &getFailingStore{MemoryObjectStore: memory, failKey: metadataKey(t, volumeID)}
+	svc.store = failing
+	svc.metadata = ebsmetadata.NewStore(failing, "test-bucket")
+
+	_, err := svc.ForceDetachVolume(context.Background(), &ec2.DetachVolumeInput{
+		VolumeId: aws.String(volumeID),
+	}, testVolAccountID)
+	require.Error(t, err)
+	assert.Equal(t, awserrors.ErrorServerInternal, awserrors.ValidErrorCodeFromError(err))
+
+	meta, getErr := ebsmetadata.NewStore(memory, "test-bucket").GetVolume(context.Background(), testVolAccountID, volumeID)
+	require.NoError(t, getErr)
+	assert.Equal(t, "in-use", meta.State, "a failed force-detach must not clear the attachment")
+	assert.Equal(t, "i-0123456789abcdef0", meta.AttachedInstance)
+
+	_, err = svc.ForceDetachVolume(context.Background(), &ec2.DetachVolumeInput{
+		VolumeId: aws.String("vol-doesnotexist00"),
+	}, testVolAccountID)
+	assert.Equal(t, awserrors.ErrorInvalidVolumeNotFound, awserrors.ValidErrorCodeFromError(err))
+
+	_, err = svc.ForceDetachVolume(context.Background(), &ec2.DetachVolumeInput{
+		VolumeId: aws.String(volumeID),
+	}, "999999999999")
+	assert.Equal(t, awserrors.ErrorInvalidVolumeNotFound, awserrors.ValidErrorCodeFromError(err))
+}

spinifex/domains/ec2/volume/service_impl.go

@@ -289,6 +289,7 @@ var describeVolumesValidFilters = map[string]bool{
 	"attachment.status":      true,
 	"attachment.device":      true,
 	"availability-zone":      true,
+	"snapshot-id":            true,
 }
 
 // DescribeVolumes lists EBS volumes by reading config.json files from S3.
@@ -401,6 +402,10 @@ func volumeMatchesFilters(vol *ec2.Volume, filters map[string][]string) bool {
 			if vol.AvailabilityZone != nil {
 				field = *vol.AvailabilityZone
 			}
+		case "snapshot-id":
+			if vol.SnapshotId != nil {
+				field = *vol.SnapshotId
+			}
 		default:
 			return false
 		}
@@ -1056,7 +1061,10 @@ func (s *VolumeServiceImpl) ForceDetachVolume(ctx context.Context, input *ec2.De
 
 	meta, err := s.metadata.GetVolume(ctx, accountID, volumeID)
 	if err != nil {
-		return nil, errors.New(awserrors.ErrorInvalidVolumeNotFound)
+		if objectstore.IsNoSuchKeyError(err) {
+			return nil, errors.New(awserrors.ErrorInvalidVolumeNotFound)
+		}
+		return nil, awserrors.Errorf(awserrors.ErrorServerInternal, "get volume metadata %s: %v", volumeID, err)
 	}
 	if meta.TenantID != accountID {
 		return nil, errors.New(awserrors.ErrorInvalidVolumeNotFound)
@@ -1091,10 +1099,15 @@ func (s *VolumeServiceImpl) DeleteVolume(ctx context.Context, input *ec2.DeleteV
 		return nil, err
 	}
 
-	// AWS-faithful: an absent volume returns InvalidVolume.NotFound.
+	// AWS-faithful: an absent volume returns InvalidVolume.NotFound. Any other
+	// read failure means the document's state is unknown, so it must not be
+	// reported or treated as already deleted.
 	meta, err := s.metadata.GetVolume(ctx, accountID, volumeID)
 	if err != nil {
-		return nil, errors.New(awserrors.ErrorInvalidVolumeNotFound)
+		if objectstore.IsNoSuchKeyError(err) {
+			return nil, errors.New(awserrors.ErrorInvalidVolumeNotFound)
+		}
+		return nil, awserrors.Errorf(awserrors.ErrorServerInternal, "get volume metadata %s: %v", volumeID, err)
 	}
 	if meta.TenantID != accountID {
 		return nil, errors.New(awserrors.ErrorInvalidVolumeNotFound)

spinifex/domains/ec2/volume/service_impl_test.go

@@ -2081,6 +2081,40 @@ func TestDescribeVolumes_FilterByAZ(t *testing.T) {
 	assert.Equal(t, "vol-az1", *out.Volumes[0].VolumeId)
 }
 
+func TestDescribeVolumes_FilterBySnapshotId(t *testing.T) {
+	store := objectstore.NewMemoryObjectStore()
+	svc := newTestVolumeServiceWithStore("ap-southeast-2a", store)
+
+	createVolumeInStoreWithMeta(t, svc, store, "vol-fromsnap", ebsmetadata.Volume{
+		VolumeID: "vol-fromsnap", CapacityGiB: 10, State: "available",
+		SnapshotID: "snap-src", TenantID: "000000000001",
+	})
+	createVolumeInStoreWithMeta(t, svc, store, "vol-othersnap", ebsmetadata.Volume{
+		VolumeID: "vol-othersnap", CapacityGiB: 10, State: "available",
+		SnapshotID: "snap-other", TenantID: "000000000001",
+	})
+	createVolumeInStoreWithMeta(t, svc, store, "vol-blank", ebsmetadata.Volume{
+		VolumeID: "vol-blank", CapacityGiB: 10, State: "available", TenantID: "000000000001",
+	})
+
+	describe := func(values ...string) []string {
+		t.Helper()
+		out, err := svc.DescribeVolumes(context.Background(), &ec2.DescribeVolumesInput{
+			Filters: []*ec2.Filter{{Name: aws.String("snapshot-id"), Values: aws.StringSlice(values)}},
+		}, "000000000001")
+		require.NoError(t, err)
+		ids := make([]string, 0, len(out.Volumes))
+		for _, v := range out.Volumes {
+			ids = append(ids, *v.VolumeId)
+		}
+		return ids
+	}
+
+	assert.Equal(t, []string{"vol-fromsnap"}, describe("snap-src"))
+	assert.ElementsMatch(t, []string{"vol-fromsnap", "vol-othersnap"}, describe("snap-src", "snap-other"))
+	assert.Empty(t, describe("snap-unknown"))
+}
+
 func TestDescribeVolumes_FilterMultipleValues_OR(t *testing.T) {
 	store := objectstore.NewMemoryObjectStore()
 	svc := newTestVolumeServiceWithStore("ap-southeast-2a", store)

spinifex/domains/network/reconcile/lock.go

@@ -41,12 +41,12 @@ func AcquireLeader(ctx context.Context, nc *nats.Conn, bucket, holder string) (f
 	})
 	if err != nil {
 		slog.ErrorContext(ctx, "reconcile/lock: lease config invalid",
-			"holder", holder, "bucket", bucket, "err", err)
+			"node", holder, "bucket", bucket, "err", err)
 		return nil, false
 	}
 	if !lease.TryAcquire(ctx) {
 		slog.InfoContext(ctx, "reconcile/lock: not leader, skipping reconcile",
-			"holder", holder, "bucket", bucket)
+			"node", holder, "bucket", bucket)
 		return nil, false
 	}
 	return func() { lease.Release(ctx) }, true

spinifex/domains/network/reconcile/lock_test.go

@@ -1,7 +1,10 @@
 package reconcile
 
 import (
+	"bytes"
 	"context"
+	"encoding/json"
+	"log/slog"
 	"testing"
 	"time"
 
@@ -103,3 +106,34 @@ func TestAcquireLeader_ReleaseDoesNotDeleteSuccessorKey(t *testing.T) {
 	require.NoError(t, err, "stale release deleted the successor's leader key")
 	assert.Equal(t, "node-2", string(entry.Value()), "successor's key must survive a stale release")
 }
+
+// A node that loses the election logs itself as node: "holder=node-2" on
+// node-2's log reads as though node-2 holds the lock.
+func TestAcquireLeader_NotLeaderLogNamesLocalNode(t *testing.T) {
+	_, nc, _ := testutil.StartTestJetStream(t)
+
+	release, ok := AcquireLeader(t.Context(), nc, KVBucketVPCDReconcile, "node-1")
+	require.True(t, ok)
+	defer release()
+
+	var buf bytes.Buffer
+	prev := slog.Default()
+	slog.SetDefault(slog.New(slog.NewJSONHandler(&buf, &slog.HandlerOptions{Level: slog.LevelInfo})))
+	t.Cleanup(func() { slog.SetDefault(prev) })
+
+	_, ok = AcquireLeader(t.Context(), nc, KVBucketVPCDReconcile, "node-2")
+	require.False(t, ok)
+
+	var line map[string]any
+	dec := json.NewDecoder(&buf)
+	for dec.More() {
+		var rec map[string]any
+		require.NoError(t, dec.Decode(&rec))
+		if rec["msg"] == "reconcile/lock: not leader, skipping reconcile" {
+			line = rec
+		}
+	}
+	require.NotNil(t, line, "a losing acquire must log that it is skipping")
+	assert.Equal(t, "node-2", line["node"])
+	assert.NotEqual(t, "node-2", line["holder"], "the local node is not the holder")
+}

spinifex/foundation/aws/errors/awserrors.go

@@ -658,13 +658,22 @@ func HasDependencies(kind, id string) error {
 // retryableError wraps a registered code with a suggested Retry-After
 // duration, for a 503 that a client can expect to clear on its own within a
 // bounded window (e.g. a warm-up race) rather than an open-ended outage.
-// Error() returns only the code, so code resolution is unaffected.
+// Error() returns only the code, or the wrapped error's text, so code
+// resolution is unaffected.
 type retryableError struct {
 	code       string
+	err        error
 	retryAfter time.Duration
 }
 
-func (e *retryableError) Error() string { return e.code }
+func (e *retryableError) Error() string {
+	if e.err != nil {
+		return e.err.Error()
+	}
+	return e.code
+}
+
+func (e *retryableError) Unwrap() error { return e.err }
 
 // RetryAfter wraps code (a registered ErrorLookup key) with a suggested
 // Retry-After duration for ErrorHandler to surface as a response header.
@@ -672,6 +681,13 @@ func RetryAfter(code string, d time.Duration) error {
 	return &retryableError{code: code, retryAfter: d}
 }
 
+// WithRetryAfter attaches a suggested Retry-After duration to err without
+// changing its text, unlike joining err with RetryAfter, which appends the
+// bare code on a second line.
+func WithRetryAfter(err error, d time.Duration) error {
+	return &retryableError{err: err, retryAfter: d}
+}
+
 // ResolveRetryAfter returns the Retry-After duration attached to err via
 // RetryAfter, if any, walking the same unwrap tree ResolveErrorDetail does.
 func ResolveRetryAfter(err error) (time.Duration, bool) {