test(lambda): sample heap only across the measured lifetimes and join the sampler

The sampler now starts after the post-GC baseline and stops after both
replies are read, and cleanup joins it if the test fails first.

.github/CODEOWNERS

@@ -1,15 +0,0 @@
-# CODEOWNERS — GitHub PR review enforcement
-#
-# Format: <path pattern>  <owner> [<owner> …]
-# Order matters — last matching pattern wins.
-#
-# Set the matching branch protection rule on `dev` and `main` to
-# "Require review from Code Owners" or this file is advisory only.
-
-# Network layer is governed by its invariant tests.
-# Changes here must respect those clauses
-# see spinifex/domains/network/invariants/
-# mulga docs/development/proposals/region-local-architecture/0006-spinifex-network-layer-contract.md (proposal, not accepted)
-/spinifex/domains/network/          @joshsiv-mulga
-/spinifex/runtime/roles/vpcd/ @joshsiv-mulga
-/spinifex/domains/network/invariants/ @joshsiv-mulga

.github/pull_request_template.md

@@ -6,20 +6,6 @@
 
 <!-- The notable changes, as bullets. -->
 
-## Account-scoped lifecycle impact
-
-<!-- Complete this section only when the PR changes account-scoped resource lifecycle, account termination, resource cleanup, or a lifecycle boundary. Otherwise write: N/A. -->
-
-- Lifecycle owner and resource family:
-- Owner-local ordering, recovery, retained-state, or force impact:
-- Cross-domain dependency or ownership impact:
-- Evidence added or rechecked:
-- ADR-0008 lifecycle register / ADR-0001 migration-record / bead updates:
-
-## Mulga design
-
-<!-- One line, so a reviewer knows whether the design records in the private mulga repository need to change. Write either "Updated: <mulga PR or bead id>" or "Not needed: <reason>". This is a reminder, not a gate. -->
-
 ## Community Contribution License
 
 - [ ] All community contributions in this pull request are licensed to Mulga under the terms of the [Apache License 2.0](https://www.apache.org/licenses/LICENSE-2.0). By creating this pull request I represent that I have the right to license these contributions to Mulga under the Apache License 2.0.

.github/scripts/e2e-select.sh

@@ -0,0 +1,148 @@
+#!/usr/bin/env bash
+# e2e-select.sh push <before> <sha> | suites "<e2e-x ...>"
+# Prints the e2e gate's single_legs, multi_suites and multi_timeout outputs.
+# Exits non-zero on anything it cannot read, which fails the e2e status.
+set -euo pipefail
+
+SCRIPTS_DIR=$(CDPATH='' cd -- "$(dirname -- "$0")" && pwd)
+# shellcheck source=suite-sets.sh
+. "${SCRIPTS_DIR}/suite-sets.sh"
+
+# Provision, install and teardown around the suites; the floor is what the
+# jobs have always had, so a budget can only be raised from it.
+SETUP_ALLOWANCE_MIN=15
+FLOOR_MIN=60
+
+BASE_RUNNER=ci-single
+SERVICE_RUNNER=ci-pseudo
+
+log() { printf '[e2e-select] %s\n' "$*" >&2; }
+die() { log "error: $*"; exit 1; }
+
+# in_list <word> <list>
+in_list() {
+  local w
+  for w in $2; do [ "$w" = "$1" ] && return 0; done
+  return 1
+}
+
+# keep <list> <eligible>: the words of <list> that are in <eligible>, deduplicated.
+keep() {
+  local out="" s
+  for s in $1; do
+    in_list "$s" "$2" || continue
+    in_list "$s" "$out" || out="${out:+$out }$s"
+  done
+  printf '%s' "$out"
+}
+
+# budget <list>: max over the suites' own timeouts plus the setup allowance.
+# A sum would demand hours for a set that shares one job and never takes them.
+budget() {
+  local want="${FLOOR_MIN}" s t
+  for s in $1; do
+    t="$(e2e_suite_timeout "$s")"
+    t=$((${t%m} + SETUP_ALLOWANCE_MIN))
+    [ "$t" -gt "$want" ] && want="$t"
+  done
+  printf '%s' "$want"
+}
+
+prefixed() {
+  local out="" s
+  for s in $1; do out="${out:+$out }e2e-$s"; done
+  printf '%s' "$out"
+}
+
+# leg <name> <runner> <suites>: one matrix entry, or nothing for no suites.
+leg() {
+  [ -n "$3" ] || return 0
+  jq -nc --arg leg "$1" --arg runner "$2" --arg suites "$(prefixed "$3")" \
+    --argjson timeout "$(budget "$3")" \
+    '{leg: $leg, runner: $runner, suites: $suites, timeout: $timeout}'
+}
+
+# emit <base-single> <service-single> <multi>
+emit() {
+  local legs
+  legs=$( { leg base "${BASE_RUNNER}" "$1"; leg service "${SERVICE_RUNNER}" "$2"; } | jq -sc .)
+  log "base leg: ${1:-none}"
+  log "service leg: ${2:-none}"
+  log "multi: ${3:-none}"
+  printf 'single_legs=%s\n' "$legs"
+  printf 'multi_suites=%s\n' "$(prefixed "$3")"
+  printf 'multi_timeout=%s\n' "$(budget "$3")"
+}
+
+# select_for_paths <newline-separated paths> [all]: base set plus matches.
+# With "all", every mapped suite is selected, for a push with no baseline.
+select_for_paths() {
+  local matched="" p s
+  if [ "${2:-}" = all ]; then
+    matched=$(e2e_path_suites_all | tr '\n' ' ')
+  else
+    while IFS= read -r p; do
+      [ -n "$p" ] || continue
+      for s in $(e2e_suites_for_path "$p"); do
+        in_list "$s" "$matched" || { matched="${matched:+$matched }$s"; log "${p} -> ${s}"; }
+      done
+    done <<<"$1"
+  fi
+  # A matched suite runs where it is eligible, preferring one node, so route53
+  # goes to the service leg and lb to the multi job.
+  local extra_single="" extra_multi=""
+  for s in $matched; do
+    if in_list "$s" "${E2E_SUITES_SINGLE}"; then
+      extra_single="${extra_single:+$extra_single }$s"
+    elif in_list "$s" "${E2E_SUITES_MULTI}"; then
+      extra_multi="${extra_multi:+$extra_multi }$s"
+    else
+      die "E2E_PATH_SUITES names '${s}', which no topology can run"
+    fi
+  done
+  local base_single base_multi
+  base_single=$(keep "${E2E_SUITES_PUSH_BASE}" "${E2E_SUITES_SINGLE}")
+  base_multi=$(keep "${E2E_SUITES_PUSH_BASE}" "${E2E_SUITES_MULTI}")
+  # A matched suite already in the base leg runs there, not twice.
+  extra_single=$(for s in $extra_single; do in_list "$s" "$base_single" || printf '%s ' "$s"; done)
+  emit "$base_single" "$(keep "$extra_single" "${E2E_SUITES_SINGLE}")" \
+    "$(keep "$base_multi $extra_multi" "${E2E_SUITES_MULTI}")"
+}
+
+mode="${1:-}"
+case "$mode" in
+  push)
+    [ $# -eq 3 ] || die "usage: $0 push <before> <sha>"
+    before="$2" sha="$3"
+    git cat-file -e "${sha}^{commit}" 2>/dev/null || die "commit ${sha} is not in this checkout"
+    if [ -z "${before//0/}" ]; then
+      log "no previous commit for this push; selecting every mapped suite"
+      select_for_paths "" all
+      exit 0
+    fi
+    if ! git cat-file -e "${before}^{commit}" 2>/dev/null; then
+      git fetch -q --filter=blob:none origin "${before}" 2>/dev/null || true
+    fi
+    if ! git cat-file -e "${before}^{commit}" 2>/dev/null; then
+      log "previous commit ${before} is gone; selecting every mapped suite"
+      select_for_paths "" all
+      exit 0
+    fi
+    # A rename counts as both paths, so moving code out of a service selects it.
+    changed=$(git diff --no-renames --name-only "${before}" "${sha}") || die "git diff ${before} ${sha} failed"
+    log "$(printf '%s\n' "$changed" | grep -c . || true) paths changed ${before:0:9}..${sha:0:9}"
+    select_for_paths "$changed"
+    ;;
+  suites)
+    [ $# -eq 2 ] || die "usage: $0 suites \"<e2e-x ...>\""
+    if [ -z "${2// /}" ]; then
+      select_for_paths ""
+      exit 0
+    fi
+    requested=$(printf '%s' "$2" | sed 's/e2e-//g')
+    emit "$(keep "$requested" "${E2E_SUITES_SINGLE}")" "" "$(keep "$requested" "${E2E_SUITES_MULTI}")"
+    ;;
+  *)
+    die "usage: $0 push <before> <sha> | suites \"<list>\""
+    ;;
+esac

.github/scripts/go-build-cache.sh

@@ -0,0 +1,45 @@
+#!/usr/bin/env bash
+# Trims GOCACHE to the entries one CI run used. Go refreshes an entry's mtime
+# on use only when it is over an hour old, so "mark" backdates every restored
+# entry first and "trim" drops whatever this run neither read nor wrote.
+set -euo pipefail
+
+cache="$(go env GOCACHE)"
+mark="${RUNNER_TEMP:?}/go-build-cache.mark"
+
+# stale prints each entry this run did not touch and no touched action entry
+# names. Go skips rewriting an output it already holds, so a fresh action
+# entry can point at a backdated output.
+stale() {
+  local keep="${RUNNER_TEMP}/go-build-cache.keep"
+  find "$cache" -mindepth 2 -maxdepth 2 -name '*-a' -newer "$mark" -exec cat {} + |
+    awk '{ print "/" $3 "-d" }' >"$keep"
+  find "$cache" -mindepth 2 -maxdepth 2 ! -newer "$mark" | { grep -vFf "$keep" || true; }
+}
+
+case "${1:-}" in
+mark)
+  mkdir -p "$cache"
+  find "$cache" -mindepth 2 -maxdepth 2 -exec touch -m -d '2 hours ago' {} +
+  touch "$mark"
+  ;;
+stale | trim)
+  if [[ ! -f "$mark" ]]; then
+    echo "go-build-cache: no mark, leaving the cache untrimmed" >&2
+    exit 0
+  fi
+  if [[ "$1" == stale ]]; then
+    stale
+    exit 0
+  fi
+  before=$(du -sm "$cache" | cut -f1)
+  # An entry is a file, or a directory holding one cached executable whose
+  # mtime Go refreshes in place of the binary's, so it goes as a unit.
+  stale | xargs -r rm -rf --
+  echo "trimmed ${before}MB -> $(du -sm "$cache" | cut -f1)MB"
+  ;;
+*)
+  echo "usage: $0 mark|stale|trim" >&2
+  exit 2
+  ;;
+esac

.github/scripts/hypervisor-disk-preflight_test.sh

@@ -1,87 +0,0 @@
-#!/bin/sh
-# hypervisor-disk-preflight_test.sh — drives hypervisor-disk-preflight.sh against
-# a stub ssh that answers for each hypervisor from a canned file, so every
-# verdict is checked without a real host.
-set -eu
-
-SCRIPT_DIR=$(CDPATH='' cd -- "$(dirname -- "$0")" && pwd)
-PREFLIGHT="$SCRIPT_DIR/hypervisor-disk-preflight.sh"
-TMP=$(mktemp -d)
-trap 'rm -rf "$TMP"' EXIT
-
-fails=0
-pass() { echo "  ok   - $1"; }
-fail() { echo "  FAIL - $1"; fails=$((fails + 1)); }
-
-# The stub answers "<root-kb> <pool-kb> <path>" from $FAKE/<host>, or exits 255
-# when that host has no file, the way ssh does for a host it cannot reach.
-mkdir -p "$TMP/bin" "$TMP/hosts"
-cat > "$TMP/bin/ssh" <<'STUB'
-#!/bin/sh
-for a in "$@"; do case $a in *@*) host=${a#*@} ;; esac; done
-cat >/dev/null
-if [ -f "$FAKE/$host" ]; then cat "$FAKE/$host"; exit 0; fi
-echo "ssh: Could not resolve hostname $host" >&2
-exit 255
-STUB
-chmod +x "$TMP/bin/ssh"
-PATH="$TMP/bin:$PATH"
-FAKE="$TMP/hosts"
-export PATH FAKE
-
-cat > "$TMP/env.tfvars" <<'TFVARS'
-nodes = [
-  { name = "wattle", wan_ip = "192.0.2.1", storage_pool = "drive1-pool", base_image = "/x.qcow2" },
-  # { name = "retired", wan_ip = "192.0.2.9", storage_pool = "drive1-pool", base_image = "/x.qcow2" },
-  { name = "banksia", wan_ip = "192.0.2.2", storage_pool = "nvme1-pool", base_image = "/x.qcow2" },
-]
-TFVARS
-
-GiB=1048576
-plenty="$((50 * GiB)) $((200 * GiB)) /mnt/drive1/pool"
-
-# run <label> <want-rc> <want-text>
-run() {
-    rc=0
-    out=$(bash "$PREFLIGHT" "$TMP/env.tfvars" 2>&1) || rc=$?
-    if [ "$rc" -ne "$2" ]; then
-        fail "$1: exit $rc, want $2"
-        echo "$out" | sed 's/^/        /'
-        return
-    fi
-    case $out in
-        *"$3"*) pass "$1" ;;
-        *) fail "$1: output lacks '$3'"; echo "$out" | sed 's/^/        /' ;;
-    esac
-}
-
-echo "-- hypervisor-disk-preflight.sh"
-
-echo "$plenty" > "$FAKE/wattle"
-echo "$plenty" > "$FAKE/banksia"
-run "every host has room" 0 "banksia: root fs 51200 MiB free"
-
-echo "0 $((200 * GiB)) /mnt/drive1/pool" > "$FAKE/wattle"
-run "a full root filesystem fails fast" 1 "::error::hypervisor wattle has 0 MiB free on the filesystem holding /etc/libvirt/qemu"
-
-echo "$plenty" > "$FAKE/wattle"
-echo "$((50 * GiB)) $((3 * GiB)) /mnt/nvme1/pool" > "$FAKE/banksia"
-run "a nearly full pool fails fast" 1 "::error::hypervisor banksia has 3 GiB free in storage pool nvme1-pool"
-
-echo "$plenty" > "$FAKE/banksia"
-rm -f "$FAKE/wattle"
-run "an unreachable host is reported, not failed" 0 "::warning::hypervisor disk preflight: cannot reach wattle"
-
-echo "$plenty" > "$FAKE/wattle"
-echo "$((50 * GiB)) - -" > "$FAKE/banksia"
-run "a pool libvirt does not know is reported, not failed" 0 "pool nvme1-pool (-) not found"
-
-echo "0 0 /x" > "$FAKE/retired"
-echo "$plenty" > "$FAKE/banksia"
-run "a commented-out node is never checked" 0 "wattle: root fs"
-
-if [ "$fails" -ne 0 ]; then
-    echo "$fails failure(s)"
-    exit 1
-fi
-echo "  hypervisor-disk-preflight ok"

.github/scripts/mulga_design_check.py

@@ -1,36 +0,0 @@
-#!/usr/bin/env python3
-"""Warn when a PR body has no filled "Mulga design" line. Reads PR_BODY; never fails the job."""
-
-import os
-import re
-
-COMMENT_RE = re.compile(r"<!--.*?-->", re.S)
-SECTION_RE = re.compile(r"^#{1,6}\s*Mulga design\s*$", re.I)
-LINE_RE = re.compile(r"^\s*(Updated|Not needed)\s*:\s*(.*?)\s*$", re.I)
-PLACEHOLDER_RE = re.compile(r"<[^<>]*>")
-
-
-def filled(body):
-    inside = False
-    for line in COMMENT_RE.sub("", body or "").splitlines():
-        if SECTION_RE.match(line):
-            inside = True
-        elif inside and line.lstrip().startswith("#"):
-            return False
-        elif inside:
-            m = LINE_RE.match(line)
-            if m and PLACEHOLDER_RE.sub("", m.group(2)).strip():
-                return True
-    return False
-
-
-def main():
-    if filled(os.environ.get("PR_BODY", "")):
-        print("The Mulga design line is filled in.")
-    else:
-        print('::warning title=Mulga design::Add a line under "## Mulga design": "Updated: <mulga PR or bead id>" or "Not needed: <reason>".')
-    return 0
-
-
-if __name__ == "__main__":
-    raise SystemExit(main())

.github/scripts/suite-sets.sh

@@ -66,6 +66,67 @@ E2E_SUITES_MULTI="multinode lb cert quota route53 kvquorum lbrecovery instancere
 E2E_SUITES_SINGLE_RE="$(printf '%s' "${E2E_SUITES_SINGLE}" | tr ' ' '|')"
 E2E_SUITES_MULTI_RE="$(printf '%s' "${E2E_SUITES_MULTI}" | tr ' ' '|')"
 
+# The push gate always runs this set. A push adds to it the suites whose code
+# it changed, by E2E_PATH_SUITES below; a dispatch names its own set instead.
+E2E_SUITES_PUSH_BASE="single multinode cert iam partialblock"
+
+# "suite prefix" pairs: a push changing a path under prefix adds suite. The
+# recovery and fault suites stay nightly, and shared code adds nothing because
+# the base set covers it. Image directories carry each guest's own scripts.
+E2E_PATH_SUITES="
+rds cmd/rds-agent/
+rds internal/rdsgw/
+rds spinifex/agents/rds/
+rds spinifex/domains/rds/
+rds tests/e2e/rds/
+rds scripts/images/rds-mariadb/
+rds scripts/images/rds-postgres/
+ecs cmd/ecs-agent/
+ecs contracts/ecs/
+ecs internal/ecsgw/
+ecs spinifex/agents/ecs/
+ecs spinifex/domains/ecs/
+ecs tests/e2e/ecs/
+ecs scripts/images/ecs-agent/
+eks cmd/eks-
+eks contracts/eks/
+eks internal/eksgw/
+eks spinifex/agents/eks/
+eks spinifex/domains/eks/
+eks tests/e2e/eks/
+eks scripts/images/eks-node/
+rds scripts/images/common/
+ecs scripts/images/common/
+eks scripts/images/common/
+rds scripts/build-system-image.sh
+ecs scripts/build-system-image.sh
+eks scripts/build-system-image.sh
+route53 spinifex/domains/route53/
+route53 tests/e2e/route53/
+lb cmd/lb-agent/
+lb spinifex/agents/elbv2/
+lb spinifex/domains/elbv2/
+lb tests/e2e/lb/
+"
+
+# Every suite E2E_PATH_SUITES can name, once each, in table order.
+#   e2e_path_suites_all
+e2e_path_suites_all() {
+  printf '%s\n' "${E2E_PATH_SUITES}" | awk 'NF == 2 && !seen[$1]++ { print $1 }'
+}
+
+# The suites a change to <path> triggers, one per line.
+#   e2e_suites_for_path <path>
+e2e_suites_for_path() {
+  local path="$1" suite prefix
+  while read -r suite prefix; do
+    [ -n "${prefix}" ] || continue
+    case "${path}" in
+      "${prefix}"*) printf '%s\n' "${suite}" ;;
+    esac
+  done <<<"${E2E_PATH_SUITES}"
+}
+
 # What every e2e-nightly permutation cell must prove, deliberately narrower
 # than the two sets above: those cells have a 35-minute budget and exist to show
 # that every install / network / host-OS combination boots and serves. The

.github/scripts/test_mulga_design_check.py

@@ -1,23 +0,0 @@
-import unittest
-
-from mulga_design_check import filled
-
-
-class FilledTest(unittest.TestCase):
-    def test_accepts_either_form(self):
-        self.assertTrue(filled("## Mulga design\nUpdated: mulga#275"))
-        self.assertTrue(filled("## Mulga design\n\nnot needed: internal refactor\n## Next"))
-
-    def test_rejects_unanswered(self):
-        self.assertFalse(filled(""))
-        self.assertFalse(filled("## Mulga design\n<!-- Updated: x -->"))
-        self.assertFalse(filled("## Mulga design\nUpdated: <mulga PR or bead id>"))
-        self.assertFalse(filled("## Mulga design\nUpdated:"))
-
-    def test_scoped_to_the_section(self):
-        self.assertFalse(filled("Updated: mulga#1"))
-        self.assertFalse(filled("## Mulga design\n## Other\nUpdated: mulga#1"))
-
-
-if __name__ == "__main__":
-    unittest.main()

.github/scripts/wait-for-memory_test.sh

@@ -1,47 +0,0 @@
-#!/bin/sh
-# wait-for-memory_test.sh — drives wait-for-memory.sh against a fake meminfo
-# that a background writer changes mid-wait, so each verdict is checked without
-# a loaded host.
-set -eu
-
-SCRIPT_DIR=$(CDPATH='' cd -- "$(dirname -- "$0")" && pwd)
-WAIT="$SCRIPT_DIR/wait-for-memory.sh"
-TMP=$(mktemp -d)
-trap 'rm -rf "$TMP"' EXIT
-
-fails=0
-pass() { echo "  ok   - $1"; }
-fail() { echo "  FAIL - $1"; fails=$((fails + 1)); }
-
-meminfo() { printf 'MemTotal:       65536000 kB\nMemAvailable:   %s kB\n' "$(( $1 * 1024 ))" > "$TMP/meminfo"; }
-MEMINFO="$TMP/meminfo"
-MEM_WAIT_POLL_S=1
-export MEMINFO MEM_WAIT_POLL_S
-
-meminfo 30000
-if MEM_WAIT_TIMEOUT_S=5 bash "$WAIT" 26624 > "$TMP/out" 2>&1; then
-  pass "enough memory returns at once"
-else
-  fail "enough memory returns at once: $(cat "$TMP/out")"
-fi
-
-meminfo 21577
-( sleep 2; meminfo 30000 ) &
-if MEM_WAIT_TIMEOUT_S=20 bash "$WAIT" 26624 > "$TMP/out" 2>&1 && grep -q waiting "$TMP/out"; then
-  pass "memory freed mid-wait is admitted"
-else
-  fail "memory freed mid-wait is admitted: $(cat "$TMP/out")"
-fi
-wait
-
-meminfo 21577
-if MEM_WAIT_TIMEOUT_S=2 bash "$WAIT" 26624 > "$TMP/out" 2>&1; then
-  fail "memory that never frees fails: exited 0"
-elif grep -q '::error::need 26624 MiB available, have 21577 MiB' "$TMP/out" && grep -q 'RSS' "$TMP/out"; then
-  pass "memory that never frees fails and names the biggest processes"
-else
-  fail "memory that never frees fails and names the biggest processes: $(cat "$TMP/out")"
-fi
-
-[ "$fails" -eq 0 ] || { echo "wait-for-memory: $fails failed"; exit 1; }
-echo "wait-for-memory: all passed"

.github/workflows/actions.lock

@@ -50,8 +50,6 @@ workflows:
     '.github/workflows/image-scripts.yml':
         - 'actions/checkout@v7.0.1'
         - 'actions/setup-go@v7.0.0'
-    '.github/workflows/mulga-design.yml':
-        - 'actions/checkout@v7.0.1'
     '.github/workflows/perf-3node.yml':
         - 'actions/checkout@v7.0.1'
         - 'actions/create-github-app-token@v3.3.0'

.github/workflows/codex-review.yml

@@ -0,0 +1,21 @@
+# Codex review gate: calls bluebottle's reusable workflows from main, so a PR cannot change its own review.
+name: codex-review
+
+on:
+  pull_request_target:
+    branches: [dev]
+    types: [opened, synchronize, reopened, ready_for_review, edited]
+  issue_comment:
+    types: [created]
+
+permissions: {}
+
+jobs:
+  review:
+    if: github.event_name == 'pull_request_target'
+    uses: mulgadc/bluebottle/.github/workflows/codex-pr-review.yml@main
+    secrets: inherit
+  waive:
+    if: github.event_name == 'issue_comment' && github.event.issue.pull_request
+    uses: mulgadc/bluebottle/.github/workflows/codex-waive.yml@main
+    secrets: inherit

.github/workflows/dev-release.yml

@@ -30,6 +30,10 @@ concurrency:
 jobs:
   prerelease:
     runs-on: ubuntu-26.04
+    env:
+      # Dev prereleases, newest published first: GitHub lists releases by tag name
+      # as text, so list order is not build order. The install Worker uses this rule.
+      DEV_RELEASES_JQ: '[.[] | select(.draft == false and .prerelease and .published_at != null and (.tag_name | test("-dev(\\.|$)")))] | sort_by(.published_at) | reverse'
     steps:
       # Full history and tags, because the tag name is git describe and a shallow
       # clone has no tag to describe against.
@@ -164,12 +168,12 @@ jobs:
               exit 1
           fi
           echo "latest is still $latest"
-          dev="$(gh api "repos/$GITHUB_REPOSITORY/releases?per_page=20" --jq 'map(select(.prerelease))[0].tag_name')"
+          dev="$(gh api "repos/$GITHUB_REPOSITORY/releases?per_page=100" --jq "$DEV_RELEASES_JQ | .[0].tag_name")"
           [ "$dev" = "$TAG" ] || { echo "::error::the dev channel resolves to $dev, not $TAG"; exit 1; }
 
-      # Kept rather than unbounded: the Worker reads the first page of releases to
-      # resolve the channel, so enough prereleases would eventually push the real
-      # ones off it. Five leaves a few recent builds to roll back to.
+      # Kept rather than unbounded: the Worker reads one page of releases to resolve
+      # the channel, so enough prereleases would eventually push the newest off it.
+      # Five leaves a few recent builds to roll back to.
       - name: Prune old dev prereleases
         if: steps.version.outputs.exists != 'true'
         env:
@@ -177,7 +181,7 @@ jobs:
         run: |
           set -euo pipefail
           gh api "repos/$GITHUB_REPOSITORY/releases?per_page=100" \
-              --jq 'map(select(.prerelease and (.tag_name | endswith("-dev")))) | .[5:] | .[].tag_name' \
+              --jq "$DEV_RELEASES_JQ | .[5:] | .[].tag_name" \
               > stale.txt
           while read -r tag; do
               [ -n "$tag" ] || continue

.github/workflows/e2e-baremetal.yml

@@ -71,9 +71,16 @@ jobs:
       ARTIFACT_DIR: /tmp/spinifex-e2e-nightly-cell-${{ inputs.cell || '20' }}-${{ github.run_id }}
       SPINIFEX_OTEL_ATTRS: ${{ inputs.otel-attrs || format('ci.run_id={0},ci.workflow={1},ci.cell={2},ci.matrix=baremetal/pxe/real-hw', github.run_id, github.workflow, inputs.cell || '20') }}
     steps:
+      # A cancelled job's root steps outlive it, since the runner cannot reach
+      # them once it kills sudo, and would keep writing here and to the cache.
+      # A rootfs killed before bin/bash exists is one the build refuses to reuse.
       - name: Clean Workspace
         run: |
+          sudo systemctl stop spinifex-bm-pxe.scope spinifex-bm-e2e.scope 2>/dev/null || true
           sudo umount -R -l "${BM_ISO_CACHE}/rootfs" 2>/dev/null || true
+          if [ -d "${BM_ISO_CACHE}/rootfs" ] && [ ! -x "${BM_ISO_CACHE}/rootfs/bin/bash" ]; then
+            sudo rm -rf --one-file-system "${BM_ISO_CACHE}/rootfs"
+          fi
           if [ ! -c /dev/null ]; then sudo rm -f /dev/null; sudo mknod -m 666 /dev/null c 1 3; fi
           sudo rm -rf --one-file-system "$GITHUB_WORKSPACE"/*
 
@@ -113,10 +120,15 @@ jobs:
             "$GITHUB_WORKSPACE" \
             "${RUNNER_TEMP}/spinifex-tests.tar" \
             "${RUNNER_TEMP}/bm-bins"
+          # Built here, as the runner user, so the PXE build finds it current.
+          make -C "$GITHUB_WORKSPACE" build-microvm-image
 
       - name: Build PXE Artifacts
         run: |
-          sudo make -C mulga/scripts/iso-builder \
+          # -o build-spinifex: the binaries come from Build Artifacts, and a
+          # rebuild here would write root-owned files into the workspace.
+          sudo systemd-run --scope --collect --quiet --unit=spinifex-bm-pxe -- \
+            make -C mulga/scripts/iso-builder -o build-spinifex \
             build-rootfs build-squashfs build-initramfs \
             CACHE="${BM_ISO_CACHE}" \
             SPINIFEX_BIN="${RUNNER_TEMP}/bm-bins/spx" \
@@ -124,6 +136,10 @@ jobs:
             SPINIFEX_SRC="${GITHUB_WORKSPACE}" \
             VIPERBLOCK_SRC="${RUNNER_TEMP}/bm-bins/viperblock" \
             VERSION="$(cd "$GITHUB_WORKSPACE" && git describe --tags --always --exclude '*-dev' 2>/dev/null || echo dev)"
+          if root_owned="$(find "$GITHUB_WORKSPACE" -user root -print -quit)" && [ -n "$root_owned" ]; then
+            echo "::error::the PXE build wrote root-owned files into the workspace, e.g. $root_owned"
+            exit 1
+          fi
 
       - name: Write ePDU credentials
         run: |
@@ -195,13 +211,18 @@ jobs:
           done
           export BM_TEST_TIMEOUT
           echo "suites=${BM_SUITES} test-timeout=${BM_TEST_TIMEOUT}"
-          sudo -E mulga/scripts/baremetal/run-baremetal-e2e.sh \
+          sudo -E systemd-run --scope --collect --quiet --unit=spinifex-bm-e2e -- \
+            mulga/scripts/baremetal/run-baremetal-e2e.sh \
             --vmlinuz   "${BM_ISO_CACHE}/vmlinuz" \
             --initrd    "${BM_ISO_CACHE}/initrd.img" \
             --squashfs  "${BM_ISO_CACHE}/spinifex.squashfs" \
             --tests-tar "${RUNNER_TEMP}/spinifex-tests.tar" \
             2>&1 | tee "${ARTIFACT_DIR}/test-baremetal.log"
 
+      - name: Stop this job's root steps
+        if: always()
+        run: sudo systemctl stop spinifex-bm-pxe.scope spinifex-bm-e2e.scope 2>/dev/null || true
+
       - name: Generate JUnit report
         if: always()
         run: |

.github/workflows/e2e-nightly.yml

@@ -6,11 +6,32 @@ name: E2E Nightly
 # `main` and `dev` as separate nightlies, and it would always execute `main`'s
 # copy of this file — leaving a fix to the workflow itself untestable until it
 # had already shipped. A dispatch takes both the file and the source from its ref.
+#
+# Cells (matrix cells are defined in .github/e2e-cells.json):
+#
+#   Single node                       Three nodes
+#    1  Debian 13, DHCP                 8  Debian 13, DHCP
+#    2  Ubuntu 26.04, static            9  Ubuntu 26.04, static
+#    3  Ubuntu 26.04, DHCP             10  Ubuntu 26.04, DHCP
+#   17  tofu examples                  29  storage fault injection
+#   18  reboot                         30  routed NAT
+#   19  routed NAT                     31  instance recovery
+#   25  eks                            32  load balancer recovery
+#   26  ecs                            33  KV bucket quorum under node loss
+#   27  rds                            37  Oracle Linux 10, DHCP
+#   34  Oracle Linux 10, DHCP          38  Oracle Linux 10, routed NAT
+#   35  Oracle Linux 10, reboot
+#   36  Oracle Linux 10, routed NAT
+#
+#   Standalone jobs
+#   20  bare metal (PXE)               23  predastore throughput
+#   21  ISO, single node               24  predastore fault injection
+#   22  ISO, three nodes               28  guest disk performance (bare metal)
 on:
   workflow_dispatch:
     inputs:
       cells:
-        description: "Comma-separated cell numbers to run, no spaces (blank = all). 1-3, 8-10, 17-19 tofu permutation matrix, 34-36 Oracle Linux single node (basic, reboot, routed NAT), 37-38 Oracle Linux on three nodes (multi-node subset with IPsec, routed NAT), 20 baremetal, 21-22 ISO, 23 predastore throughput, 24 predastore fault injection, 25-27 service suites (eks, ecs, rds), 28 guest disk performance, 29 storage fault injection, 30 routed NAT on three nodes, 31 instance recovery on three nodes, 32 load balancer recovery on three nodes, 33 KV bucket quorum under node loss on three nodes"
+        description: "Comma-separated cell numbers, no spaces, e.g. 1,25,27 (blank = all). The cell list is at the top of .github/workflows/e2e-nightly.yml."
         required: false
 
 permissions:
@@ -1017,7 +1038,7 @@ jobs:
         id: siblings
         env:
           BRANCH: ${{ env.GIT_BRANCH }}
-        run: bash .github/resolve-sibling-refs.sh
+        run: bash .github/scripts/resolve-sibling-refs.sh
         working-directory: mulga/spinifex
 
       - uses: actions/checkout@v7.0.1
@@ -1275,7 +1296,7 @@ jobs:
         id: siblings
         env:
           BRANCH: ${{ env.GIT_BRANCH }}
-        run: bash .github/resolve-sibling-refs.sh
+        run: bash .github/scripts/resolve-sibling-refs.sh
         working-directory: mulga/spinifex
 
       - uses: actions/checkout@v7.0.1
@@ -1423,7 +1444,7 @@ jobs:
         id: siblings
         env:
           BRANCH: ${{ github.ref_name }}
-        run: bash .github/resolve-sibling-refs.sh
+        run: bash .github/scripts/resolve-sibling-refs.sh
         working-directory: spinifex
 
       - name: Checkout Predastore
@@ -1547,7 +1568,7 @@ jobs:
         id: siblings
         env:
           BRANCH: ${{ github.ref_name }}
-        run: bash .github/resolve-sibling-refs.sh
+        run: bash .github/scripts/resolve-sibling-refs.sh
         working-directory: spinifex
 
       - name: Checkout Predastore

.github/workflows/e2e.yml

@@ -12,13 +12,15 @@ on:
       - go.sum
       - Makefile
       - tests/e2e/**
+      - scripts/images/**
+      - scripts/build-system-image.sh
       - .github/workflows/e2e.yml
   workflow_dispatch:
     inputs:
       suites:
         description: "Space-separated Make targets to run (e.g. 'e2e-single e2e-multinode e2e-cert e2e-lb e2e-eks e2e-iam e2e-rds')"
         required: false
-        default: "e2e-single e2e-multinode e2e-cert e2e-lb e2e-iam"
+        default: "e2e-single e2e-multinode e2e-cert e2e-iam"
       test_run:
         description: "Optional regex for -test.run (e.g. '^TestMultinodeIPSec$'). Empty = run all tests in suite."
         required: false
@@ -31,86 +33,75 @@ permissions:
 
 env:
   GIT_BRANCH: ${{ github.ref_name }}
-  SUITES: ${{ inputs.suites || 'e2e-single e2e-multinode e2e-cert e2e-lb e2e-iam e2e-partialblock' }}
   TEST_RUN: ${{ inputs.test_run || '' }}
   SPINIFEX_OTLP_ENDPOINT: ${{ vars.SPINIFEX_OTLP_ENDPOINT || 'http://192.168.1.13:8200' }}
   SPINIFEX_OTEL_ATTRS: ci.run_id=${{ github.run_id }},ci.workflow=${{ github.workflow }}
 
 jobs:
-  # The inner timeout must fire before the outer one. `go test` reports where it
-  # stopped and the suite's own failure; a job cancelled on timeout reports only
-  # "The operation was canceled" and takes the artifacts with it.
-  #
-  # A job budget cannot be a literal here because the suite set is a dispatch
-  # input: storagefault asks for 90m and diskperf for 180m, both eligible on one
-  # node, so any fixed number is either wrong for them or wasteful for everyone
-  # else. Resolving it from suite-sets.sh keeps the one source of truth and
-  # leaves the default push set at exactly the 60m it has always had.
+  # A push runs the base set plus the suites its changed paths map to in
+  # suite-sets.sh; a dispatch runs its suites input, or the base set if empty.
+  # Budgets come from the suites' own timeouts there.
   plan:
-    name: Resolve Job Budgets
+    name: Select Suites
     runs-on: ubuntu-26.04
     timeout-minutes: 5
     outputs:
-      single_timeout: ${{ steps.budgets.outputs.single_timeout }}
-      multi_timeout: ${{ steps.budgets.outputs.multi_timeout }}
+      single_legs: ${{ steps.select.outputs.single_legs }}
+      multi_suites: ${{ steps.select.outputs.multi_suites }}
+      multi_timeout: ${{ steps.select.outputs.multi_timeout }}
     steps:
+      # Full history without file contents: the diff needs both commits' trees.
       - uses: actions/checkout@v7.0.1
         with:
           path: spinifex
+          fetch-depth: 0
+          filter: blob:none
           sparse-checkout: .github/scripts
 
-      - name: Resolve Budgets
-        id: budgets
+      - name: Select suites
+        id: select
+        working-directory: spinifex
         env:
-          # Provision, install and teardown around the suites, plus the headroom
-          # a multi-suite job needs for whatever ran before the one that hangs.
-          SETUP_ALLOWANCE_MIN: 15
-          # What the job has always had. Kept as a floor so resolving the budget
-          # can only ever raise it, never quietly shorten an existing run.
-          FLOOR_MIN: 60
+          EVENT: ${{ github.event_name }}
+          BEFORE: ${{ github.event.before }}
+          IN_SUITES: ${{ inputs.suites }}
         run: |
-          . "${GITHUB_WORKSPACE}/spinifex/.github/scripts/suite-sets.sh"
-          # Only max() is a sound bound: the suites share one job, so a sum would
-          # demand hours for a set that has never taken them. Max is what makes
-          # the hanging suite panic on its own timeout first, which is the point.
-          budget() {
-            local want="${FLOOR_MIN}" s t
-            for s in $(echo "${SUITES}" | sed 's/e2e-//g' | tr ' ' '\n' | grep -xE "$1"); do
-              t="$(e2e_suite_timeout "${s}")"
-              t=$(( ${t%m} + SETUP_ALLOWANCE_MIN ))
-              [ "${t}" -gt "${want}" ] && want="${t}"
-            done
-            echo "${want}"
-          }
-          single=$(budget "${E2E_SUITES_SINGLE_RE}")
-          multi=$(budget "${E2E_SUITES_MULTI_RE}")
-          echo "single_timeout=${single}" >> "$GITHUB_OUTPUT"
-          echo "multi_timeout=${multi}" >> "$GITHUB_OUTPUT"
-          echo "suites='${SUITES}' -> single ${single}m, multi ${multi}m"
+          set -euo pipefail
+          SELECT=.github/scripts/e2e-select.sh
+          if [ "${EVENT}" = push ]; then
+            bash "${SELECT}" push "${BEFORE}" "${GITHUB_SHA}" | tee -a "$GITHUB_OUTPUT"
+          else
+            bash "${SELECT}" suites "${IN_SUITES}" | tee -a "$GITHUB_OUTPUT"
+          fi
 
   e2e_single:
-    name: Single-Node E2E
-    # The dispatch check gates on exactly the suites this job's Run step
-    # allow-lists, i.e. E2E_SUITES_SINGLE in .github/scripts/suite-sets.sh —
-    # an `if:` expression cannot source that file, so this one list is
-    # hand-mirrored and must be updated with it. Gating on a suite the job
-    # won't run — lb, which only executes in the multinode job — would
-    # provision a tofu VM and pay the Spinifex install only to run nothing and
-    # tear down. Push triggers keep running via the null-check on
-    # inputs.suites.
-    if: github.event_name != 'workflow_dispatch' || contains(github.event.inputs.suites, 'e2e-single') || contains(github.event.inputs.suites, 'e2e-iam') || contains(github.event.inputs.suites, 'e2e-cert') || contains(github.event.inputs.suites, 'e2e-eks') || contains(github.event.inputs.suites, 'e2e-ecs') || contains(github.event.inputs.suites, 'e2e-storagegrowth') || contains(github.event.inputs.suites, 'e2e-partialblock') || contains(github.event.inputs.suites, 'e2e-rds')
+    # The base leg keeps the job and artifact names it always had: mulga's
+    # registry intake finds a run's results by them.
+    name: ${{ matrix.leg == 'base' && 'Single-Node E2E' || format('Single-Node E2E ({0})', matrix.leg) }}
+    # `base` runs on ci-single; `service`, on ci-pseudo, carries the matched
+    # single-node suites so a long one stays off the base leg's critical path.
+    if: needs.plan.outputs.single_legs != '[]'
     needs: plan
-    runs-on: [self-hosted, ci-single]
-    timeout-minutes: ${{ fromJSON(needs.plan.outputs.single_timeout) }}
+    runs-on: [self-hosted, "${{ matrix.runner }}"]
+    timeout-minutes: ${{ matrix.timeout }}
+    strategy:
+      fail-fast: false
+      matrix:
+        include: ${{ fromJSON(needs.plan.outputs.single_legs) }}
     env:
-      TOFU_LOG: /tmp/e2e-tofu-single-${{ github.run_id }}.log
-      ARTIFACT_DIR: /tmp/spinifex-e2e-artifacts-${{ github.run_id }}
+      SUITES: ${{ matrix.suites }}
+      # Both legs can run on banksia at once, so every host path carries the leg.
+      TOFU_LOG: /tmp/e2e-tofu-single-${{ github.run_id }}-${{ matrix.leg }}.log
+      ARTIFACT_DIR: /tmp/spinifex-e2e-artifacts-${{ github.run_id }}-${{ matrix.leg }}
+      # The base leg keeps the cell name it always had, so its history in
+      # Elastic and telemetry continues; any other leg appends its name.
+      E2E_CELL: ${{ matrix.leg == 'base' && 'single' || format('single-{0}', matrix.leg) }}
       # Job-level env wins over the workflow-level default. This job and
       # e2e_multi share one run id and both name their hosts node1/node2/node3,
       # so without ci.cell their telemetry collapses into one bucket. This job
       # is where "Bootstrap via Binary Install" bakes SPINIFEX_OTEL_ATTRS into
       # the VM's telemetry.env, so the label has to be set here.
-      SPINIFEX_OTEL_ATTRS: ci.run_id=${{ github.run_id }},ci.workflow=${{ github.workflow }},ci.cell=single
+      SPINIFEX_OTEL_ATTRS: ci.run_id=${{ github.run_id }},ci.workflow=${{ github.workflow }},ci.cell=${{ matrix.leg == 'base' && 'single' || format('single-{0}', matrix.leg) }}
     steps:
       - name: Reap Orphan Procs
         run: |
@@ -180,7 +171,7 @@ jobs:
       - name: Resolve CI job id
         uses: $/.github/actions/otel-job-id
         with:
-          job-name: Single-Node E2E
+          job-name: ${{ matrix.leg == 'base' && 'Single-Node E2E' || format('Single-Node E2E ({0})', matrix.leg) }}
 
       - name: Build Install Artifacts
         working-directory: mulga/scripts/tofu-cluster
@@ -224,6 +215,12 @@ jobs:
           echo "wan_ip=${WAN_IP}" >> "$GITHUB_OUTPUT"
           echo "ssh_key=${SSH_KEY}" >> "$GITHUB_OUTPUT"
 
+      # build-system-image.sh reuses a raw image under ten minutes old from
+      # its build directory, and every runner on banksia shares the default
+      # one in /tmp, so a job could test another commit's guest agent.
+      - name: Use a per-job image build directory
+        run: echo "SYSTEM_IMAGE_BUILD_DIR=${RUNNER_TEMP}/system-image-build" >> "$GITHUB_ENV"
+
       # The eks suite's CreateCluster boots a control-plane VM from the
       # eks-node AMI, so that AMI must exist in the VM's image store before the
       # suite runs. build-system-image.sh customizes the image with libguestfs
@@ -247,7 +244,7 @@ jobs:
           group_open "make build-eks-node-image"
           make build-eks-node-image
           group_close
-          ok "raw: $(du -h /tmp/eks-node-image-build/eks-node-alpine.raw | cut -f1)"
+          ok "raw: $(du -h "${SYSTEM_IMAGE_BUILD_DIR}/eks-node-image-build/eks-node-alpine.raw" | cut -f1)"
 
       - name: Import eks-node AMI on VM
         if: contains(env.SUITES, 'e2e-eks')
@@ -259,13 +256,16 @@ jobs:
           SSH_OPTS=(-o StrictHostKeyChecking=no -i "${{ steps.vm.outputs.ssh_key }}")
           VM=tf-user@${{ steps.vm.outputs.wan_ip }}
           banner "import eks-node AMI to VM image store"
+          # The VM's /tmp is a tmpfs and scp writes each raw at its full size,
+          # so every import removes its file before the next image arrives.
           scp "${SSH_OPTS[@]}" \
-            /tmp/eks-node-image-build/eks-node-alpine.raw \
+            "${SYSTEM_IMAGE_BUILD_DIR}/eks-node-image-build/eks-node-alpine.raw" \
             "$VM":/tmp/eks-node-alpine.raw
           ssh "${SSH_OPTS[@]}" "$VM" \
             "spx admin images import --file /tmp/eks-node-alpine.raw \
                --distro alpine --version ${ALPINE_VERSION} --arch x86_64 \
-               --boot-mode bios --ami-name ${AMI_NAME} --tag ${SYSTEM_TAG}" >/dev/null 2>&1
+               --boot-mode bios --ami-name ${AMI_NAME} --tag ${SYSTEM_TAG}; \
+               rc=\$?; rm -f /tmp/eks-node-alpine.raw; exit \$rc" >/dev/null 2>&1
           ok "eks-node AMI imported"
 
       # The eks data-plane subtest (TestEKS/KubectlGetNodes) drives the full
@@ -299,6 +299,40 @@ jobs:
           EOF
           ok "kubectl installed on VM"
 
+      # The ecs suite boots its container instance from the ecs-node AMI, found
+      # by its managed-by tag, so the tag matters. The build directory is named
+      # for the manifest's IMAGE_NAME (ecs-agent), not for the make target.
+      - name: Build ecs-node AMI (libguestfs, runner)
+        if: contains(env.SUITES, 'e2e-ecs')
+        working-directory: spinifex
+        run: |
+          set -euo pipefail
+          # direct backend = no libvirt. Set here rather than in the script:
+          # it has to be in the environment the build itself runs in.
+          export LIBGUESTFS_BACKEND=direct
+          "${GITHUB_WORKSPACE}/spinifex/.github/scripts/ensure-libguestfs.sh"
+          make build-ecs-node-image
+
+      - name: Import ecs-node AMI on VM
+        if: contains(env.SUITES, 'e2e-ecs')
+        run: |
+          set -euo pipefail
+          source "${GITHUB_WORKSPACE}/spinifex/.github/scripts/ci-fmt.sh"
+          eval "$(grep -E '^(ALPINE_VERSION|AMI_NAME|SYSTEM_TAG)=' \
+            spinifex/scripts/images/ecs-agent/manifest.conf)"
+          SSH_OPTS=(-o StrictHostKeyChecking=no -i "${{ steps.vm.outputs.ssh_key }}")
+          VM=tf-user@${{ steps.vm.outputs.wan_ip }}
+          banner "import ecs-node AMI to VM image store"
+          scp "${SSH_OPTS[@]}" \
+            "${SYSTEM_IMAGE_BUILD_DIR}/ecs-agent-image-build/ecs-agent-alpine.raw" \
+            "$VM":/tmp/ecs-agent-alpine.raw
+          ssh "${SSH_OPTS[@]}" "$VM" \
+            "spx admin images import --file /tmp/ecs-agent-alpine.raw \
+               --distro alpine --version ${ALPINE_VERSION} --arch x86_64 \
+               --boot-mode bios --ami-name ${AMI_NAME} --tag ${SYSTEM_TAG}; \
+               rc=\$?; rm -f /tmp/ecs-agent-alpine.raw; exit \$rc" >/dev/null 2>&1
+          ok "ecs-node AMI imported"
+
       # CreateDBInstance boots the DB VM from an RDS AMI, so both engines' images
       # must be in the VM's image store before the suite runs — same split as the
       # eks-node pair above: build on the runner (needs source + network), import
@@ -321,7 +355,7 @@ jobs:
             group_open "make build-${IMAGE}-image"
             make "build-${IMAGE}-image"
             group_close
-            ok "raw: $(du -h "/tmp/${IMAGE}-image-build/${IMAGE}-alpine.raw" | cut -f1)"
+            ok "raw: $(du -h "${SYSTEM_IMAGE_BUILD_DIR}/${IMAGE}-image-build/${IMAGE}-alpine.raw" | cut -f1)"
           done
 
       - name: Import RDS AMIs on VM
@@ -343,12 +377,13 @@ jobs:
             for tag in ${SYSTEM_TAG}; do TAG_ARGS="${TAG_ARGS} --tag ${tag}"; done
             banner "import ${IMAGE} AMI to VM image store"
             scp "${SSH_OPTS[@]}" \
-              "/tmp/${IMAGE}-image-build/${IMAGE}-alpine.raw" \
+              "${SYSTEM_IMAGE_BUILD_DIR}/${IMAGE}-image-build/${IMAGE}-alpine.raw" \
               "$VM":"/tmp/${IMAGE}-alpine.raw"
             ssh "${SSH_OPTS[@]}" "$VM" \
               "spx admin images import --file /tmp/${IMAGE}-alpine.raw \
                  --distro alpine --version ${ALPINE_VERSION} --arch x86_64 \
-                 --boot-mode bios --ami-name ${AMI_NAME}${TAG_ARGS}" >/dev/null 2>&1
+                 --boot-mode bios --ami-name ${AMI_NAME}${TAG_ARGS}; \
+                 rc=\$?; rm -f /tmp/${IMAGE}-alpine.raw; exit \$rc" >/dev/null 2>&1
             ok "${IMAGE} AMI imported"
           done
 
@@ -526,7 +561,7 @@ jobs:
         uses: $/.github/actions/junit-to-es
         with:
           junit-glob: ${{ env.ARTIFACT_DIR }}/junit-*.xml
-          cell: single
+          cell: ${{ env.E2E_CELL }}
 
       - name: Analyze e2e failures
         if: always()
@@ -546,7 +581,7 @@ jobs:
       - uses: actions/upload-artifact@v7.0.2
         if: always()
         with:
-          name: e2e-artifacts-${{ github.run_id }}
+          name: ${{ matrix.leg == 'base' && format('e2e-artifacts-{0}', github.run_id) || format('e2e-artifacts-{0}-{1}', github.run_id, matrix.leg) }}
           path: ${{ env.ARTIFACT_DIR }}
           if-no-files-found: warn
           retention-days: 14
@@ -589,7 +624,7 @@ jobs:
         if: always()
         uses: actions/upload-artifact@v7.0.2
         with:
-          name: tofu-log-single-${{ github.run_id }}
+          name: ${{ matrix.leg == 'base' && format('tofu-log-single-{0}', github.run_id) || format('tofu-log-single-{0}-{1}', github.run_id, matrix.leg) }}
           path: ${{ env.TOFU_LOG }}
           if-no-files-found: ignore
           retention-days: 14
@@ -605,15 +640,13 @@ jobs:
 
   e2e_multi:
     name: Multi-Node E2E
-    # The dispatch check gates on exactly the suites this job's Run step
-    # allow-lists (multinode|lb|cert) so lb — which only runs here — still
-    # provisions the 3-node env when requested on its own, and the env stays
-    # unprovisioned when none of them are asked for.
-    if: github.event_name != 'workflow_dispatch' || contains(github.event.inputs.suites, 'e2e-multinode') || contains(github.event.inputs.suites, 'e2e-lb') || contains(github.event.inputs.suites, 'e2e-cert')
+    # Runs when the plan job selected any multi-node suite, and only then.
+    if: needs.plan.outputs.multi_suites != ''
     needs: plan
     runs-on: [self-hosted, ci-multi]
     timeout-minutes: ${{ fromJSON(needs.plan.outputs.multi_timeout) }}
     env:
+      SUITES: ${{ needs.plan.outputs.multi_suites }}
       TOFU_LOG: /tmp/e2e-tofu-multi-${{ github.run_id }}.log
       ARTIFACT_DIR: /tmp/spinifex-e2e-multi-artifacts-${{ github.run_id }}
       # See the matching note on e2e_single: ci.cell is what separates this
@@ -1070,7 +1103,10 @@ jobs:
   e2e_status:
     name: Report E2E Status
     if: always()
-    needs: [e2e_single, e2e_multi]
+    # plan is a dependency because a failed plan skips both e2e jobs, and a
+    # skipped job counts as passed below: without it a broken selection would
+    # report green having run nothing.
+    needs: [plan, e2e_single, e2e_multi]
     runs-on: ubuntu-26.04
     timeout-minutes: 5
     steps:
@@ -1079,15 +1115,16 @@ jobs:
           GH_TOKEN: ${{ github.token }}
           REPORT_SHA: ${{ github.sha }}
         run: |
-          single_ok=true; multi_ok=true
+          plan_ok=true; single_ok=true; multi_ok=true
+          [ "${{ needs.plan.result }}" = success ] || plan_ok=false
           case "${{ needs.e2e_single.result }}" in success|skipped) ;; *) single_ok=false ;; esac
           case "${{ needs.e2e_multi.result }}" in success|skipped) ;; *) multi_ok=false ;; esac
-          if $single_ok && $multi_ok; then
+          if $plan_ok && $single_ok && $multi_ok; then
             STATE="success"
             DESC="E2E suites passed"
           else
             STATE="failure"
-            DESC="E2E failed: single=${{ needs.e2e_single.result }} multi=${{ needs.e2e_multi.result }}"
+            DESC="E2E failed: plan=${{ needs.plan.result }} single=${{ needs.e2e_single.result }} multi=${{ needs.e2e_multi.result }}"
           fi
           gh api "repos/${{ github.repository }}/statuses/${REPORT_SHA}" \
             -f state="$STATE" \
@@ -1104,7 +1141,7 @@ jobs:
   # status-report check.
   e2e_es_summary:
     name: ES CI Run Summary
-    needs: [e2e_single, e2e_multi]
+    needs: [plan, e2e_single, e2e_multi]
     if: always()
     runs-on: [self-hosted, ci-single]
     timeout-minutes: 5
@@ -1116,6 +1153,11 @@ jobs:
       # this host, so no credentials are needed for the POST.
       MULGA_ES_SINK: ${{ vars.MULGA_ES_SINK || '192.168.1.13:9200' }}
     steps:
+      # The workspace outlives the run on a self-hosted runner, and the status
+      # files are collected recursively, so an earlier run's would count here.
+      - name: Clear previous summary downloads
+        run: rm -rf "${GITHUB_WORKSPACE}/es-summary-single" "${GITHUB_WORKSPACE}/es-summary-multi"
+
       # Either leg may not have run (e.g. a workflow_dispatch that only
       # requested single-node suites), so a missing artifact is tolerated —
       # the aggregation script below treats an absent suite-status.txt as
@@ -1124,7 +1166,7 @@ jobs:
         uses: actions/download-artifact@v8.0.2
         continue-on-error: true
         with:
-          name: e2e-artifacts-${{ github.run_id }}
+          pattern: e2e-artifacts-${{ github.run_id }}*
           path: ${{ github.workspace }}/es-summary-single
 
       - name: Download multi-node artifacts
@@ -1144,6 +1186,7 @@ jobs:
           BRANCH: ${{ github.head_ref || github.ref_name }}
           SHA: ${{ github.sha }}
           SERVER_URL: ${{ github.server_url }}
+          PLAN_RESULT: ${{ needs.plan.result }}
           SINGLE_RESULT: ${{ needs.e2e_single.result }}
           MULTI_RESULT: ${{ needs.e2e_multi.result }}
           SINGLE_DIR: ${{ github.workspace }}/es-summary-single
@@ -1151,7 +1194,8 @@ jobs:
         run: |
           set -euo pipefail
 
-          SINGLE_STATUS="${SINGLE_DIR}/suite-status.txt"
+          # One subdirectory per single-node leg's artifact.
+          SINGLE_STATUS=$(find "${SINGLE_DIR}" -name suite-status.txt 2>/dev/null | sort || true)
           MULTI_STATUS="${MULTI_DIR}/suite-status.txt"
 
           # Collect "<suite>: FAILED" lines from both legs' suite-status.txt
@@ -1160,7 +1204,9 @@ jobs:
           # rather than a bash array so this stays correct under `set -u`
           # even when nothing failed.
           FAILED_SUITES=""
-          for f in "$SINGLE_STATUS" "$MULTI_STATUS"; do
+          # SINGLE_STATUS is a newline-separated list, so it is split here.
+          # shellcheck disable=SC2086
+          for f in ${SINGLE_STATUS} "$MULTI_STATUS"; do
             [ -f "$f" ] || continue
             while IFS= read -r line; do
               case "$line" in
@@ -1180,7 +1226,7 @@ jobs:
           case "$MULTI_RESULT" in success | skipped) ;; *) multi_ok=false ;; esac
 
           STATE=success
-          if [ "$single_ok" != true ] || [ "$multi_ok" != true ] || [ "$FAILED_COUNT" -gt 0 ]; then
+          if [ "$PLAN_RESULT" != success ] || [ "$single_ok" != true ] || [ "$multi_ok" != true ] || [ "$FAILED_COUNT" -gt 0 ]; then
             STATE=failure
           fi
 

.github/workflows/mulga-design.yml

@@ -1,33 +0,0 @@
-name: Mulga design
-
-# Reminds the author to say whether the Mulga design records need to change.
-# It reads only the pull request body, uses no secrets, and warns without ever
-# failing: the check is a prompt, not a gate.
-
-on:
-  pull_request:
-    branches:
-      - dev
-    types: [opened, edited, reopened]
-
-permissions:
-  contents: read
-
-jobs:
-  mulga_design:
-    name: Mulga design line
-    if: ${{ !endsWith(github.actor, '[bot]') }}
-    runs-on: ubuntu-26.04
-    steps:
-      - name: Checkout Code
-        uses: actions/checkout@v7.0.1
-        with:
-          sparse-checkout: .github/scripts
-
-      - name: Check the Mulga design line
-        env:
-          PR_BODY: ${{ github.event.pull_request.body }}
-        working-directory: .github/scripts
-        run: |
-          python3 -m unittest -q test_mulga_design_check
-          python3 mulga_design_check.py

.github/workflows/perf-3node.yml

@@ -113,7 +113,7 @@ jobs:
         id: siblings
         env:
           BRANCH: ${{ env.GIT_BRANCH }}
-        run: bash .github/resolve-sibling-refs.sh
+        run: bash .github/scripts/resolve-sibling-refs.sh
         working-directory: spinifex
 
       - uses: actions/checkout@v7.0.1

.github/workflows/publish-system-ami.yml

@@ -84,7 +84,7 @@ jobs:
 
       - name: Resolve sibling release tags from go.mod
         id: siblings
-        run: bash spinifex/.github/resolve-release-sibling-refs.sh spinifex/go.mod
+        run: bash spinifex/.github/scripts/resolve-release-sibling-refs.sh spinifex/go.mod
 
       - name: Checkout Viperblock
         uses: actions/checkout@v7.0.1

.github/workflows/release.yml

@@ -18,7 +18,7 @@ jobs:
 
       - name: Resolve sibling release tags from go.mod
         id: siblings
-        run: bash spinifex/.github/resolve-release-sibling-refs.sh spinifex/go.mod
+        run: bash spinifex/.github/scripts/resolve-release-sibling-refs.sh spinifex/go.mod
 
       - name: Checkout Predastore
         uses: actions/checkout@v7.0.1
@@ -85,7 +85,7 @@ jobs:
 
       - name: Resolve sibling release tags from go.mod
         id: siblings
-        run: bash spinifex/.github/resolve-release-sibling-refs.sh spinifex/go.mod
+        run: bash spinifex/.github/scripts/resolve-release-sibling-refs.sh spinifex/go.mod
 
       - name: Checkout Predastore
         uses: actions/checkout@v7.0.1
@@ -335,7 +335,7 @@ jobs:
 
       - name: Resolve sibling release tags from go.mod
         id: siblings
-        run: bash spinifex/.github/resolve-release-sibling-refs.sh spinifex/go.mod
+        run: bash spinifex/.github/scripts/resolve-release-sibling-refs.sh spinifex/go.mod
 
       - name: Checkout Predastore
         uses: actions/checkout@v7.0.1

.github/workflows/segscan-oracle.yml

@@ -34,7 +34,7 @@ jobs:
         id: siblings
         env:
           BRANCH: ${{ github.head_ref || github.ref_name }}
-        run: bash .github/resolve-sibling-refs.sh
+        run: bash .github/scripts/resolve-sibling-refs.sh
         working-directory: spinifex
 
       - name: Checkout Viperblock

.github/workflows/spinifex.yml

@@ -20,6 +20,12 @@ on:
 permissions:
   contents: read
 
+# A newer push to the same pull request supersedes its run. Pushes and dispatches
+# get a group of their own, so every commit on dev and main keeps its result.
+concurrency:
+  group: ${{ github.workflow }}-${{ github.event_name == 'pull_request' && format('pr-{0}', github.event.pull_request.number) || github.run_id }}
+  cancel-in-progress: true
+
 jobs:
   build_and_test:
     name: Build and Test
@@ -36,7 +42,7 @@ jobs:
         id: siblings
         env:
           BRANCH: ${{ github.head_ref || github.ref_name }}
-        run: bash .github/resolve-sibling-refs.sh
+        run: bash .github/scripts/resolve-sibling-refs.sh
         working-directory: spinifex
 
       - name: Checkout Viperblock
@@ -99,6 +105,13 @@ jobs:
           key: gobuild-${{ github.job }}-${{ github.run_id }}
           restore-keys: gobuild-${{ github.job }}-
 
+      # Go refreshes an entry's mtime on use only when it is over an hour old, so
+      # restored entries are backdated past that before any Go command runs and
+      # the trim below can tell this run's working set from the last run's.
+      - name: Mark Go build cache
+        if: ${{ github.ref == 'refs/heads/dev' || github.ref == 'refs/heads/main' }}
+        run: spinifex/.github/scripts/go-build-cache.sh mark
+
       - name: Setup Go Workspace
         run: go work init ./spinifex ./viperblock ./predastore ./northstar ./bluebottle
 
@@ -131,19 +144,12 @@ jobs:
       - name: Check Diff Coverage
         run: cd spinifex && scripts/diff-coverage.sh coverage.out --base HEAD~1
 
-      # Compiled output stays valid whatever the tests said, and skipping the
-      # save on a red run would leave the next one rebuilding it, so this is
-      # gated on the branch rather than on success.
-      # Go bumps an entry's mtime when a run uses it, so this run's working set
-      # survives and objects superseded by later commits do not. Runs after the
-      # work, so it costs this job nothing and only shrinks what the next reads.
+      # Keeps what this run read or wrote and drops the rest, so the saved entry
+      # is one run's working set however often dev is pushed. A red run may have
+      # stopped early and used only part of the set, so it saves untrimmed.
       - name: Trim Go build cache
-        if: ${{ !cancelled() && (github.ref == 'refs/heads/dev' || github.ref == 'refs/heads/main') }}
-        run: |
-          C="$(go env GOCACHE)"
-          before=$(du -sm "$C" | cut -f1)
-          find "$C" -type f -mmin +240 -delete
-          echo "trimmed ${before}MB -> $(du -sm "$C" | cut -f1)MB"
+        if: ${{ success() && (github.ref == 'refs/heads/dev' || github.ref == 'refs/heads/main') }}
+        run: spinifex/.github/scripts/go-build-cache.sh trim
 
       - name: Save Go build cache
         if: ${{ !cancelled() && (github.ref == 'refs/heads/dev' || github.ref == 'refs/heads/main') }}
@@ -154,6 +160,7 @@ jobs:
 
   race_detection:
     name: Race Detection
+    if: ${{ github.event_name != 'pull_request' }}
     runs-on: ubuntu-26.04
 
     steps:
@@ -166,7 +173,7 @@ jobs:
         id: siblings
         env:
           BRANCH: ${{ github.head_ref || github.ref_name }}
-        run: bash .github/resolve-sibling-refs.sh
+        run: bash .github/scripts/resolve-sibling-refs.sh
         working-directory: spinifex
 
       - name: Checkout Viperblock
@@ -229,22 +236,25 @@ jobs:
           key: gobuild-${{ github.job }}-${{ github.run_id }}
           restore-keys: gobuild-${{ github.job }}-
 
+      # Go refreshes an entry's mtime on use only when it is over an hour old, so
+      # restored entries are backdated past that before any Go command runs and
+      # the trim below can tell this run's working set from the last run's.
+      - name: Mark Go build cache
+        if: ${{ github.ref == 'refs/heads/dev' || github.ref == 'refs/heads/main' }}
+        run: spinifex/.github/scripts/go-build-cache.sh mark
+
       - name: Setup Go Workspace
         run: go work init ./spinifex ./viperblock ./predastore ./northstar ./bluebottle
 
       - name: Run Tests with Race Detector
         run: make -C spinifex test-race GOTESTFLAGS=-count=1
 
-      # Go bumps an entry's mtime when a run uses it, so this run's working set
-      # survives and objects superseded by later commits do not. Runs after the
-      # work, so it costs this job nothing and only shrinks what the next reads.
+      # Keeps what this run read or wrote and drops the rest, so the saved entry
+      # is one run's working set however often dev is pushed. A red run may have
+      # stopped early and used only part of the set, so it saves untrimmed.
       - name: Trim Go build cache
-        if: ${{ !cancelled() && (github.ref == 'refs/heads/dev' || github.ref == 'refs/heads/main') }}
-        run: |
-          C="$(go env GOCACHE)"
-          before=$(du -sm "$C" | cut -f1)
-          find "$C" -type f -mmin +240 -delete
-          echo "trimmed ${before}MB -> $(du -sm "$C" | cut -f1)MB"
+        if: ${{ success() && (github.ref == 'refs/heads/dev' || github.ref == 'refs/heads/main') }}
+        run: spinifex/.github/scripts/go-build-cache.sh trim
 
       - name: Save Go build cache
         if: ${{ !cancelled() && (github.ref == 'refs/heads/dev' || github.ref == 'refs/heads/main') }}
@@ -268,7 +278,7 @@ jobs:
         id: siblings
         env:
           BRANCH: ${{ github.head_ref || github.ref_name }}
-        run: bash .github/resolve-sibling-refs.sh
+        run: bash .github/scripts/resolve-sibling-refs.sh
         working-directory: spinifex
 
       - name: Checkout Viperblock
@@ -331,6 +341,13 @@ jobs:
           key: gobuild-${{ github.job }}-${{ github.run_id }}
           restore-keys: gobuild-${{ github.job }}-
 
+      # Go refreshes an entry's mtime on use only when it is over an hour old, so
+      # restored entries are backdated past that before any Go command runs and
+      # the trim below can tell this run's working set from the last run's.
+      - name: Mark Go build cache
+        if: ${{ github.ref == 'refs/heads/dev' || github.ref == 'refs/heads/main' }}
+        run: spinifex/.github/scripts/go-build-cache.sh mark
+
       - name: Setup Go Workspace
         run: go work init ./spinifex ./viperblock ./predastore ./northstar ./bluebottle
 
@@ -347,16 +364,16 @@ jobs:
         if: ${{ !cancelled() }}
         run: make -C spinifex test-integration
 
-      # Go bumps an entry's mtime when a run uses it, so this run's working set
-      # survives and objects superseded by later commits do not. Runs after the
-      # work, so it costs this job nothing and only shrinks what the next reads.
+      - name: Run Gated Suites
+        if: ${{ !cancelled() }}
+        run: make -C spinifex test-gated GOTESTFLAGS=-count=1
+
+      # Keeps what this run read or wrote and drops the rest, so the saved entry
+      # is one run's working set however often dev is pushed. A red run may have
+      # stopped early and used only part of the set, so it saves untrimmed.
       - name: Trim Go build cache
-        if: ${{ !cancelled() && (github.ref == 'refs/heads/dev' || github.ref == 'refs/heads/main') }}
-        run: |
-          C="$(go env GOCACHE)"
-          before=$(du -sm "$C" | cut -f1)
-          find "$C" -type f -mmin +240 -delete
-          echo "trimmed ${before}MB -> $(du -sm "$C" | cut -f1)MB"
+        if: ${{ success() && (github.ref == 'refs/heads/dev' || github.ref == 'refs/heads/main') }}
+        run: spinifex/.github/scripts/go-build-cache.sh trim
 
       - name: Save Go build cache
         if: ${{ !cancelled() && (github.ref == 'refs/heads/dev' || github.ref == 'refs/heads/main') }}

.github/workflows/ui.yml

@@ -30,6 +30,12 @@ on:
 permissions:
   contents: read
 
+# A newer push to the same pull request supersedes its run. Pushes and dispatches
+# get a group of their own, so every commit on dev and main keeps its result.
+concurrency:
+  group: ${{ github.workflow }}-${{ github.event_name == 'pull_request' && format('pr-{0}', github.event.pull_request.number) || github.run_id }}
+  cancel-in-progress: true
+
 jobs:
   ui_build_and_test:
     name: UI Build and Test

.gitignore

@@ -27,6 +27,7 @@ build/microvm/sqs-worker/.stage.*/
 build/microvm/lambda-runtime/vmlinuz
 build/microvm/lambda-runtime/initramfs.cpio.gz
 build/microvm/lambda-runtime/.stage.*/
+build/microvm/lambda-testlink-runtime/
 build/qboot/
 !spinifex/runtime/roles/spinifexui/frontend/dist/
 tmp/

Makefile

@@ -42,11 +42,15 @@ ifdef QUIET
   _COVQ  = 2>&1 | { grep -Ev '^\s*(ok|PASS|\?|=== RUN|--- PASS:)\s' | grep -v 'coverage: 0\.0%' || true; }; exit $${PIPESTATUS[0]}
   _RACEQ = 2>&1 | { grep -Ev '^\s*(ok|PASS|\?|=== RUN|--- PASS:)\s' || true; }; exit $${PIPESTATUS[0]}
   _SECQ  = >
+  _HDR   = :
+  _RUNSH = sh "$$t" >"$$log" 2>&1 || { echo "-- $$t"; cat "$$log"; exit 1; }
 else
   _Q     =
   _COVQ  =
   _RACEQ =
   _SECQ  = 2>&1 | tee
+  _HDR   = echo -e
+  _RUNSH = echo "-- $$t"; sh "$$t" || exit 1
 endif
 
 build: go_build build-installer build-lb-agent generate-aws-model-coverage
@@ -78,6 +82,36 @@ go_build:
 	@echo -e "\n....Building $(GO_PROJECT_NAME)"
 	go build $(GO_BUILD_MOD) -ldflags "$(LDFLAGS)" -o ./bin/$(GO_PROJECT_NAME) cmd/spinifex/main.go
 
+# Controlled-test only: builds the tagged daemon and its hard-kill initializer
+# together, pinning the initializer's digest into that daemon build. The daemon
+# keeps its symbols so the pin can be read back from the binary.
+# LAMBDA_IMAGE_DIGEST is the test-link runtime image digest that
+# build-lambda-testlink-image prints; it is read raw and validated before use.
+LAMBDA_TESTLINK_DIR := ./bin/lambda-testlink
+LAMBDA_TESTLINK_PKG := github.com/mulgadc/spinifex/spinifex/runtime/daemon
+LAMBDA_TESTLINK_PIN := $(LAMBDA_TESTLINK_PKG).lambdaHardKillSHA256
+LAMBDA_TESTLINK_GW := github.com/mulgadc/spinifex/spinifex/runtime/roles/awsgw
+LAMBDA_DIGEST_RE := ^[0-9a-f]{64}$$
+build-lambda-testlink: override export LAMBDA_IMAGE_DIGEST := $(value LAMBDA_IMAGE_DIGEST)
+build-lambda-testlink:
+	@[[ "$$LAMBDA_IMAGE_DIGEST" =~ $(LAMBDA_DIGEST_RE) ]] || { echo "LAMBDA_IMAGE_DIGEST must be 64 lowercase hex digits"; exit 1; }
+	@host=$$(uname -m); goos=$$(go env GOOS); goarch=$$(go env GOARCH); \
+	test "$$host" = x86_64 && test "$$goos/$$goarch" = linux/amd64 || { echo "the test link supports only an x86_64 host building linux/amd64, not $$host building $$goos/$$goarch"; exit 1; }
+	@rm -f $(LAMBDA_TESTLINK_DIR)/hardkill.so $(LAMBDA_TESTLINK_DIR)/$(GO_PROJECT_NAME)
+	@mkdir -p $(LAMBDA_TESTLINK_DIR).tmp
+	@rm -f $(LAMBDA_TESTLINK_DIR).tmp/hardkill.so $(LAMBDA_TESTLINK_DIR).tmp/$(GO_PROJECT_NAME)
+	cc -std=c17 -O2 -Wall -Wextra -Werror -shared -fPIC -o $(LAMBDA_TESTLINK_DIR).tmp/hardkill.so spinifex/runtime/compute/lambda/testlink/hardkill/hardkill.c -lrt
+	scripts/check-glibc-ceiling.sh 2.34 $(LAMBDA_TESTLINK_DIR).tmp/hardkill.so
+	@digest=$$(sha256sum $(LAMBDA_TESTLINK_DIR).tmp/hardkill.so | cut -d' ' -f1); \
+	test $${#digest} -eq 64 || { echo "initializer digest missing"; exit 1; }; \
+	echo "initializer sha256 $$digest"; \
+	pin="-X $(LAMBDA_TESTLINK_PIN)=$$digest -X $(LAMBDA_TESTLINK_PKG).lambdaImageDigest=$$LAMBDA_IMAGE_DIGEST -X $(LAMBDA_TESTLINK_GW).lambdaImageDigest=$$LAMBDA_IMAGE_DIGEST"; \
+	LAMBDA_TESTLINK_INITIALIZER=$(abspath $(LAMBDA_TESTLINK_DIR).tmp/hardkill.so) go test -count=1 -tags lambdatestlink -ldflags "$$pin" -run '^TestLambda(HardKillPin|ImageDigestPin)$$' ./spinifex/runtime/daemon/ ./spinifex/runtime/roles/awsgw/ && \
+	go build $(GO_BUILD_MOD) -tags lambdatestlink -ldflags "$(subst -s -w ,,$(LDFLAGS)) $$pin" -o $(LAMBDA_TESTLINK_DIR).tmp/$(GO_PROJECT_NAME) cmd/spinifex/main.go
+	@mkdir -p $(LAMBDA_TESTLINK_DIR)
+	mv $(LAMBDA_TESTLINK_DIR).tmp/hardkill.so $(LAMBDA_TESTLINK_DIR).tmp/$(GO_PROJECT_NAME) $(LAMBDA_TESTLINK_DIR)/
+	@rmdir $(LAMBDA_TESTLINK_DIR).tmp
+
 build-installer:
 	@echo -e "\n....Building spinifex-installer"
 	go build -ldflags "-s -w" -o ./bin/spinifex-installer cmd/installer/main.go
@@ -193,6 +227,24 @@ bin/lambda-agent: go.mod go.sum $$(call go_local_srcs,./cmd/lambda-agent)
 	GOWORK=off CGO_ENABLED=0 GOOS=linux GOARCH=amd64 go build -trimpath -ldflags '-s -w' -o $@ ./cmd/lambda-agent
 
 build-lambda-runtime-microvm-image: $(LAMBDA_RUNTIME_MICROVM_ARTIFACTS) ## Build the Lambda runtime microVM kernel and initramfs (not shipped)
+	GOWORK=off go run ./cmd/lambda-image-agent-tag $(LAMBDA_RUNTIME_MICROVM_OUT_DIR)/initramfs.cpio.gz absent
+
+# Controlled-test only: a separately built guest whose agent has the
+# lambdatestlink tag, in its own outputs, never touching bin/lambda-agent or the
+# default image. It prints the digest to pass as LAMBDA_IMAGE_DIGEST.
+LAMBDA_TESTLINK_IMAGE_DIR := $(MICROVM_OUT_DIR)/lambda-testlink-runtime
+build-lambda-testlink-image: ## Build the test-link Lambda guest image and print its digest (not shipped)
+	@host=$$(uname -m); goos=$$(go env GOOS); goarch=$$(go env GOARCH); \
+	test "$$host" = x86_64 && test "$$goos/$$goarch" = linux/amd64 || { echo "the test link supports only an x86_64 host building linux/amd64, not $$host building $$goos/$$goarch"; exit 1; }
+	$(MAKE) --no-print-directory $(MICROVM_OUT_DIR)/qboot.rom
+	@mkdir -p bin $(LAMBDA_TESTLINK_IMAGE_DIR)
+	GOWORK=off GOFLAGS= CGO_ENABLED=0 GOOS=linux GOARCH=amd64 go build -trimpath -tags lambdatestlink -ldflags '-s -w' -o bin/lambda-agent-testlink ./cmd/lambda-agent
+	./scripts/build-lambda-runtime-microvm-image.sh bin/lambda-agent-testlink $(LAMBDA_TESTLINK_IMAGE_DIR)
+	GOWORK=off go run ./cmd/lambda-image-agent-tag $(LAMBDA_TESTLINK_IMAGE_DIR)/initramfs.cpio.gz present
+	install -m 0644 $(MICROVM_OUT_DIR)/qboot.rom $(LAMBDA_TESTLINK_IMAGE_DIR)/qboot.rom
+	@digest=$$(GOWORK=off go run -tags lambdatestlink ./cmd/lambda-image-digest $(LAMBDA_TESTLINK_IMAGE_DIR)) && \
+	[[ "$$digest" =~ $(LAMBDA_DIGEST_RE) ]] || { echo "image digest not computed"; exit 1; }; \
+	echo "LAMBDA_IMAGE_DIGEST=$$digest"
 
 # qboot is the microvm firmware. The source tarball ships with each release
 # because qboot is GPL-2.0.
@@ -240,27 +292,30 @@ preflight:
 NODE_SCRIPTS := scripts/setup.sh scripts/setup-ovn.sh scripts/install-node.sh scripts/smoke-test.sh scripts/node-reset.sh scripts/uninstall-spx.sh scripts/lib/distro.sh scripts/lib/stage-sqs-store-microvm.sh
 
 test-build-scripts:
-	@echo -e "\n....Running build/scripts/**/*_test.sh...."
-	@for t in $$(find build/scripts -name '*_test.sh' | sort); do \
-		echo "-- $$t"; \
-		sh "$$t" || exit 1; \
+	@$(_HDR) "\n....Running build/scripts/**/*_test.sh...."
+	@log=$$(mktemp); trap 'rm -f "$$log"' EXIT; for t in $$(find build/scripts -name '*_test.sh' | sort); do \
+		$(_RUNSH); \
 	done
-	@echo -e "\n....Running scripts/*_test.sh, scripts/lib/*_test.sh and scripts/appliance/*_test.sh...."
-	@for t in $$( (find scripts -maxdepth 1 -name '*_test.sh'; find scripts/lib scripts/appliance -name '*_test.sh') | sort); do \
-		echo "-- $$t"; \
-		sh "$$t" || exit 1; \
+	@$(_HDR) "\n....Running scripts/*_test.sh, scripts/lib/*_test.sh and scripts/appliance/*_test.sh...."
+	@log=$$(mktemp); trap 'rm -f "$$log"' EXIT; for t in $$( (find scripts -maxdepth 1 -name '*_test.sh'; find scripts/lib scripts/appliance -name '*_test.sh') | sort); do \
+		$(_RUNSH); \
 	done
-	@echo -e "\n....Running .github/scripts/*_test.sh...."
-	@for t in $$(find .github/scripts -maxdepth 1 -name '*_test.sh' | sort); do \
-		echo "-- $$t"; \
-		sh "$$t" || exit 1; \
+	@$(_HDR) "\n....Running scripts/terraform/latitude-spx/*_test.sh...."
+	@log=$$(mktemp); trap 'rm -f "$$log"' EXIT; for t in $$(find scripts/terraform/latitude-spx -maxdepth 1 -name '*_test.sh' | sort); do \
+		$(_RUNSH); \
 	done
-	shellcheck -S warning .github/scripts/hypervisor-disk-preflight.sh .github/scripts/hypervisor-disk-preflight_test.sh \
-		.github/scripts/wait-for-memory.sh .github/scripts/wait-for-memory_test.sh
-	@echo -e "\n....Running shellcheck over build/scripts/**/*.sh + the node scripts...."
-	shellcheck -S warning $$(find build/scripts -name '*.sh' | sort) $(NODE_SCRIPTS)
-	shellcheck -S warning -x scripts/build-oracle-appliance.sh $$(find scripts/appliance -name '*.sh' | sort)
-	@echo "  test-build-scripts ok"
+	@$(_HDR) "\n....Running .github/scripts/*_test.sh...."
+	@log=$$(mktemp); trap 'rm -f "$$log"' EXIT; for t in $$(find .github/scripts -maxdepth 1 -name '*_test.sh' | sort); do \
+		$(_RUNSH); \
+	done
+	$(_Q)shellcheck -S warning .github/scripts/hypervisor-disk-preflight.sh \
+		.github/scripts/wait-for-memory.sh \
+		.github/scripts/go-build-cache.sh \
+		.github/scripts/e2e-select.sh
+	@$(_HDR) "\n....Running shellcheck over build/scripts/**/*.sh + the node scripts...."
+	$(_Q)shellcheck -S warning $$(find build/scripts -name '*.sh' | sort) $(NODE_SCRIPTS)
+	$(_Q)shellcheck -S warning -x scripts/build-oracle-appliance.sh $$(find scripts/appliance -name '*.sh' | sort)
+	@$(_HDR) "  test-build-scripts ok"
 
 # E2E harness unit tests. Build-tagged `e2e` so they're skipped by the
 # default `go test ./spinifex/...`. Runs with mocked AWS clients — no
@@ -291,6 +346,16 @@ test-integration:
 	@$(MAKE) --no-print-directory generate-aws-model-coverage
 	@echo "AWS operation coverage: $(AWS_MODEL_OPERATION_COVERAGE_DIR)"
 
+# Suites the unit tiers skip: the three-node SQS R3 scenarios behind
+# SPINIFEX_SQS_R3 and the lambdatestlink-tagged tests. CI runs these on every PR.
+test-gated:
+	@echo -e "\n....Running env- and tag-gated suites...."
+	$(_Q)LOG_IGNORE=1 SPINIFEX_SQS_R3=1 go test $(GOTESTFLAGS) -timeout 180s -run '^TestR3$$' ./spinifex/domains/sqs/message/jetstream/ $(_RACEQ)
+	$(_Q)LOG_IGNORE=1 go test $(GOTESTFLAGS) -tags=lambdatestlink -timeout 120s ./cmd/lambda-agent/ ./spinifex/agents/lambda/hostlink/ ./spinifex/runtime/compute/lambda/guestconn/ ./spinifex/runtime/compute/lambda/testlink/ ./spinifex/domains/network/host/ $(_RACEQ)
+	$(_Q)LOG_IGNORE=1 go test $(GOTESTFLAGS) -tags=lambdatestlink -timeout 120s -run '^TestLambdaTestLink(Executor|Reply)_' ./spinifex/runtime/roles/awsgw/ $(_RACEQ)
+	$(_Q)LOG_IGNORE=1 go test $(GOTESTFLAGS) -tags=lambdatestlink,lambdatestlinkseam -timeout 120s -run '^TestLambdaTestLinkSeam_' ./spinifex/runtime/roles/awsgw/ $(_RACEQ)
+	$(_Q)LOG_IGNORE=1 go test $(GOTESTFLAGS) -tags=lambdatestlink -timeout 120s -run '^TestLambdaTestLink' ./spinifex/runtime/daemon/ $(_RACEQ)
+
 generate-aws-model-coverage:
 	@go run ./cmd/aws-model-coverage -out $(AWS_MODEL_OPERATION_COVERAGE_DIR) -json $(AWS_MODEL_OPERATION_COVERAGE_DIR)/coverage.json
 
@@ -326,8 +391,7 @@ test-cover:
 # Run unit tests with race detector
 test-race:
 	@echo -e "\n....Running tests with race detector for $(GO_PROJECT_NAME)...."
-	$(_Q)LOG_IGNORE=1 SPINIFEX_SQS_R3=1 go test $(GOTESTFLAGS) -race -timeout 300s ./spinifex/... ./cmd/... ./internal/... ./contracts/... ./architecture/... $(_RACEQ)
-	$(_Q)LOG_IGNORE=1 go test $(GOTESTFLAGS) -race -tags=lambdatestlink -timeout 120s ./cmd/lambda-agent/ ./spinifex/agents/lambda/hostlink/ $(_RACEQ)
+	$(_Q)LOG_IGNORE=1 go test $(GOTESTFLAGS) -race -timeout 300s ./spinifex/... ./cmd/... ./internal/... ./contracts/... ./architecture/... $(_RACEQ)
 
 # Unit tests for in-repo GitHub Actions (e.g. .github/actions/e2e-analyze).
 # Kept out of `test-cover` so coverage % isn't diluted by CI-only tooling.
@@ -538,7 +602,7 @@ distro-arm64:
 distro-clean:
 	rm -rf dist/
 
-.PHONY: test-package-check build build-ui test-ui-e2e ui-e2e-container ui-baselines build-installer build-lb-agent build-ecs-agent build-system-image build-eks-node-image import-eks-node-image publish-eks-node-image build-ecs-node-image import-ecs-node-image build-rds-postgres-image import-rds-postgres-image build-rds-mariadb-image import-rds-mariadb-image build-microvm-image build-sqs-worker-microvm-image build-lambda-runtime-microvm-image install-microvm install-sqs-worker-microvm go_build preflight test test-cover test-race diff-coverage bench fuzz test-actions test-images test-build-scripts test-harness test-integration generate-aws-model-coverage aws-model-coverage test-segscan-oracle \
+.PHONY: test-package-check build build-lambda-testlink build-lambda-testlink-image build-ui test-ui-e2e ui-e2e-container ui-baselines build-installer build-lb-agent build-ecs-agent build-system-image build-eks-node-image import-eks-node-image publish-eks-node-image build-ecs-node-image import-ecs-node-image build-rds-postgres-image import-rds-postgres-image build-rds-mariadb-image import-rds-mariadb-image build-microvm-image build-sqs-worker-microvm-image build-lambda-runtime-microvm-image install-microvm install-sqs-worker-microvm go_build preflight test test-cover test-race diff-coverage bench fuzz test-actions test-images test-build-scripts test-harness test-integration test-gated generate-aws-model-coverage aws-model-coverage test-segscan-oracle \
 	deploy reinstall clean \
 	install-system install-go install-aws quickinstall \
 	lint fix govulncheck nilaway \

README.md

@@ -70,8 +70,8 @@ The platform serves **405 AWS API operations** across the services below. See ex
 | **RDS** | Databases | PostgreSQL and MariaDB, with automated backups | Available |
 | **Route 53** | DNS | [Northstar](https://github.com/mulgadc/northstar): authoritative DNS with hosted zones | Available |
 | **Bedrock** | AI deployment | vLLM for text generation and Hugging Face TEI for embeddings and reranking | Q4 2026 |
-| **Lambda** | Serverless functions | | Q1 2027 |
-| **SQS** | Message queues | | Q1 2027 |
+| **Lambda** | Serverless functions | QEMU `microvm` execution environments | Q1 2027 |
+| **SQS** | Message queues | A dedicated NATS JetStream store in a per-Region QEMU `microvm` guest | Q1 2027 |
 
 Roadmap services ship under the same AWS API surface, so code written for AWS today works the moment they land. [Track what's shipped in the release notes](https://github.com/mulgadc/spinifex/releases).
 

architecture/layering/allowlist_test.go

@@ -15,9 +15,6 @@ var allowlist = []allowance{
 	{From: "operator/accountlifecycle", To: "domains/ec2/awsapi/igw", Debt: "accountteardown (dependency)", Reason: "EC2 reapers call the awsapi action adapters; needs domain-owned cleanup capabilities"},
 	{From: "operator/accountlifecycle", To: "domains/ec2/awsapi/image", Debt: "accountteardown (dependency)", Reason: "EC2 reapers call the awsapi action adapters; needs domain-owned cleanup capabilities"},
 	{From: "operator/accountlifecycle", To: "domains/ec2/awsapi/instance", Debt: "accountteardown (dependency)", Reason: "EC2 reapers call the awsapi action adapters; needs domain-owned cleanup capabilities"},
-	{From: "operator/accountlifecycle", To: "domains/ec2/awsapi/key", Debt: "accountteardown (dependency)", Reason: "EC2 reapers call the awsapi action adapters; needs domain-owned cleanup capabilities"},
-	{From: "operator/accountlifecycle", To: "domains/ec2/awsapi/launchtemplate", Debt: "accountteardown (dependency)", Reason: "EC2 reapers call the awsapi action adapters; needs domain-owned cleanup capabilities"},
-	{From: "operator/accountlifecycle", To: "domains/ec2/awsapi/placementgroup", Debt: "accountteardown (dependency)", Reason: "EC2 reapers call the awsapi action adapters; needs domain-owned cleanup capabilities"},
 	{From: "operator/accountlifecycle", To: "domains/ec2/awsapi/routetable", Debt: "accountteardown (dependency)", Reason: "EC2 reapers call the awsapi action adapters; needs domain-owned cleanup capabilities"},
 	{From: "operator/accountlifecycle", To: "domains/ec2/awsapi/snapshot", Debt: "accountteardown (dependency)", Reason: "EC2 reapers call the awsapi action adapters; needs domain-owned cleanup capabilities"},
 	{From: "operator/accountlifecycle", To: "domains/ec2/awsapi/volume", Debt: "accountteardown (dependency)", Reason: "EC2 reapers call the awsapi action adapters; needs domain-owned cleanup capabilities"},

cmd/lambda-agent/connect.go

@@ -6,6 +6,7 @@ import (
 	"errors"
 	"fmt"
 
+	lambdav1 "github.com/mulgadc/spinifex/contracts/lambda/v1"
 	"github.com/mulgadc/spinifex/spinifex/agents/lambda/hostlink"
 )
 
@@ -29,20 +30,20 @@ func defaultStartup() startup {
 
 // connect loads the identity, seals fw_cfg whether or not the load succeeded,
 // then dials the host. Under TLS 1.3 a host refusal is seen only on first read.
-func (s startup) connect(ctx context.Context) (*tls.Conn, int, error) {
+func (s startup) connect(ctx context.Context) (*tls.Conn, lambdav1.Scope, int, error) {
 	id, loadErr := hostlink.LoadIdentity(s.fwCfgDir)
 	if err := hostlink.SealFwCfg(s.fwCfgRoot, s.unload); err != nil {
-		return nil, exitRefused, errors.Join(loadErr, err)
+		return nil, "", exitRefused, errors.Join(loadErr, err)
 	}
 	if loadErr != nil {
-		return nil, exitRefused, loadErr
+		return nil, "", exitRefused, loadErr
 	}
 	conn, err := hostlink.Dial(ctx, id)
 	if err != nil {
 		if ctx.Err() != nil {
 			err = fmt.Errorf("%w: %w", err, context.Cause(ctx))
 		}
-		return nil, exitLink, err
+		return nil, "", exitLink, err
 	}
-	return conn, 0, nil
+	return conn, id.Scope, 0, nil
 }

cmd/lambda-agent/grant_testlink.go

@@ -0,0 +1,66 @@
+//go:build lambdatestlink
+
+package main
+
+import (
+	"errors"
+	"fmt"
+	"os"
+	"path/filepath"
+	"strconv"
+	"strings"
+	"time"
+
+	lambdav1 "github.com/mulgadc/spinifex/contracts/lambda/v1"
+	"github.com/mulgadc/spinifex/spinifex/agents/lambda/hostlink"
+)
+
+const entryGrant = "test-grant"
+
+// testGrant is the grant the host launched this VM under.
+type testGrant struct {
+	grantID, attemptID, environmentID string
+	generation                        uint64
+}
+
+// readGrant reads the test-grant fw_cfg entry: grant ID, attempt ID,
+// environment ID and generation, one per line.
+func readGrant(fwCfgDir string) (testGrant, error) {
+	b, err := os.ReadFile(filepath.Join(fwCfgDir, entryGrant, "raw"))
+	if err != nil {
+		return testGrant{}, fmt.Errorf("read %s: %w", entryGrant, err)
+	}
+	lines := strings.Split(strings.TrimSuffix(string(b), "\n"), "\n")
+	if len(lines) != 4 {
+		return testGrant{}, errors.New(entryGrant + " entry malformed")
+	}
+	g := testGrant{grantID: lines[0], attemptID: lines[1], environmentID: lines[2]}
+	g.generation, err = strconv.ParseUint(lines[3], 10, 64)
+	if err != nil || g.generation == 0 || g.grantID == "" || g.attemptID == "" || g.environmentID == "" {
+		return testGrant{}, errors.New(entryGrant + " entry malformed")
+	}
+	return g, nil
+}
+
+// checkBinding refuses a run unless the grant names the authenticated environment
+// generation, the request carries that grant and its deadline is still ahead. The
+// clock check is a local guard only: it cannot detect skew or a superseded grant.
+func checkBinding(scope lambdav1.Scope, g testGrant, req hostlink.Request, now time.Time) error {
+	parts := strings.Split(string(scope), "/")
+	if len(parts) != 5 {
+		return errors.New("authenticated scope malformed")
+	}
+	if parts[3] != g.environmentID || parts[4] != strconv.FormatUint(g.generation, 10) {
+		return errors.New("grant is not for the authenticated environment generation")
+	}
+	if req.GrantID != g.grantID || req.AttemptID != g.attemptID {
+		return errors.New("request does not carry this VM's grant")
+	}
+	if req.DeadlineMs <= 0 {
+		return errors.New("request has no deadline")
+	}
+	if !now.Before(time.UnixMilli(req.DeadlineMs)) {
+		return errors.New("request deadline already passed")
+	}
+	return nil
+}

cmd/lambda-agent/helpers_test.go

@@ -14,7 +14,7 @@ import (
 	"github.com/stretchr/testify/require"
 )
 
-const scope lambdav1.Scope = "ap-southeast-2/000000000001/fn-1/env-1/gen-1"
+const scope lambdav1.Scope = "ap-southeast-2/000000000001/fn-1/env-1/1"
 
 // guest is a fake fw_cfg tree; unload removes it as unloading the module would.
 type guest struct {

cmd/lambda-agent/package_testlink.go

@@ -0,0 +1,286 @@
+//go:build lambdatestlink
+
+package main
+
+import (
+	"archive/zip"
+	"bytes"
+	"crypto/sha256"
+	"encoding/hex"
+	"errors"
+	"fmt"
+	"io"
+	"io/fs"
+	"os"
+	"os/exec"
+	"path"
+	"path/filepath"
+	"strconv"
+	"strings"
+	"time"
+	"unicode"
+
+	"golang.org/x/sys/unix"
+)
+
+// Dev-only code delivery for the test path: the host attaches the ZIP as a
+// read-only disk and passes its exact size and SHA-256 through fw_cfg. This
+// is not the production delivery, and these bounds are not AWS limits.
+const (
+	entryPackageSize   = "test-package-size"
+	entryPackageSHA256 = "test-package-sha256"
+	entryHandler       = "test-handler"
+
+	devMaxZipBytes    = 16 << 20
+	devMaxUnzipBytes  = 64 << 20
+	devMaxEntries     = 1024
+	packageDevice     = "/dev/vda"
+	taskTmpfsSizeOpts = "mode=0755,size=72m"
+)
+
+type packageSpec struct {
+	size    int64
+	sha256  [sha256.Size]byte
+	handler string
+}
+
+type unpackLimits struct {
+	entries int
+	bytes   int64
+}
+
+var devLimits = unpackLimits{entries: devMaxEntries, bytes: devMaxUnzipBytes}
+
+// readPackageSpec reads the expected size and SHA-256 from fw_cfg; it must
+// run before fw_cfg is sealed.
+func readPackageSpec(fwCfgDir string) (packageSpec, error) {
+	var spec packageSpec
+	read := func(name string) (string, error) {
+		b, err := os.ReadFile(filepath.Join(fwCfgDir, name, "raw"))
+		if err != nil {
+			return "", fmt.Errorf("read %s: %w", name, err)
+		}
+		return strings.TrimSpace(string(b)), nil
+	}
+	s, err := read(entryPackageSize)
+	if err != nil {
+		return spec, err
+	}
+	spec.size, err = strconv.ParseInt(s, 10, 64)
+	if err != nil || spec.size <= 0 || spec.size > devMaxZipBytes {
+		return spec, fmt.Errorf("package size %q outside 1..%d bytes", s, devMaxZipBytes)
+	}
+	h, err := read(entryPackageSHA256)
+	if err != nil {
+		return spec, err
+	}
+	b, err := hex.DecodeString(h)
+	if err != nil || len(b) != sha256.Size {
+		return spec, errors.New("package sha256 is not 64 hex digits")
+	}
+	copy(spec.sha256[:], b)
+	if spec.handler, err = read(entryHandler); err != nil {
+		return spec, err
+	}
+	if spec.handler == "" || len(spec.handler) > 128 || strings.IndexFunc(spec.handler, unicode.IsSpace) >= 0 ||
+		strings.ContainsRune(spec.handler, 0) {
+		return spec, errors.New("handler must be 1..128 characters with no whitespace")
+	}
+	return spec, nil
+}
+
+// taskPackage puts the package under dir: a fresh nosuid,nodev tmpfs, made
+// read-only once the files are in place.
+type taskPackage struct {
+	device     string
+	dir        string
+	loadDriver func() error
+	mountDir   func(dir string) error
+	sealDir    func(dir string) error
+}
+
+func devTaskPackage() taskPackage {
+	return taskPackage{device: packageDevice, dir: taskRoot, loadDriver: loadBlockDriver, mountDir: mountTaskDir, sealDir: sealTaskDir}
+}
+
+func (p taskPackage) install(spec packageSpec) error {
+	if err := p.loadDriver(); err != nil {
+		return fmt.Errorf("load block driver: %w", err)
+	}
+	dev, err := os.Open(p.device)
+	if err != nil {
+		return fmt.Errorf("open package disk: %w", err)
+	}
+	defer dev.Close()
+	if err := p.mountDir(p.dir); err != nil {
+		return fmt.Errorf("mount %s: %w", p.dir, err)
+	}
+	if err := unpack(dev, spec, p.dir, devLimits); err != nil {
+		return err
+	}
+	if err := p.sealDir(p.dir); err != nil {
+		return fmt.Errorf("make %s read-only: %w", p.dir, err)
+	}
+	return nil
+}
+
+// unpack reads exactly spec.size bytes once, checks their SHA-256, then
+// validates every entry before writing any of them under dir.
+func unpack(r io.ReaderAt, spec packageSpec, dir string, lim unpackLimits) error {
+	buf := make([]byte, spec.size)
+	if _, err := io.ReadFull(io.NewSectionReader(r, 0, spec.size), buf); err != nil {
+		return fmt.Errorf("read package: %w", err)
+	}
+	if sha256.Sum256(buf) != spec.sha256 {
+		return errors.New("package sha256 mismatch")
+	}
+	zr, err := zip.NewReader(bytes.NewReader(buf), spec.size)
+	if err != nil {
+		return fmt.Errorf("open package: %w", err)
+	}
+	if len(zr.File) > lim.entries {
+		return fmt.Errorf("package has %d entries, limit %d", len(zr.File), lim.entries)
+	}
+	names := make([]string, len(zr.File))
+	var declared int64
+	for i, f := range zr.File {
+		name, err := entryName(f.Name)
+		if err != nil {
+			return err
+		}
+		names[i] = name
+		switch mode := f.Mode(); {
+		case mode.IsDir():
+		case mode.IsRegular():
+			// Compared against what remains, so a huge size cannot wrap the sum.
+			if f.UncompressedSize64 > uint64(lim.bytes-declared) {
+				return fmt.Errorf("%q: unpacked size over the %d-byte limit", f.Name, lim.bytes)
+			}
+			declared += int64(f.UncompressedSize64)
+		default:
+			return fmt.Errorf("%q: %s entries are refused", f.Name, mode.Type())
+		}
+	}
+
+	root, err := os.OpenRoot(dir)
+	if err != nil {
+		return err
+	}
+	defer root.Close()
+	var written int64
+	for i, f := range zr.File {
+		n, err := extract(root, names[i], f, lim.bytes-written)
+		written += n
+		if err != nil {
+			return fmt.Errorf("%q: %w", f.Name, err)
+		}
+	}
+	return nil
+}
+
+// extract writes one entry, refusing it if its actual bytes exceed budget
+// whatever its header claims.
+func extract(root *os.Root, name string, f *zip.File, budget int64) (int64, error) {
+	if f.Mode().IsDir() {
+		return 0, mkdirs(root, name)
+	}
+	if d := path.Dir(name); d != "." {
+		if err := mkdirs(root, d); err != nil {
+			return 0, err
+		}
+	}
+	perm := os.FileMode(0o644)
+	if f.Mode()&0o111 != 0 {
+		perm = 0o755
+	}
+	in, err := f.Open()
+	if err != nil {
+		return 0, err
+	}
+	defer in.Close()
+	out, err := root.OpenFile(name, os.O_CREATE|os.O_EXCL|os.O_WRONLY, perm)
+	if err != nil {
+		return 0, err
+	}
+	// The umask applies at create; fchmod does not, so the mode is exact.
+	if err := out.Chmod(perm); err != nil {
+		out.Close()
+		return 0, err
+	}
+	n, err := io.Copy(out, io.LimitReader(in, budget+1))
+	if cerr := out.Close(); err == nil {
+		err = cerr
+	}
+	if err == nil && n > budget {
+		err = errors.New("unpacked bytes over the limit")
+	}
+	return n, err
+}
+
+// mkdirs creates each component of dir as mode 0755 whatever the umask. An
+// existing component must be a real directory, never a symlink.
+func mkdirs(root *os.Root, dir string) error {
+	cur := ""
+	for _, part := range strings.Split(dir, "/") {
+		cur = path.Join(cur, part)
+		if err := root.Mkdir(cur, 0o755); err != nil {
+			if !errors.Is(err, fs.ErrExist) {
+				return err
+			}
+			fi, err := root.Lstat(cur)
+			if err != nil {
+				return err
+			}
+			if !fi.IsDir() {
+				return fmt.Errorf("%q exists and is not a directory", cur)
+			}
+		}
+		if err := root.Chmod(cur, 0o755); err != nil {
+			return err
+		}
+	}
+	return nil
+}
+
+// entryName returns a clean relative path, refusing anything that could name
+// a location outside the task root.
+func entryName(name string) (string, error) {
+	if name == "" || strings.ContainsAny(name, "\\\x00") || path.IsAbs(name) {
+		return "", fmt.Errorf("%q: entry name refused", name)
+	}
+	c := path.Clean(name)
+	if c == "." || c == ".." || strings.HasPrefix(c, "../") {
+		return "", fmt.Errorf("%q: entry name refused", name)
+	}
+	return c, nil
+}
+
+// loadBlockDriver loads virtio_blk, which the image ships but init does not
+// load, then waits briefly for the package disk to appear.
+func loadBlockDriver() error {
+	if err := exec.Command("/sbin/modprobe", "virtio_blk").Run(); err != nil {
+		var uts unix.Utsname
+		if uerr := unix.Uname(&uts); uerr != nil {
+			return uerr
+		}
+		ko := "/lib/modules/" + unix.ByteSliceToString(uts.Release[:]) + "/kernel/drivers/block/virtio_blk.ko"
+		if err := exec.Command("/sbin/insmod", ko).Run(); err != nil {
+			return fmt.Errorf("modprobe and insmod virtio_blk failed: %w", err)
+		}
+	}
+	for range 50 {
+		if _, err := os.Stat(packageDevice); err == nil {
+			return nil
+		}
+		time.Sleep(20 * time.Millisecond)
+	}
+	return fmt.Errorf("%s did not appear", packageDevice)
+}
+
+func mountTaskDir(dir string) error {
+	return unix.Mount("tmpfs", dir, "tmpfs", unix.MS_NOSUID|unix.MS_NODEV, taskTmpfsSizeOpts)
+}
+
+func sealTaskDir(dir string) error {
+	return unix.Mount("", dir, "", unix.MS_REMOUNT|unix.MS_RDONLY|unix.MS_NOSUID|unix.MS_NODEV, "")
+}

cmd/lambda-agent/package_testlink_test.go

@@ -0,0 +1,215 @@
+//go:build lambdatestlink
+
+package main
+
+import (
+	"archive/zip"
+	"bytes"
+	"crypto/sha256"
+	"hash/crc32"
+	"io/fs"
+	"math"
+	"os"
+	"path/filepath"
+	"sort"
+	"strings"
+	"syscall"
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+// testZip builds a ZIP; an entry named bootstrap is marked executable.
+func testZip(t *testing.T, files map[string]string) []byte {
+	t.Helper()
+	var buf bytes.Buffer
+	zw := zip.NewWriter(&buf)
+	names := make([]string, 0, len(files))
+	for n := range files {
+		names = append(names, n)
+	}
+	sort.Strings(names)
+	for _, n := range names {
+		h := &zip.FileHeader{Name: n, Method: zip.Deflate}
+		h.SetMode(0o644)
+		if n == "bootstrap" {
+			h.SetMode(0o755)
+		}
+		w, err := zw.CreateHeader(h)
+		require.NoError(t, err)
+		_, err = w.Write([]byte(files[n]))
+		require.NoError(t, err)
+	}
+	require.NoError(t, zw.Close())
+	return buf.Bytes()
+}
+
+// rawZip writes entries with headers set by the caller, so a test can claim
+// sizes or modes the data does not have.
+func rawZip(t *testing.T, entries ...*zip.FileHeader) []byte {
+	t.Helper()
+	var buf bytes.Buffer
+	zw := zip.NewWriter(&buf)
+	for _, h := range entries {
+		w, err := zw.CreateRaw(h)
+		require.NoError(t, err)
+		_, err = w.Write([]byte("x"))
+		require.NoError(t, err)
+	}
+	require.NoError(t, zw.Close())
+	return buf.Bytes()
+}
+
+func specFor(b []byte) packageSpec {
+	return packageSpec{size: int64(len(b)), sha256: sha256.Sum256(b)}
+}
+
+// Not parallel: the umask is process-wide, so the test sets a restrictive one
+// and restores it.
+func TestUnpackInstallsFilesWithFixedModes(t *testing.T) {
+	old := syscall.Umask(0o077)
+	t.Cleanup(func() { syscall.Umask(old) })
+	b := testZip(t, map[string]string{"bootstrap": "#!/bin/sh\n", "lib/data.txt": "data"})
+	dir := t.TempDir()
+	require.NoError(t, unpack(bytes.NewReader(b), specFor(b), dir, devLimits))
+
+	pfi, err := os.Stat(filepath.Join(dir, "lib"))
+	require.NoError(t, err)
+	assert.Equal(t, fs.FileMode(0o755), pfi.Mode().Perm(), "implicit parent dir")
+
+	fi, err := os.Stat(filepath.Join(dir, "bootstrap"))
+	require.NoError(t, err)
+	assert.Equal(t, fs.FileMode(0o755), fi.Mode().Perm(), "executable")
+	fi, err = os.Stat(filepath.Join(dir, "lib", "data.txt"))
+	require.NoError(t, err)
+	assert.Equal(t, fs.FileMode(0o644), fi.Mode().Perm())
+	got, err := os.ReadFile(filepath.Join(dir, "lib", "data.txt"))
+	require.NoError(t, err)
+	assert.Equal(t, "data", string(got))
+}
+
+func TestUnpackRefusesBeforeWritingAnything(t *testing.T) {
+	symlink := &zip.FileHeader{Name: "link", Method: zip.Store, CRC32: crc32.ChecksumIEEE([]byte("x")), CompressedSize64: 1, UncompressedSize64: 1}
+	symlink.SetMode(fs.ModeSymlink | 0o777)
+	fifo := &zip.FileHeader{Name: "pipe", Method: zip.Store, CRC32: crc32.ChecksumIEEE([]byte("x")), CompressedSize64: 1, UncompressedSize64: 1}
+	fifo.SetMode(fs.ModeNamedPipe | 0o644)
+	// The second size would wrap a plain running sum back under the limit.
+	big := func(name string, n uint64) *zip.FileHeader {
+		return &zip.FileHeader{Name: name, Method: zip.Store, CRC32: crc32.ChecksumIEEE([]byte("x")), CompressedSize64: 1, UncompressedSize64: n}
+	}
+	cases := map[string]struct {
+		zip  []byte
+		want string
+	}{
+		"parent traversal": {testZip(t, map[string]string{"a": "ok", "../escape": "x"}), "entry name refused"},
+		"nested traversal": {testZip(t, map[string]string{"a": "ok", "lib/../../escape": "x"}), "entry name refused"},
+		"absolute path":    {testZip(t, map[string]string{"a": "ok", "/etc/escape": "x"}), "entry name refused"},
+		"backslash":        {testZip(t, map[string]string{"a": "ok", `..\escape`: "x"}), "entry name refused"},
+		"symlink":          {rawZip(t, big("a", 1), symlink), "entries are refused"},
+		"named pipe":       {rawZip(t, big("a", 1), fifo), "entries are refused"},
+		"declared size":    {rawZip(t, big("a", devMaxUnzipBytes+1)), "over the"},
+		"wrapping sizes":   {rawZip(t, big("a", devMaxUnzipBytes), big("b", math.MaxUint64)), "over the"},
+	}
+	for name, tc := range cases {
+		t.Run(name, func(t *testing.T) {
+			dir := t.TempDir()
+			err := unpack(bytes.NewReader(tc.zip), specFor(tc.zip), dir, devLimits)
+			assert.ErrorContains(t, err, tc.want)
+			left, rerr := os.ReadDir(dir)
+			require.NoError(t, rerr)
+			assert.Empty(t, left, "files written before the refusal")
+		})
+	}
+}
+
+func TestUnpackRefusesTooManyEntries(t *testing.T) {
+	b := testZip(t, map[string]string{"a": "1", "b": "2", "c": "3"})
+	err := unpack(bytes.NewReader(b), specFor(b), t.TempDir(), unpackLimits{entries: 2, bytes: devMaxUnzipBytes})
+	assert.ErrorContains(t, err, "limit 2")
+}
+
+// The header claims a size within the limit, but the data is longer: the
+// actual bytes are refused, whatever the header says.
+func TestUnpackRefusesBytesBeyondTheLimit(t *testing.T) {
+	data := bytes.Repeat([]byte("a"), 64)
+	var buf bytes.Buffer
+	zw := zip.NewWriter(&buf)
+	h := &zip.FileHeader{Name: "a", Method: zip.Store, CRC32: crc32.ChecksumIEEE(data), CompressedSize64: uint64(len(data)), UncompressedSize64: 8}
+	w, err := zw.CreateRaw(h)
+	require.NoError(t, err)
+	_, err = w.Write(data)
+	require.NoError(t, err)
+	require.NoError(t, zw.Close())
+	b := buf.Bytes()
+
+	dir := t.TempDir()
+	err = unpack(bytes.NewReader(b), specFor(b), dir, unpackLimits{entries: 4, bytes: 16})
+	require.Error(t, err)
+	fi, serr := os.Stat(filepath.Join(dir, "a"))
+	if serr == nil {
+		assert.LessOrEqual(t, fi.Size(), int64(17))
+	}
+}
+
+func TestUnpackChecksSizeAndHash(t *testing.T) {
+	b := testZip(t, map[string]string{"bootstrap": "x"})
+	wrongHash := specFor(b)
+	wrongHash.sha256[0] ^= 1
+	assert.ErrorContains(t, unpack(bytes.NewReader(b), wrongHash, t.TempDir(), devLimits), "sha256 mismatch")
+
+	longer := specFor(b)
+	longer.size++
+	assert.ErrorContains(t, unpack(bytes.NewReader(b), longer, t.TempDir(), devLimits), "read package")
+}
+
+func TestReadPackageSpecBounds(t *testing.T) {
+	sum := "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855"
+	cases := map[string]struct {
+		size, hash, handler string
+		want                string
+	}{
+		"valid":            {"10", sum, "bootstrap.h", ""},
+		"zero":             {"0", sum, "bootstrap.h", "outside"},
+		"too large":        {"16777217", sum, "bootstrap.h", "outside"},
+		"not hex":          {"10", "zz", "bootstrap.h", "64 hex digits"},
+		"short":            {"10", sum[:62], "bootstrap.h", "64 hex digits"},
+		"empty handler":    {"10", sum, "", "handler"},
+		"handler space":    {"10", sum, "a b", "handler"},
+		"128-char handler": {"10", sum, strings.Repeat("h", 128), ""},
+		"129-char handler": {"10", sum, strings.Repeat("h", 129), "handler"},
+	}
+	for name, tc := range cases {
+		t.Run(name, func(t *testing.T) {
+			dir := t.TempDir()
+			for n, v := range map[string]string{entryPackageSize: tc.size, entryPackageSHA256: tc.hash, entryHandler: tc.handler} {
+				require.NoError(t, os.MkdirAll(filepath.Join(dir, n), 0o700))
+				require.NoError(t, os.WriteFile(filepath.Join(dir, n, "raw"), []byte(v), 0o600))
+			}
+			spec, err := readPackageSpec(dir)
+			if tc.want == "" {
+				require.NoError(t, err)
+				assert.Equal(t, int64(10), spec.size)
+				assert.Equal(t, tc.handler, spec.handler)
+				return
+			}
+			assert.ErrorContains(t, err, tc.want)
+		})
+	}
+}
+
+// A pre-existing symlink where a directory belongs is refused, not followed
+// or chmodded.
+func TestMkdirsRefusesSymlinkComponent(t *testing.T) {
+	dir := t.TempDir()
+	outside := t.TempDir()
+	require.NoError(t, os.Chmod(outside, 0o700))
+	require.NoError(t, os.Symlink(outside, filepath.Join(dir, "lib")))
+	root, err := os.OpenRoot(dir)
+	require.NoError(t, err)
+	defer root.Close()
+	assert.ErrorContains(t, mkdirs(root, "lib/sub"), "not a directory")
+	fi, err := os.Stat(outside)
+	require.NoError(t, err)
+	assert.Equal(t, fs.FileMode(0o700), fi.Mode().Perm(), "symlink target was chmodded")
+}

cmd/lambda-agent/start_tcp.go

@@ -25,7 +25,7 @@ func start(ctx context.Context) int {
 
 // run authenticates to the host and holds the connection until it closes.
 func (s startup) run(ctx context.Context) (int, error) {
-	conn, code, err := s.connect(ctx)
+	conn, _, code, err := s.connect(ctx)
 	if err != nil {
 		return code, err
 	}

cmd/lambda-agent/start_testlink.go

@@ -5,8 +5,8 @@ package main
 import (
 	"context"
 	"errors"
-	"flag"
 	"fmt"
+	"io"
 	"log/slog"
 	"net"
 	"net/http"
@@ -24,12 +24,11 @@ const (
 	taskRoot    = "/var/task"
 )
 
-var handler = flag.String("handler", os.Getenv("_HANDLER"), "function handler passed to the bootstrap")
-
 // testlink relays one invocation over the test carrier on the authenticated
 // connection; the carrier is not the guest protocol.
 type testlink struct {
 	startup
+	pkg            taskPackage
 	runtimeAddr    string
 	startBootstrap func(ctx context.Context, env []string) (*exec.Cmd, error)
 }
@@ -37,6 +36,7 @@ type testlink struct {
 func start(ctx context.Context) int {
 	t := testlink{
 		startup:     defaultStartup(),
+		pkg:         devTaskPackage(),
 		runtimeAddr: runtimeAddr,
 		startBootstrap: func(ctx context.Context, env []string) (*exec.Cmd, error) {
 			return agent.StartBootstrap(ctx, taskRoot+"/bootstrap", env, os.Stdout, os.Stderr)
@@ -50,11 +50,17 @@ func start(ctx context.Context) int {
 }
 
 func (t testlink) run(parent context.Context) (int, error) {
-	conn, code, err := t.connect(parent)
+	// The package values are fw_cfg entries, so read them before connect seals it.
+	spec, specErr := readPackageSpec(t.fwCfgDir)
+	grant, grantErr := readGrant(t.fwCfgDir)
+	conn, scope, code, err := t.connect(parent)
 	if err != nil {
-		return code, err
+		return code, errors.Join(err, specErr, grantErr)
 	}
 	defer conn.Close()
+	if err := errors.Join(specErr, grantErr); err != nil {
+		return exitRefused, err
+	}
 	link := hostlink.NewTestStream(conn)
 
 	// The handshake can return before the host refuses the guest, so nothing
@@ -69,6 +75,19 @@ func (t testlink) run(parent context.Context) (int, error) {
 	if parent.Err() != nil {
 		return exitLink, context.Cause(parent)
 	}
+	// Nothing of the tenant's is installed or run for an unbound request.
+	if err := checkBinding(scope, grant, first, time.Now()); err != nil {
+		return exitRefused, err
+	}
+	// A host that sent no credential is refused before anything is installed.
+	if first.Credential == nil || !first.Credential.Complete() {
+		return exitRefused, errors.New("first frame carries no execution credential")
+	}
+	cred := *first.Credential
+	first.Credential = nil
+	if err := t.pkg.install(spec); err != nil {
+		return exitRefused, fmt.Errorf("install package: %w", err)
+	}
 
 	srv := runtimeapi.New()
 	ln, err := net.Listen("tcp", t.runtimeAddr)
@@ -81,12 +100,7 @@ func (t testlink) run(parent context.Context) (int, error) {
 
 	ctx, cancel := context.WithCancelCause(parent)
 	defer cancel(nil)
-	cmd, err := t.startBootstrap(ctx, []string{
-		"AWS_LAMBDA_RUNTIME_API=" + ln.Addr().String(),
-		"_HANDLER=" + *handler,
-		"LAMBDA_TASK_ROOT=" + taskRoot,
-		"PATH=/bin:/sbin:/usr/bin",
-	})
+	cmd, err := t.startBootstrap(ctx, bootstrapEnv(ln.Addr().String(), spec.handler, cred))
 	if err != nil {
 		return exitLink, fmt.Errorf("start bootstrap: %w", err)
 	}
@@ -94,7 +108,11 @@ func (t testlink) run(parent context.Context) (int, error) {
 	go func() {
 		defer close(exited)
 		err := cmd.Wait()
-		cancel(fmt.Errorf("bootstrap exited: %w", err))
+		// Once a response is being received the result stands, so only an
+		// exit before that cancels the invocation.
+		if srv.BootstrapExited() {
+			cancel(fmt.Errorf("bootstrap exited: %w", err))
+		}
 	}()
 
 	err = agent.Run(ctx, &oneInvocation{Link: link, first: &first}, srv)
@@ -106,11 +124,32 @@ func (t testlink) run(parent context.Context) (int, error) {
 	return exitClosed, nil
 }
 
+// bootstrapEnv is the bootstrap's whole environment: the runtime variables and
+// the execution role's three credential variables.
+func bootstrapEnv(runtimeAPI, handler string, cred hostlink.Credential) []string {
+	return []string{
+		"AWS_LAMBDA_RUNTIME_API=" + runtimeAPI,
+		"_HANDLER=" + handler,
+		"LAMBDA_TASK_ROOT=" + taskRoot,
+		"PATH=/bin:/sbin:/usr/bin",
+		"AWS_ACCESS_KEY_ID=" + cred.AccessKeyID,
+		"AWS_SECRET_ACCESS_KEY=" + cred.SecretAccessKey,
+		"AWS_SESSION_TOKEN=" + cred.SessionToken,
+	}
+}
+
+// shutdownGrace is how long past the invocation deadline the agent waits for
+// the host's shutdown frame once its result is written.
+const shutdownGrace = 2 * time.Second
+
 // oneInvocation serves the host's first request and then stops the agent:
-// each test VM runs a single invocation.
+// each test VM runs a single invocation. After the result is written it waits
+// for the host's shutdown frame or EOF, so the result is not cut off by the
+// poweroff that follows the agent's exit.
 type oneInvocation struct {
 	hostlink.Link
-	first *hostlink.Request
+	first      *hostlink.Request
+	deadlineMs int64
 }
 
 var _ hostlink.Link = (*oneInvocation)(nil)
@@ -119,10 +158,36 @@ func (o *oneInvocation) Next(ctx context.Context) (hostlink.Request, error) {
 	if err := ctx.Err(); err != nil {
 		return hostlink.Request{}, err
 	}
-	if o.first == nil {
-		return hostlink.Request{}, hostlink.ErrShutdown
+	if o.first != nil {
+		req := *o.first
+		o.first = nil
+		o.deadlineMs = req.DeadlineMs
+		return req, nil
 	}
-	req := *o.first
-	o.first = nil
-	return req, nil
+	return hostlink.Request{}, o.awaitShutdown(ctx)
+}
+
+// awaitShutdown reads the host's next frame and always ends the agent: with
+// ErrShutdown for a shutdown frame, EOF, a stray frame or an expired wait.
+func (o *oneInvocation) awaitShutdown(ctx context.Context) error {
+	end := time.UnixMilli(o.deadlineMs)
+	if now := time.Now(); end.Before(now) {
+		end = now
+	}
+	wctx, cancel := context.WithDeadline(ctx, end.Add(shutdownGrace))
+	defer cancel()
+	_, err := o.Link.Next(wctx)
+	switch {
+	case errors.Is(err, hostlink.ErrShutdown), errors.Is(err, io.EOF):
+		return hostlink.ErrShutdown
+	case err == nil:
+		slog.Warn("lambda-agent: frame after the invocation was not dispatched")
+		return hostlink.ErrShutdown
+	case ctx.Err() != nil:
+		return err
+	case wctx.Err() != nil:
+		slog.Warn("lambda-agent: no shutdown from the host; exiting")
+		return hostlink.ErrShutdown
+	}
+	return err
 }

cmd/lambda-agent/start_testlink_test.go

@@ -5,10 +5,15 @@ package main
 import (
 	"bytes"
 	"context"
+	"crypto/sha256"
+	"encoding/hex"
 	"io"
 	"net/http"
 	"os"
 	"os/exec"
+	"path/filepath"
+	"strconv"
+	"strings"
 	"testing"
 	"time"
 
@@ -21,12 +26,18 @@ import (
 
 const helperEnv = "LAMBDA_AGENT_TEST_BOOTSTRAP"
 
+// helperHandlerOut names the file the helper bootstrap writes its _HANDLER to.
+const helperHandlerOut = "LAMBDA_AGENT_TEST_HANDLER_OUT"
+
 // TestHelperBootstrap is the fake function bootstrap, run as a subprocess: it
 // echoes one invocation through the Runtime API, then waits for the next.
 func TestHelperBootstrap(t *testing.T) {
 	if os.Getenv(helperEnv) != "1" {
 		return
 	}
+	if out := os.Getenv(helperHandlerOut); out != "" {
+		require.NoError(t, os.WriteFile(out, []byte(os.Getenv("_HANDLER")), 0o600))
+	}
 	base := "http://" + os.Getenv("AWS_LAMBDA_RUNTIME_API") + "/2018-06-01/runtime/invocation/"
 	resp, err := http.Get(base + "next")
 	require.NoError(t, err)
@@ -37,24 +48,67 @@ func TestHelperBootstrap(t *testing.T) {
 	post, err := http.Post(base+id+"/response", "application/json", bytes.NewReader(body))
 	require.NoError(t, err)
 	post.Body.Close()
+	if os.Getenv(helperExitAfterPost) == "1" {
+		return
+	}
 	_, _ = http.Get(base + "next")
 }
 
+// helperExitAfterPost makes the helper bootstrap exit once its response is posted.
+const helperExitAfterPost = "LAMBDA_AGENT_TEST_EXIT_AFTER_POST"
+
+// testGrantEntry grants g-1 for attempt a-1 on the fixture scope's env-1/1.
+const testGrantEntry = "g-1\na-1\nenv-1\n1\n"
+
+// testHandler is the handler the fixture's host configures.
+const testHandler = "bootstrap.fixture"
+
 type bootstrapStarts struct {
 	n       int
 	started chan struct{}
+	task    string
+	sealed  bool
+	// handlerOut receives the _HANDLER the bootstrap process itself saw.
+	handlerOut string
+	// env is the environment the agent gave the bootstrap.
+	env []string
 }
 
-func (b *bootstrapStarts) testlink(g *guest) testlink {
+// testlink builds a guest whose fw_cfg also carries a controlled package,
+// installed into a temp dir without mounting anything.
+func (b *bootstrapStarts) testlink(t *testing.T, entries map[string][]byte) testlink {
+	t.Helper()
+	zipBytes := testZip(t, map[string]string{"bootstrap": "#!/bin/sh\n", "lib/data.txt": "data"})
+	disk := filepath.Join(t.TempDir(), "disk")
+	require.NoError(t, os.WriteFile(disk, zipBytes, 0o600))
+	sum := sha256.Sum256(zipBytes)
+	entries[entryPackageSize] = []byte(strconv.Itoa(len(zipBytes)))
+	entries[entryPackageSHA256] = []byte(hex.EncodeToString(sum[:]))
+	if _, ok := entries[entryHandler]; !ok {
+		entries[entryHandler] = []byte(testHandler)
+	}
+	if _, ok := entries[entryGrant]; !ok {
+		entries[entryGrant] = []byte(testGrantEntry)
+	}
+	g := newGuest(t, entries)
 	b.started = make(chan struct{}, 1)
+	b.task = t.TempDir()
+	b.handlerOut = filepath.Join(t.TempDir(), "handler")
 	return testlink{
-		startup:     g.startup(),
+		startup: g.startup(),
+		pkg: taskPackage{
+			device: disk, dir: b.task,
+			loadDriver: func() error { return nil },
+			mountDir:   func(string) error { return nil },
+			sealDir:    func(string) error { b.sealed = true; return nil },
+		},
 		runtimeAddr: "127.0.0.1:0",
 		startBootstrap: func(ctx context.Context, env []string) (*exec.Cmd, error) {
 			b.n++
+			b.env = env
 			b.started <- struct{}{}
 			cmd := exec.CommandContext(ctx, os.Args[0], "-test.run=^TestHelperBootstrap$")
-			cmd.Env = append(env, helperEnv+"=1")
+			cmd.Env = append(env, helperEnv+"=1", helperHandlerOut+"="+b.handlerOut)
 			return cmd, cmd.Start()
 		},
 	}
@@ -80,7 +134,7 @@ func TestTestlinkServesOneInvocationThenStops(t *testing.T) {
 	h := newHost(t)
 	h.accept(ctx)
 	var b bootstrapStarts
-	done := runTestlinkAsync(ctx, b.testlink(newGuest(t, h.entries())))
+	done := runTestlinkAsync(ctx, b.testlink(t, h.entries()))
 
 	conn := h.guest(t)
 	defer conn.Close()
@@ -90,7 +144,7 @@ func TestTestlinkServesOneInvocationThenStops(t *testing.T) {
 	case <-time.After(100 * time.Millisecond):
 	}
 	hs := hostlink.NewTestStream(conn)
-	req := hostlink.Request{GrantID: "g-1", AttemptID: "a-1", RequestID: "r-1", DeadlineMs: time.Now().Add(5 * time.Second).UnixMilli(), Payload: []byte(`{"a":40,"b":2}`)}
+	req := hostlink.Request{Credential: &testCredential, GrantID: "g-1", AttemptID: "a-1", RequestID: "r-1", DeadlineMs: time.Now().Add(5 * time.Second).UnixMilli(), Payload: []byte(`{"a":40,"b":2}`)}
 	require.NoError(t, hs.HostSend(req))
 	got, err := hs.HostResult(ctx)
 	require.NoError(t, err)
@@ -109,15 +163,161 @@ func TestTestlinkServesOneInvocationThenStops(t *testing.T) {
 		assert.Error(t, err, "agent answered a second invocation")
 	}
 	assert.Equal(t, 1, b.n)
+	seen, err := os.ReadFile(b.handlerOut)
+	require.NoError(t, err, "bootstrap process did not report its handler")
+	assert.Equal(t, testHandler, string(seen), "bootstrap process not given the configured handler")
+	assert.True(t, b.sealed, "task dir not made read-only")
+	assert.FileExists(t, filepath.Join(b.task, "lib", "data.txt"))
+	assert.ElementsMatch(t, []string{
+		"AWS_LAMBDA_RUNTIME_API=" + envValue(b.env, "AWS_LAMBDA_RUNTIME_API"),
+		"_HANDLER=" + testHandler,
+		"LAMBDA_TASK_ROOT=" + taskRoot,
+		"PATH=/bin:/sbin:/usr/bin",
+		"AWS_ACCESS_KEY_ID=" + testCredential.AccessKeyID,
+		"AWS_SECRET_ACCESS_KEY=" + testCredential.SecretAccessKey,
+		"AWS_SESSION_TOKEN=" + testCredential.SessionToken,
+	}, b.env, "the bootstrap gets exactly the runtime and credential variables")
+	assert.NotEmpty(t, envValue(b.env, "AWS_LAMBDA_RUNTIME_API"))
 }
 
+var testCredential = hostlink.Credential{AccessKeyID: "ASIAGUESTKEY", SecretAccessKey: "guest-secret", SessionToken: "guest-token"}
+
+func envValue(env []string, name string) string {
+	for _, kv := range env {
+		if v, ok := strings.CutPrefix(kv, name+"="); ok {
+			return v
+		}
+	}
+	return ""
+}
+
+// A first frame without a complete credential installs and runs nothing.
+func TestTestlinkFrameWithoutCredentialNeverStartsBootstrap(t *testing.T) {
+	for name, cred := range map[string]*hostlink.Credential{
+		"none":       nil,
+		"incomplete": {AccessKeyID: "ASIAGUESTKEY"},
+	} {
+		t.Run(name, func(t *testing.T) {
+			ctx, cancel := context.WithTimeout(context.Background(), 10*time.Second)
+			defer cancel()
+			h := newHost(t)
+			h.accept(ctx)
+			var b bootstrapStarts
+			done := runTestlinkAsync(ctx, b.testlink(t, h.entries()))
+			conn := h.guest(t)
+			defer conn.Close()
+			require.NoError(t, hostlink.NewTestStream(conn).HostSend(hostlink.Request{Credential: cred, GrantID: "g-1", AttemptID: "a-1",
+				RequestID: "r-1", DeadlineMs: time.Now().Add(5 * time.Second).UnixMilli(), Payload: []byte(`{}`)}))
+
+			res := wait(t, done, "agent to stop")
+			assert.Equal(t, exitRefused, res.code)
+			assert.ErrorContains(t, res.err, "no execution credential")
+			assert.Zero(t, b.n)
+			assert.False(t, b.sealed)
+			assert.NoFileExists(t, filepath.Join(b.task, "bootstrap"), "package installed")
+		})
+	}
+}
+
+// A bootstrap that posts its response and exits at once does not cost the
+// host the result; the agent then waits for the host's shutdown.
+func TestTestlinkBootstrapExitingAfterResponseStillDelivers(t *testing.T) {
+	ctx, cancel := context.WithTimeout(context.Background(), 10*time.Second)
+	defer cancel()
+	h := newHost(t)
+	h.accept(ctx)
+	var b bootstrapStarts
+	tl := b.testlink(t, h.entries())
+	start := tl.startBootstrap
+	tl.startBootstrap = func(ctx context.Context, env []string) (*exec.Cmd, error) {
+		return start(ctx, append(env, helperExitAfterPost+"=1"))
+	}
+	done := runTestlinkAsync(ctx, tl)
+
+	conn := h.guest(t)
+	defer conn.Close()
+	hs := hostlink.NewTestStream(conn)
+	payload := bytes.Repeat([]byte("0123456789"), 300_000)
+	req := hostlink.Request{Credential: &testCredential, GrantID: "g-1", AttemptID: "a-1", RequestID: "r-1", DeadlineMs: time.Now().Add(5 * time.Second).UnixMilli(), Payload: payload}
+	require.NoError(t, hs.HostSend(req))
+	got, err := hs.HostResult(ctx)
+	require.NoError(t, err)
+	assert.Equal(t, hostlink.Response{GrantID: "g-1", AttemptID: "a-1", RequestID: "r-1", Payload: payload}, got)
+
+	require.NoError(t, hs.HostShutdown())
+	res := wait(t, done, "agent to stop after the host's shutdown")
+	assert.Equal(t, exitClosed, res.code)
+	assert.NoError(t, res.err)
+}
+
+// The agent stays up for the host's shutdown rather than exiting as soon as
+// its result is written, and never dispatches a stray frame.
+func TestOneInvocationAfterResult(t *testing.T) {
+	first := hostlink.Request{RequestID: "r-1", DeadlineMs: time.Now().Add(time.Minute).UnixMilli()}
+	cases := map[string]struct {
+		next func(context.Context) (hostlink.Request, error)
+		want error
+	}{
+		"shutdown frame": {func(context.Context) (hostlink.Request, error) { return hostlink.Request{}, hostlink.ErrShutdown }, hostlink.ErrShutdown},
+		"eof":            {func(context.Context) (hostlink.Request, error) { return hostlink.Request{}, io.EOF }, hostlink.ErrShutdown},
+		"a second request is not dispatched": {func(context.Context) (hostlink.Request, error) {
+			return hostlink.Request{RequestID: "r-2"}, nil
+		}, hostlink.ErrShutdown},
+	}
+	for name, tc := range cases {
+		t.Run(name, func(t *testing.T) {
+			o := &oneInvocation{Link: stubLink{tc.next}, first: &first}
+			_, err := o.Next(context.Background())
+			require.NoError(t, err)
+			req, err := o.Next(context.Background())
+			require.ErrorIs(t, err, tc.want)
+			assert.Empty(t, req.RequestID)
+		})
+	}
+	t.Run("host never sends shutdown", func(t *testing.T) {
+		past := first
+		past.DeadlineMs = time.Now().Add(-time.Second).UnixMilli()
+		o := &oneInvocation{Link: stubLink{func(ctx context.Context) (hostlink.Request, error) {
+			<-ctx.Done()
+			return hostlink.Request{}, ctx.Err()
+		}}, first: &past}
+		_, err := o.Next(context.Background())
+		require.NoError(t, err)
+		start := time.Now()
+		_, err = o.Next(context.Background())
+		require.ErrorIs(t, err, hostlink.ErrShutdown)
+		assert.GreaterOrEqual(t, time.Since(start), shutdownGrace-100*time.Millisecond, "waited the grace")
+	})
+	t.Run("parent cancel is not a shutdown", func(t *testing.T) {
+		o := &oneInvocation{Link: stubLink{func(ctx context.Context) (hostlink.Request, error) {
+			<-ctx.Done()
+			return hostlink.Request{}, ctx.Err()
+		}}, first: &first}
+		_, _ = o.Next(context.Background())
+		ctx, cancel := context.WithCancel(context.Background())
+		cancel()
+		_, err := o.Next(ctx)
+		require.ErrorIs(t, err, context.Canceled)
+	})
+}
+
+type stubLink struct {
+	next func(context.Context) (hostlink.Request, error)
+}
+
+var _ hostlink.Link = stubLink{}
+
+func (l stubLink) Next(ctx context.Context) (hostlink.Request, error) { return l.next(ctx) }
+
+func (stubLink) Complete(context.Context, hostlink.Response) error { return nil }
+
 func TestTestlinkShutdownBeforeInvocationNeverStartsBootstrap(t *testing.T) {
 	ctx, cancel := context.WithTimeout(context.Background(), 10*time.Second)
 	defer cancel()
 	h := newHost(t)
 	h.accept(ctx)
 	var b bootstrapStarts
-	done := runTestlinkAsync(ctx, b.testlink(newGuest(t, h.entries())))
+	done := runTestlinkAsync(ctx, b.testlink(t, h.entries()))
 
 	conn := h.guest(t)
 	defer conn.Close()
@@ -142,7 +342,7 @@ func TestTestlinkHostRefusalNeverStartsBootstrap(t *testing.T) {
 	entries[lambdav1.EntryGuestKey], entries[lambdav1.EntryGuestCert] = wrong[lambdav1.EntryGuestKey], wrong[lambdav1.EntryGuestCert]
 	var b bootstrapStarts
 
-	res := wait(t, runTestlinkAsync(ctx, b.testlink(newGuest(t, entries))), "agent to stop")
+	res := wait(t, runTestlinkAsync(ctx, b.testlink(t, entries)), "agent to stop")
 	assert.Equal(t, exitLink, res.code)
 	require.Error(t, res.err)
 	assert.Contains(t, res.err.Error(), "bad certificate")
@@ -155,7 +355,7 @@ func TestTestlinkHostCloseBeforeFrameNeverStartsBootstrap(t *testing.T) {
 	h := newHost(t)
 	h.accept(ctx)
 	var b bootstrapStarts
-	done := runTestlinkAsync(ctx, b.testlink(newGuest(t, h.entries())))
+	done := runTestlinkAsync(ctx, b.testlink(t, h.entries()))
 
 	require.NoError(t, h.guest(t).Close())
 	res := wait(t, done, "agent to stop")
@@ -177,3 +377,130 @@ func TestOneInvocationHonoursCancelledContext(t *testing.T) {
 	require.NoError(t, err)
 	assert.Equal(t, "r-1", req.RequestID)
 }
+
+// A package that fails its check is refused after authentication, and the
+// bootstrap never starts.
+func TestTestlinkBadPackageNeverStartsBootstrap(t *testing.T) {
+	ctx, cancel := context.WithTimeout(context.Background(), 10*time.Second)
+	defer cancel()
+	h := newHost(t)
+	h.accept(ctx)
+	var b bootstrapStarts
+	tl := b.testlink(t, h.entries())
+	tl.pkg.device = filepath.Join(t.TempDir(), "other")
+	require.NoError(t, os.WriteFile(tl.pkg.device, testZip(t, map[string]string{"bootstrap": "x"}), 0o600))
+	done := runTestlinkAsync(ctx, tl)
+
+	conn := h.guest(t)
+	defer conn.Close()
+	require.NoError(t, hostlink.NewTestStream(conn).HostSend(hostlink.Request{Credential: &testCredential, GrantID: "g-1", AttemptID: "a-1", RequestID: "r-1",
+		DeadlineMs: time.Now().Add(5 * time.Second).UnixMilli()}))
+	res := wait(t, done, "agent to stop")
+	assert.Equal(t, exitRefused, res.code)
+	assert.ErrorContains(t, res.err, "install package")
+	assert.Zero(t, b.n)
+	assert.False(t, b.sealed)
+}
+
+// Missing package values still seal fw_cfg, and the agent refuses before
+// waiting for any frame.
+func TestTestlinkMissingPackageSpecRefuses(t *testing.T) {
+	ctx, cancel := context.WithTimeout(context.Background(), 10*time.Second)
+	defer cancel()
+	h := newHost(t)
+	h.accept(ctx)
+	var b bootstrapStarts
+	tl := b.testlink(t, h.entries())
+	require.NoError(t, os.RemoveAll(filepath.Join(tl.fwCfgDir, entryPackageSHA256)))
+
+	res := wait(t, runTestlinkAsync(ctx, tl), "agent to stop")
+	assert.Equal(t, exitRefused, res.code)
+	assert.ErrorContains(t, res.err, "read "+entryPackageSHA256)
+	assert.Zero(t, b.n)
+	_, err := os.Stat(tl.fwCfgRoot)
+	assert.ErrorIs(t, err, os.ErrNotExist, "fw_cfg not sealed")
+}
+
+// A request that does not carry this VM's grant, or a grant for another
+// environment generation than the one authenticated, installs and runs
+// nothing of the tenant's.
+func TestTestlinkUnboundRequestNeverStartsBootstrap(t *testing.T) {
+	cases := map[string]struct {
+		grant            string
+		grantID, attempt string
+		deadlineMs       int64
+		want             string
+	}{
+		"frame with no deadline":      {testGrantEntry, "g-1", "a-1", -1, "no deadline"},
+		"frame with expired deadline": {testGrantEntry, "g-1", "a-1", time.Now().Add(-time.Second).UnixMilli(), "deadline already passed"},
+		"frame for another grant":     {testGrantEntry, "g-2", "a-1", 0, "does not carry this VM's grant"},
+		"frame for another attempt":   {testGrantEntry, "g-1", "a-2", 0, "does not carry this VM's grant"},
+		"frame with no grant":         {testGrantEntry, "", "", 0, "does not carry this VM's grant"},
+		"grant for another env":       {"g-1\na-1\nenv-2\n1\n", "g-1", "a-1", 0, "not for the authenticated environment"},
+		"grant for another gen":       {"g-1\na-1\nenv-1\n2\n", "g-1", "a-1", 0, "not for the authenticated environment"},
+	}
+	for name, tc := range cases {
+		t.Run(name, func(t *testing.T) {
+			ctx, cancel := context.WithTimeout(context.Background(), 10*time.Second)
+			defer cancel()
+			h := newHost(t)
+			h.accept(ctx)
+			var b bootstrapStarts
+			entries := h.entries()
+			entries[entryGrant] = []byte(tc.grant)
+			done := runTestlinkAsync(ctx, b.testlink(t, entries))
+			conn := h.guest(t)
+			defer conn.Close()
+			deadline := time.Now().Add(5 * time.Second).UnixMilli()
+			switch {
+			case tc.deadlineMs < 0:
+				deadline = 0
+			case tc.deadlineMs > 0:
+				deadline = tc.deadlineMs
+			}
+			req := hostlink.Request{Credential: &testCredential, GrantID: tc.grantID, AttemptID: tc.attempt, RequestID: "r-1",
+				DeadlineMs: deadline, Payload: []byte(`{}`)}
+			require.NoError(t, hostlink.NewTestStream(conn).HostSend(req))
+
+			res := wait(t, done, "agent to stop")
+			assert.Equal(t, exitRefused, res.code)
+			assert.ErrorContains(t, res.err, tc.want)
+			assert.Zero(t, b.n)
+			assert.False(t, b.sealed)
+			assert.NoFileExists(t, filepath.Join(b.task, "bootstrap"), "package installed")
+		})
+	}
+}
+
+// A missing or malformed grant entry refuses before any frame is awaited.
+func TestTestlinkBadGrantEntryRefuses(t *testing.T) {
+	cases := map[string]*string{
+		"missing":       nil,
+		"too few lines": ptr("g-1\na-1\nenv-1\n"),
+		"generation 0":  ptr("g-1\na-1\nenv-1\n0\n"),
+		"empty grant":   ptr("\na-1\nenv-1\n1\n"),
+	}
+	for name, grant := range cases {
+		t.Run(name, func(t *testing.T) {
+			ctx, cancel := context.WithTimeout(context.Background(), 10*time.Second)
+			defer cancel()
+			h := newHost(t)
+			h.accept(ctx)
+			var b bootstrapStarts
+			entries := h.entries()
+			if grant != nil {
+				entries[entryGrant] = []byte(*grant)
+			}
+			tl := b.testlink(t, entries)
+			if grant == nil {
+				require.NoError(t, os.RemoveAll(filepath.Join(tl.fwCfgDir, entryGrant)))
+			}
+			res := wait(t, runTestlinkAsync(ctx, tl), "agent to stop")
+			assert.Equal(t, exitRefused, res.code)
+			assert.ErrorContains(t, res.err, entryGrant)
+			assert.Zero(t, b.n)
+		})
+	}
+}
+
+func ptr(s string) *string { return &s }

cmd/lambda-image-agent-tag/main.go

@@ -0,0 +1,224 @@
+// Command lambda-image-agent-tag checks whether the lambda-agent archived in a
+// built Lambda guest initramfs carries the lambdatestlink build tag. Build
+// information is recorded by the build; it is not proof of how the agent behaves.
+package main
+
+import (
+	"bufio"
+	"bytes"
+	"compress/gzip"
+	"debug/buildinfo"
+	"errors"
+	"fmt"
+	"io"
+	"os"
+	"slices"
+	"strconv"
+	"strings"
+)
+
+const (
+	agentPath    = "usr/local/bin/lambda-agent"
+	maxArchive   = 1 << 30
+	maxAgent     = 256 << 20
+	maxName      = 4096
+	newcHeader   = 110
+	trailerName  = "TRAILER!!!"
+	modeTypeMask = 0o170000
+	modeDir      = 0o040000
+	modeRegular  = 0o100000
+)
+
+var agentParents = []string{"usr", "usr/local", "usr/local/bin"}
+
+func main() {
+	if len(os.Args) != 3 || (os.Args[2] != "present" && os.Args[2] != "absent") {
+		fmt.Fprintln(os.Stderr, "usage: lambda-image-agent-tag <initramfs.cpio.gz> present|absent")
+		os.Exit(2)
+	}
+	if err := run(os.Args[1], os.Args[2] == "present"); err != nil {
+		fmt.Fprintf(os.Stderr, "%s: %v\n", os.Args[1], err)
+		os.Exit(1)
+	}
+}
+
+func run(path string, want bool) error {
+	fi, err := os.Lstat(path)
+	if err != nil {
+		return err
+	}
+	if !fi.Mode().IsRegular() {
+		return errors.New("not a regular file")
+	}
+	f, err := os.Open(path)
+	if err != nil {
+		return err
+	}
+	defer f.Close()
+	agent, err := archivedAgent(f)
+	if err != nil {
+		return err
+	}
+	return checkTag(agent, want)
+}
+
+// archivedAgent returns the single agent entry after checking listed names,
+// entry types, the agent's link count and zero-only padding after the trailer.
+// It is not a full kernel-equivalent parser; 070702 CRCs, for one, are unchecked.
+func archivedAgent(r io.Reader) ([]byte, error) {
+	zr, err := gzip.NewReader(bufio.NewReader(r))
+	if err != nil {
+		return nil, fmt.Errorf("not a readable gzip archive: %w", err)
+	}
+	raw, err := io.ReadAll(io.LimitReader(zr, maxArchive+1))
+	if err != nil {
+		return nil, fmt.Errorf("not a readable gzip archive: %w", err)
+	}
+	if len(raw) > maxArchive {
+		return nil, fmt.Errorf("archive larger than %d bytes", maxArchive)
+	}
+
+	seen := map[string]bool{}
+	var agent []byte
+	off := 0
+	for {
+		e, next, err := parseEntry(raw, off)
+		if err != nil {
+			return nil, fmt.Errorf("entry at offset %d: %w", off, err)
+		}
+		off = next
+		if e.name == trailerName {
+			break
+		}
+		if !canonical(e.name) {
+			return nil, fmt.Errorf("entry name %q is not canonical", e.name)
+		}
+		if seen[e.name] {
+			return nil, fmt.Errorf("entry %q appears more than once", e.name)
+		}
+		seen[e.name] = true
+		for _, p := range agentParents {
+			if e.name == p && e.mode&modeTypeMask != modeDir {
+				return nil, fmt.Errorf("%s is not a directory", p)
+			}
+		}
+		if e.name == agentPath {
+			for _, p := range agentParents {
+				if !seen[p] {
+					return nil, fmt.Errorf("%s has no %s directory entry before it", agentPath, p)
+				}
+			}
+			if e.mode&modeTypeMask != modeRegular {
+				return nil, fmt.Errorf("%s is not a regular file", agentPath)
+			}
+			if e.nlink != 1 {
+				return nil, fmt.Errorf("%s has %d links, want 1", agentPath, e.nlink)
+			}
+			if len(e.data) > maxAgent {
+				return nil, fmt.Errorf("%s is larger than %d bytes", agentPath, maxAgent)
+			}
+			agent = e.data
+		}
+	}
+	if i := bytes.IndexFunc(raw[off:], func(r rune) bool { return r != 0 }); i >= 0 {
+		return nil, fmt.Errorf("data after the archive trailer at offset %d", off+i)
+	}
+	if agent == nil {
+		return nil, fmt.Errorf("no %s entry", agentPath)
+	}
+	return agent, nil
+}
+
+type entry struct {
+	name  string
+	mode  uint64
+	nlink uint64
+	data  []byte
+}
+
+func parseEntry(raw []byte, off int) (entry, int, error) {
+	if len(raw)-off < newcHeader {
+		return entry{}, 0, errors.New("truncated header")
+	}
+	h := raw[off : off+newcHeader]
+	if m := string(h[:6]); m != "070701" && m != "070702" {
+		return entry{}, 0, fmt.Errorf("magic %q is not newc", m)
+	}
+	field := func(i int) (uint64, error) {
+		return strconv.ParseUint(string(h[6+8*i:14+8*i]), 16, 32)
+	}
+	mode, err1 := field(1)
+	nlink, err2 := field(4)
+	size, err3 := field(6)
+	nameSize, err4 := field(11)
+	if err := errors.Join(err1, err2, err3, err4); err != nil {
+		return entry{}, 0, fmt.Errorf("bad header field: %w", err)
+	}
+	if nameSize < 2 || nameSize > maxName {
+		return entry{}, 0, fmt.Errorf("name size %d out of range", nameSize)
+	}
+	nameEnd := off + newcHeader + int(nameSize)
+	dataOff := align4(nameEnd)
+	if dataOff > len(raw) || size > uint64(len(raw)-dataOff) {
+		return entry{}, 0, errors.New("truncated entry")
+	}
+	dataEnd := dataOff + int(size)
+	if dataEnd > len(raw) || align4(dataEnd) > len(raw) {
+		return entry{}, 0, errors.New("truncated entry")
+	}
+	name := raw[off+newcHeader : nameEnd]
+	if name[len(name)-1] != 0 || bytes.IndexByte(name[:len(name)-1], 0) >= 0 {
+		return entry{}, 0, errors.New("name is not one NUL-terminated string")
+	}
+	return entry{name: string(name[:len(name)-1]), mode: mode, nlink: nlink, data: raw[dataOff:dataEnd]}, align4(dataEnd), nil
+}
+
+func align4(n int) int { return (n + 3) &^ 3 }
+
+// canonical accepts "." and relative names with no empty, "." or ".." component.
+func canonical(name string) bool {
+	if name == "." {
+		return true
+	}
+	if name == "" || strings.HasPrefix(name, "/") {
+		return false
+	}
+	for c := range strings.SplitSeq(name, "/") {
+		if c == "" || c == "." || c == ".." {
+			return false
+		}
+	}
+	return true
+}
+
+func checkTag(agent []byte, want bool) error {
+	info, err := buildinfo.Read(bytes.NewReader(agent))
+	if err != nil {
+		return fmt.Errorf("no Go build information in %s: %w", agentPath, err)
+	}
+	return decide(info, want)
+}
+
+// decide refuses build information with no recorded settings, since their
+// absence does not show the tag was absent.
+func decide(info *buildinfo.BuildInfo, want bool) error {
+	if len(info.Settings) == 0 {
+		return fmt.Errorf("%s records no build settings", agentPath)
+	}
+	if hasTag(info, "lambdatestlink") != want {
+		return fmt.Errorf("%s: lambdatestlink build tag present=%v, want %v", agentPath, !want, want)
+	}
+	return nil
+}
+
+func hasTag(info *buildinfo.BuildInfo, tag string) bool {
+	for _, s := range info.Settings {
+		if s.Key != "-tags" {
+			continue
+		}
+		if slices.Contains(strings.Split(s.Value, ","), tag) {
+			return true
+		}
+	}
+	return false
+}

cmd/lambda-image-agent-tag/main_test.go

@@ -0,0 +1,154 @@
+package main
+
+import (
+	"bytes"
+	"compress/gzip"
+	"debug/buildinfo"
+	"fmt"
+	"os"
+	"path/filepath"
+	"runtime/debug"
+	"testing"
+
+	"github.com/stretchr/testify/require"
+)
+
+type fileEntry struct {
+	name  string
+	mode  uint32
+	nlink uint32
+	data  string
+}
+
+func dir(name string) fileEntry { return fileEntry{name: name, mode: modeDir | 0o755, nlink: 2} }
+func file(name, data string) fileEntry {
+	return fileEntry{name: name, mode: modeRegular | 0o755, nlink: 1, data: data}
+}
+func symlink(name, target string) fileEntry {
+	return fileEntry{name: name, mode: 0o120777, nlink: 1, data: target}
+}
+
+// newc writes entries and a trailer in the format the kernel unpacks.
+func newc(entries ...fileEntry) []byte {
+	var b bytes.Buffer
+	for i, e := range append(entries, fileEntry{name: trailerName, nlink: 1}) {
+		fmt.Fprintf(&b, "070701%08X%08X%08X%08X%08X%08X%08X%08X%08X%08X%08X%08X%08X",
+			i+1, e.mode, 0, 0, e.nlink, 0, len(e.data), 0, 0, 0, 0, len(e.name)+1, 0)
+		b.WriteString(e.name)
+		b.WriteByte(0)
+		pad(&b)
+		b.WriteString(e.data)
+		pad(&b)
+	}
+	return b.Bytes()
+}
+
+func pad(b *bytes.Buffer) {
+	for b.Len()%4 != 0 {
+		b.WriteByte(0)
+	}
+}
+
+func gz(parts ...[]byte) []byte {
+	var b bytes.Buffer
+	for _, p := range parts {
+		w := gzip.NewWriter(&b)
+		_, _ = w.Write(p)
+		_ = w.Close()
+	}
+	return b.Bytes()
+}
+
+func withAgent(agent fileEntry) []byte {
+	return newc(dir("."), dir("usr"), dir("usr/local"), dir("usr/local/bin"), agent, file("init", "#!/bin/sh"))
+}
+
+func TestArchivedAgent_ReturnsTheOneArchivedAgent(t *testing.T) {
+	got, err := archivedAgent(bytes.NewReader(gz(withAgent(file(agentPath, "agent-bytes")))))
+	require.NoError(t, err)
+	require.Equal(t, "agent-bytes", string(got))
+
+	padded := append(withAgent(file(agentPath, "agent-bytes")), make([]byte, 1024)...)
+	got, err = archivedAgent(bytes.NewReader(gz(padded)))
+	require.NoError(t, err, "zero padding after the trailer")
+	require.Equal(t, "agent-bytes", string(got))
+}
+
+func TestArchivedAgent_Refuses(t *testing.T) {
+	parents := []fileEntry{dir("."), dir("usr"), dir("usr/local"), dir("usr/local/bin")}
+	agent := file(agentPath, "untagged")
+	small := newc(append(parents, agent)...)
+	second := newc(append(parents[1:], file(agentPath, "tagged"))...)
+	// The second archive starts inside what would be the first one's final block.
+	sameBlock := append(append([]byte{}, small...), second...)
+	for len(sameBlock)%512 != 0 {
+		sameBlock = append(sameBlock, 0)
+	}
+
+	for name, archive := range map[string][]byte{
+		"aliased duplicate":            gz(newc(append(parents, agent, file("././"+agentPath, "tagged"))...)),
+		"leading slash alias":          gz(newc(append(parents, agent, file("/"+agentPath, "tagged"))...)),
+		"dot-dot component":            gz(newc(append(parents, agent, file("usr/local/../local/bin/lambda-agent", "x"))...)),
+		"empty component":              gz(newc(append(parents, agent, file("usr//local/bin/lambda-agent", "x"))...)),
+		"exact duplicate":              gz(newc(append(parents, agent, file(agentPath, "tagged"))...)),
+		"duplicate directory":          gz(newc(dir("usr"), dir("usr"), dir("usr/local"), dir("usr/local/bin"), agent)),
+		"missing directory entry":      gz(newc(dir("usr"), dir("usr/local/bin"), agent)),
+		"directory after the agent":    gz(newc(dir("usr"), dir("usr/local"), agent, dir("usr/local/bin"))),
+		"parent is a symlink":          gz(newc(dir("usr"), dir("usr/local"), symlink("usr/local/bin", "elsewhere"), agent)),
+		"agent is a symlink":           gz(newc(append(parents, symlink(agentPath, "/bin/sh"))...)),
+		"hard-linked agent":            gz(newc(append(parents, fileEntry{name: agentPath, mode: modeRegular | 0o755, nlink: 2, data: "x"})...)),
+		"no agent":                     gz(newc(parents...)),
+		"second archive appended":      gz(small, second),
+		"second archive in same block": gz(sameBlock),
+		"non-zero byte after trailer":  gz(append(append([]byte{}, small...), 0, 0, 1)),
+		"truncated archive":            gz(small[:len(small)-20]),
+		"not newc":                     gz([]byte("not cpio at all, but long enough to fill one newc header: " + string(make([]byte, 120)))),
+		"not gzip":                     small,
+	} {
+		t.Run(name, func(t *testing.T) {
+			_, err := archivedAgent(bytes.NewReader(archive))
+			require.Error(t, err)
+		})
+	}
+}
+
+func TestRun_RefusesASymlinkedImagePath(t *testing.T) {
+	d := t.TempDir()
+	img := filepath.Join(d, "initramfs.cpio.gz")
+	require.NoError(t, os.WriteFile(img, gz(withAgent(file(agentPath, "x"))), 0o644))
+	require.NoError(t, os.Symlink(img, filepath.Join(d, "link")))
+	require.ErrorContains(t, run(filepath.Join(d, "link"), false), "not a regular file")
+}
+
+func TestHasTag_MatchesWholeTagsOnly(t *testing.T) {
+	info := func(tags string) *buildinfo.BuildInfo {
+		return &buildinfo.BuildInfo{Settings: []debug.BuildSetting{{Key: "-buildmode", Value: "exe"}, {Key: "-tags", Value: tags}}}
+	}
+	require.True(t, hasTag(info("lambdatestlink,fips140v1.0"), "lambdatestlink"))
+	require.False(t, hasTag(info("fips140v1.0"), "lambdatestlink"))
+	require.False(t, hasTag(info("lambdatestlinkx,xlambdatestlink"), "lambdatestlink"))
+	require.False(t, hasTag(&buildinfo.BuildInfo{}, "lambdatestlink"))
+}
+
+// The test binary carries real build information, so checkTag reads it and
+// refuses the opposite answer; bytes with no build information are refused.
+func TestCheckTag_ReadsRealBuildInformation(t *testing.T) {
+	exe, err := os.Executable()
+	require.NoError(t, err)
+	raw, err := os.ReadFile(exe)
+	require.NoError(t, err)
+	info, err := buildinfo.Read(bytes.NewReader(raw))
+	require.NoError(t, err)
+	tagged := hasTag(info, "lambdatestlink")
+	require.NoError(t, checkTag(raw, tagged))
+	require.Error(t, checkTag(raw, !tagged))
+	require.Error(t, checkTag([]byte("not a Go binary"), false))
+}
+
+func TestDecide_RefusesMissingBuildSettings(t *testing.T) {
+	require.ErrorContains(t, decide(&buildinfo.BuildInfo{}, false), "no build settings")
+	require.ErrorContains(t, decide(&buildinfo.BuildInfo{}, true), "no build settings")
+	untagged := &buildinfo.BuildInfo{Settings: []debug.BuildSetting{{Key: "-buildmode", Value: "exe"}}}
+	require.NoError(t, decide(untagged, false))
+	require.Error(t, decide(untagged, true))
+}

cmd/lambda-image-digest/main.go

@@ -0,0 +1,26 @@
+//go:build lambdatestlink
+
+// Command lambda-image-digest prints the test-link runtime image digest of one
+// directory, selecting the boot files the same way the daemon does. The digest
+// identifies the bytes; it does not authenticate where they came from.
+package main
+
+import (
+	"fmt"
+	"os"
+
+	"github.com/mulgadc/spinifex/spinifex/runtime/compute/lambda/testlink"
+)
+
+func main() {
+	if len(os.Args) != 2 {
+		fmt.Fprintln(os.Stderr, "usage: lambda-image-digest <image-dir>")
+		os.Exit(2)
+	}
+	digest, err := testlink.DirDigest(os.Args[1])
+	if err != nil {
+		fmt.Fprintln(os.Stderr, err)
+		os.Exit(1)
+	}
+	fmt.Println(digest)
+}

contracts/cluster/v1/facts.go

@@ -56,6 +56,8 @@ type NodeFacts struct {
 	Memory      MemoryFacts       `json:"memory"`
 	NICs        []NICFacts        `json:"nics"`
 	Planes      []PlaneFacts      `json:"planes"`
+	Disks       []DiskFacts       `json:"disks,omitempty"`
+	Filesystems []FilesystemFacts `json:"filesystems,omitempty"`
 	Sources     map[string]string `json:"sources"`
 }
 
@@ -134,6 +136,42 @@ type MemoryGroup struct {
 	Serials       []string `json:"serials,omitempty"`
 }
 
+// DiskFacts is one physical drive as sysfs reports it. Transport is nvme, sata,
+// usb or virtio, and empty when the device path names none, as a RAID
+// controller's virtual disk does.
+type DiskFacts struct {
+	Name       string `json:"name"`
+	Path       string `json:"path"`
+	SizeBytes  uint64 `json:"size_bytes"`
+	Model      string `json:"model,omitempty"`
+	Transport  string `json:"transport,omitempty"`
+	Rotational bool   `json:"rotational,omitempty"`
+	Removable  bool   `json:"removable,omitempty"`
+}
+
+// FilesystemFacts is one mounted block-device or ZFS filesystem. BindOf names
+// the directory a bind mount shows; Disk names the drives under it. Error says
+// why Usage is missing or stale.
+type FilesystemFacts struct {
+	Mount  string           `json:"mount"`
+	Source string           `json:"source"`
+	Disk   string           `json:"disk,omitempty"`
+	Type   string           `json:"type"`
+	BindOf string           `json:"bind_of,omitempty"`
+	Usage  *FilesystemUsage `json:"usage,omitempty"`
+	Error  string           `json:"error,omitempty"`
+	// ID identifies the mount on its node, so a filesystem replaced at the
+	// same path is read afresh. Empty for a mount the daemon cannot see.
+	ID string `json:"-"`
+}
+
+// FilesystemUsage is a filesystem's last statfs reading, as df counts it.
+type FilesystemUsage struct {
+	SizeBytes  uint64 `json:"size_bytes"`
+	UsedBytes  uint64 `json:"used_bytes"`
+	AvailBytes uint64 `json:"avail_bytes"`
+}
+
 // NICFacts is one physical network interface and, where LLDP reports it, the
 // switch port it is cabled to.
 type NICFacts struct {

contracts/ec2/v1/README.md

@@ -23,7 +23,7 @@ subject shape are wire compatibility. Callers and receivers must use this
 package rather than rebuild those details. `commands_test.go` preserves the
 version-1 JSON shape and subject form.
 
-This package owns shared command payloads and the drain-volume response. It
+This package owns shared command payloads. It
 does not define every action's reply body, authorize the request, store EC2
 state or implement NATS transport. Those remain responsibilities of the EC2
 domain and its runtime composition.
@@ -37,6 +37,13 @@ contract or an approved compatibility plan; it is not a package refactor.
 Only the subject is owned here today.
 The request (`account_id`) and reply (`vpc_id`, `error`) bodies are still declared as anonymous structs on each side; moving them into this package is a contract change that needs its own characterization first.
 
+## Volume held elsewhere
+
+A start refused because another node holds one of the instance's volumes is an `IncorrectState` error payload on the `ec2.start`, `ec2.start.<node>` and `ec2.cmd.<instance-id>` replies, and the AWS body cannot say why.
+The owning node adds the `X-Volume-Held-Elsewhere: true` header to that reply, the `ec2.start` forwarder relays it, and `StartInstances` wraps `ErrVolumeHeldElsewhere` around the error it rebuilds, so a caller matches it with `errors.Is` rather than reading the message.
+The owning node sets the header only from the EBS provider's `volume_held_elsewhere` code, which every provider returns for this refusal; no reader matches the message text.
+`VolumeHeldElsewhereText` is the wording viperblockd puts in that refusal's message, not a signal.
+
 ## 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.

contracts/ec2/v1/commands.go

@@ -29,7 +29,6 @@ type EC2InstanceCommand struct {
 	Attributes                EC2CommandAttributes       `json:"attributes"`
 	AttachVolumeData          *AttachVolumeData          `json:"attach_volume_data,omitempty"`
 	DetachVolumeData          *DetachVolumeData          `json:"detach_volume_data,omitempty"`
-	DrainVolumeData           *DrainVolumeData           `json:"drain_volume_data,omitempty"`
 	AttachENIData             *AttachENIData             `json:"attach_eni_data,omitempty"`
 	DetachENIData             *DetachENIData             `json:"detach_eni_data,omitempty"`
 	IamProfileAssociationData *IamProfileAssociationData `json:"iam_profile_association_data,omitempty"`
@@ -45,7 +44,6 @@ type EC2CommandAttributes struct {
 	StartInstance               bool `json:"start_instance"`
 	AttachVolume                bool `json:"attach_volume"`
 	DetachVolume                bool `json:"detach_volume"`
-	DrainVolume                 bool `json:"drain_volume,omitempty"`
 	RebootInstance              bool `json:"reboot_instance"`
 	AttachENI                   bool `json:"attach_eni"`
 	DetachENI                   bool `json:"detach_eni"`
@@ -69,33 +67,6 @@ type DetachVolumeData struct {
 	Force    bool   `json:"force,omitempty"`
 }
 
-// DrainVolumeData carries parameters for a drain-volume command, which flushes
-// the volume's in-flight writes to S3 before a snapshot reads them. The command
-// is addressed to the instance the volume is attached to because the drain
-// socket only exists on the node hosting it.
-type DrainVolumeData struct {
-	VolumeID string `json:"volume_id"`
-}
-
-const (
-	// DrainVolumeStatusDrained is the ack a node returns once the volume's
-	// in-flight writes have reached S3.
-	DrainVolumeStatusDrained = "drained"
-
-	// DrainVolumeStatusNotRunning is the ack a node returns when it still holds
-	// the instance but the VM is not running, so no process is writing to the
-	// volume and its sealed checkpoint is already current. An attachment record
-	// outlives the writer — stop deliberately leaves boot volumes attached — so
-	// the caller cannot tell this from the record alone.
-	DrainVolumeStatusNotRunning = "not-running"
-)
-
-// DrainVolumeResponse is the reply to a drain-volume command.
-type DrainVolumeResponse struct {
-	VolumeID string `json:"volume_id"`
-	Status   string `json:"status"`
-}
-
 // AttachENIData carries parameters for an attach-network-interface command.
 type AttachENIData struct {
 	NetworkInterfaceID string `json:"network_interface_id"`

contracts/ec2/v1/commands_test.go

@@ -16,7 +16,6 @@ func TestEC2InstanceCommandJSONContract(t *testing.T) {
 			StartInstance:               true,
 			AttachVolume:                true,
 			DetachVolume:                true,
-			DrainVolume:                 true,
 			RebootInstance:              true,
 			AttachENI:                   true,
 			DetachENI:                   true,
@@ -28,7 +27,6 @@ func TestEC2InstanceCommandJSONContract(t *testing.T) {
 		},
 		AttachVolumeData:          &AttachVolumeData{VolumeID: "vol-1", Device: "/dev/sdf"},
 		DetachVolumeData:          &DetachVolumeData{VolumeID: "vol-2", Device: "/dev/sdg", Force: true},
-		DrainVolumeData:           &DrainVolumeData{VolumeID: "vol-3"},
 		AttachENIData:             &AttachENIData{NetworkInterfaceID: "eni-1", DeviceIndex: 2},
 		DetachENIData:             &DetachENIData{AttachmentID: "eni-attach-1", Force: true},
 		IamProfileAssociationData: &IamProfileAssociationData{InstanceProfileArn: "arn:aws:iam::123456789012:instance-profile/example"},
@@ -47,7 +45,6 @@ func TestEC2InstanceCommandJSONContract(t *testing.T) {
 			"start_instance": true,
 			"attach_volume": true,
 			"detach_volume": true,
-			"drain_volume": true,
 			"reboot_instance": true,
 			"attach_eni": true,
 			"detach_eni": true,
@@ -59,7 +56,6 @@ func TestEC2InstanceCommandJSONContract(t *testing.T) {
 		},
 		"attach_volume_data": {"volume_id": "vol-1", "device": "/dev/sdf"},
 		"detach_volume_data": {"volume_id": "vol-2", "device": "/dev/sdg", "force": true},
-		"drain_volume_data": {"volume_id": "vol-3"},
 		"attach_eni_data": {"network_interface_id": "eni-1", "device_index": 2},
 		"detach_eni_data": {"attachment_id": "eni-attach-1", "force": true},
 		"iam_profile_association_data": {"instance_profile_arn": "arn:aws:iam::123456789012:instance-profile/example"},
@@ -69,16 +65,6 @@ func TestEC2InstanceCommandJSONContract(t *testing.T) {
 	}`, string(data))
 }
 
-func TestDrainVolumeResponseJSONContract(t *testing.T) {
-	data, err := json.Marshal(DrainVolumeResponse{
-		VolumeID: "vol-1",
-		Status:   DrainVolumeStatusDrained,
-	})
-	require.NoError(t, err)
-	require.JSONEq(t, `{"volume_id":"vol-1","status":"drained"}`, string(data))
-	require.Equal(t, "not-running", DrainVolumeStatusNotRunning)
-}
-
 func TestInstanceCommandSubjects(t *testing.T) {
 	require.Equal(t, "ec2.cmd.i-0123456789abcdef0", InstanceCommandSubject("i-0123456789abcdef0"))
 	require.Equal(t, "ec2.cmd.*", InstanceCommandSubjectWildcard)

contracts/ec2/v1/volume_held.go

@@ -1,7 +1,13 @@
 package ec2v1
 
-// VolumeHeldElsewhereText is the text of a start refused because another node
-// holds one of the instance's volumes. It crosses NATS as message text, so
-// readers match on it; the provider that refuses and every reader share this
-// string so they cannot drift.
+import "errors"
+
+// VolumeHeldElsewhereText is the message viperblockd gives a lease refusal. It
+// is wording for the operator; callers classify the refusal by its code.
 const VolumeHeldElsewhereText = "volume is leased by another owner"
+
+// ErrVolumeHeldElsewhere marks a start refused because another node still
+// holds one of the instance's volumes. The owning node sets it from the EBS
+// provider's held-elsewhere code, and a start reply carries it across NATS as
+// the X-Volume-Held-Elsewhere header rather than as message text.
+var ErrVolumeHeldElsewhere = errors.New("a volume is leased by another node")

contracts/lambda/v1/README.md

@@ -21,7 +21,12 @@ Tenant code may be able to reach the listener, but it cannot open the agent's co
 **Authentication:** mutual TLS 1.3, bound to the environment ID and realization generation, using material the user runtime cannot read (ADR-0009:S17).
 For each generation the executor creates ephemeral self-signed host and guest certificates whose single URI SAN is `spinifex-lambda://<Region>/<account>/<function ID>/<environment ID>/<generation>/<role>`, with role `host` or `guest`.
 Each end validates the peer certificate against the one pinned peer certificate and separately requires that URI SAN with the opposite role; hostname verification is not relied on.
-ALPN carries the L-GUEST major version, `spinifex-lambda-guest/1`, and an unsupported version fails the handshake.
+ALPN carries the L-GUEST major version, `spinifex-lambda-guest/1`, and an unsupported version fails the handshake. Builds with the `lambdatestlink` tag offer only `spinifex-lambda-testlink/2` on both host and guest, whose first frame carries the execution-role credential, so a test-link end and a release end fail the handshake; release builds are unchanged.
+In the same builds, the gateway reaches the host daemon on the local abstract socket `spinifex-lambda-testlink-control/3`, and each request carries the runtime-image digest the gateway was built with; the daemon refuses a missing or different digest before starting the attempt, and a gateway and daemon on different socket names never connect.
+On that socket a request's raw payload is at most 6 MiB, checked by the gateway before encoding and by the daemon after decoding.
+Its encoded frame is at most the base64 length of 6 MiB plus 64 KiB of metadata headroom and 64 bytes of field slack, and a receiver checks the frame length before allocating; metadata has no separate bound within the frame.
+The credential travels as its own frame of at most 8 KiB, and the reply has the same payload and frame bounds as the request.
+The daemon holds at most two connections at once, from accept to close, and leaves the rest unaccepted in the listen backlog.
 Under TLS 1.3 the guest's handshake can complete before the host refuses its certificate, so the guest treats the connection as accepted only after the host's first message.
 
 **Bootstrap material:** the executor passes these fw_cfg entries under `opt/spinifex/lambda/`, and the agent reads them as root and unloads `qemu_fw_cfg` before user code starts.

contracts/operator/v1/admin.go

@@ -25,6 +25,10 @@ const (
 	DeletionStateRunning   = "RUNNING"
 	DeletionStateCompleted = "COMPLETED"
 	DeletionStateFailed    = "FAILED"
+
+	// DeletionStateExercised ends a non-finalizing run: the account is left
+	// TERMINATING and a later DeleteAccount takes the same operation over.
+	DeletionStateExercised = "EXERCISED"
 )
 
 // CreateAccountRequest is the /admin/CreateAccount request body.
@@ -52,6 +56,10 @@ type DeleteAccountRequest struct {
 	ClientToken string `json:"clientToken"`
 	Force       bool   `json:"force,omitempty"`
 	DryRun      bool   `json:"dryRun,omitempty"`
+
+	// NonFinalizing quiesces the account, cleans the quiesced EC2 platform
+	// kinds and stops. No legacy stage runs and the account record stays.
+	NonFinalizing bool `json:"nonFinalizing,omitempty"`
 }
 
 // DescribeAccountDeletionRequest is the /admin/DescribeAccountDeletion body.
@@ -76,8 +84,8 @@ type ListAccountsResponse struct {
 }
 
 // IAMQuiescence is the coordinator's recorded view of IAM quiescence for one
-// fence. Recording only: it admits nothing, and QuiescedScope stays empty until
-// a final inventory re-read exists to catch a late write.
+// fence. Recording only: it admits nothing. QuiescedScope is filled only by a run
+// that guarantees a final inventory re-read, and holds as of EvaluatedAt.
 type IAMQuiescence struct {
 	OperationID   string            `json:"operationId"`
 	Fence         uint64            `json:"fence"`
@@ -136,3 +144,67 @@ const (
 	ReasonOwnerInventoryIncomplete     = "owner-inventory-incomplete"
 	ReasonFinalInventoryRereadNotWired = "final-inventory-reread-not-wired"
 )
+
+// ExerciseResult is what a non-finalizing run did after IAM quiescence.
+type ExerciseResult struct {
+	// Skipped is the fixed reason no cleanup ran; Cleanup is then absent.
+	Skipped    string           `json:"skipped,omitempty"`
+	Cleanup    *PlatformCleanup `json:"cleanup,omitempty"`
+	Accounting []KindAccounting `json:"accounting,omitempty"`
+}
+
+// PlatformCleanup is the EC2 platform cleanup result. It carries resource IDs,
+// fixed reasons and counts only, never names, tags, key material or data.
+type PlatformCleanup struct {
+	Kinds  []CleanupKind  `json:"kinds"`
+	Items  []CleanupItem  `json:"items,omitempty"`
+	Counts []CleanupCount `json:"counts,omitempty"`
+}
+
+// CleanupKind is one kind's scan status after the cleanup's own re-scan.
+type CleanupKind struct {
+	Kind              string   `json:"kind"`
+	Status            string   `json:"status"`
+	IncompleteReasons []string `json:"incompleteReasons,omitempty"`
+}
+
+// CleanupItem is one resource by its ID with what the cleanup did to it.
+type CleanupItem struct {
+	ResourceID  string `json:"resourceId"`
+	Kind        string `json:"kind"`
+	Disposition string `json:"disposition"`
+	Reason      string `json:"reason,omitempty"`
+	MemberCount int    `json:"memberCount,omitempty"`
+}
+
+// CleanupCount is the number of records of one class that have no public ID.
+type CleanupCount struct {
+	Kind        string `json:"kind"`
+	Class       string `json:"class"`
+	Disposition string `json:"disposition"`
+	Reason      string `json:"reason,omitempty"`
+	Count       int    `json:"count"`
+}
+
+// KindAccounting is the final inventory re-read held against the cleanup result
+// for one kind. Quiescence is not rewritten from it.
+type KindAccounting struct {
+	Kind      string `json:"kind"`
+	Status    string `json:"status"`
+	Reason    string `json:"reason,omitempty"`
+	Remaining int    `json:"remaining"`
+}
+
+// Accounting statuses and fixed reasons of a KindAccounting.
+const (
+	AccountingAccounted   = "accounted"
+	AccountingUnaccounted = "unaccounted"
+
+	AccountingReasonRereadIncomplete = "reread-incomplete"
+	AccountingReasonUnnamedRemainder = "unnamed-remainder"
+)
+
+// Fixed reasons an exercise ran no cleanup.
+const (
+	ExerciseSkippedNoQuiescedKind = "no-quiesced-kind"
+)

contracts/operator/v1/admin_test.go

@@ -43,6 +43,27 @@ func TestAdminWire_DeletionStates(t *testing.T) {
 	require.Equal(t, "RUNNING", DeletionStateRunning)
 	require.Equal(t, "COMPLETED", DeletionStateCompleted)
 	require.Equal(t, "FAILED", DeletionStateFailed)
+	require.Equal(t, "EXERCISED", DeletionStateExercised)
+}
+
+func TestAdminWire_NonFinalizingDeleteRequest(t *testing.T) {
+	goldenJSON(t, DeleteAccountRequest{AccountID: "000000000042", AccountName: "acme", ClientToken: "tok-3", NonFinalizing: true},
+		`{"accountId":"000000000042","accountName":"acme","clientToken":"tok-3","nonFinalizing":true}`)
+}
+
+func TestAdminWire_ExerciseResult(t *testing.T) {
+	goldenJSON(t, ExerciseResult{
+		Cleanup: &PlatformCleanup{
+			Kinds:  []CleanupKind{{Kind: "key-pair", Status: "observed-clean-at-scan-time"}},
+			Items:  []CleanupItem{{ResourceID: "pg-1", Kind: "placement-group", Disposition: "blocked", Reason: "instance-members", MemberCount: 1}},
+			Counts: []CleanupCount{{Kind: "key-pair", Class: "public-key-object-without-metadata", Disposition: "deleted", Count: 2}},
+		},
+		Accounting: []KindAccounting{{Kind: "placement-group", Status: AccountingAccounted, Remaining: 1}},
+	}, `{"cleanup":{"kinds":[{"kind":"key-pair","status":"observed-clean-at-scan-time"}],
+		"items":[{"resourceId":"pg-1","kind":"placement-group","disposition":"blocked","reason":"instance-members","memberCount":1}],
+		"counts":[{"kind":"key-pair","class":"public-key-object-without-metadata","disposition":"deleted","count":2}]},
+		"accounting":[{"kind":"placement-group","status":"accounted","remaining":1}]}`)
+	goldenJSON(t, ExerciseResult{Skipped: ExerciseSkippedNoQuiescedKind}, `{"skipped":"no-quiesced-kind"}`)
 }
 
 func TestAdminWire_ListAccountsResponse(t *testing.T) {

contracts/viperblockd/legacy/v1/README.md

@@ -51,19 +51,6 @@ request/reply replies.
 removed its local state and export. These routes and JSON shapes are deployed
 legacy compatibility behaviour, not a generic EBS-provider interface.
 
-## Fence route
-
-`VolumeFencedSubject(node)` is `ebs.<node>.fenced`, or `ebs.fenced` with no
-node for a single-node daemon. The Viperblock adapter publishes
-`VolumeFencedEvent` on it after tearing down an export whose volume lease
-moved to another node; the local daemon subscribes to stop the guest that
-was using it. Node-addressed, like mount and unmount: the guest is local, and
-no other node's daemon has anything to stop.
-
-The event is sent as a request and repeated until the daemon replies with `VolumeFencedAck`, because the fence often runs while the node's NATS server is restarting and a single publish can reach no subscriber.
-The daemon replies only when the event carries a reply subject, so a plain publish is still accepted.
-Repeats are harmless, since fencing a guest that is already fenced changes nothing.
-
 ## Abandon route
 
 `VolumeAbandonSubject(node)` is `ebs.<node>.abandon`. The daemon uses

contracts/viperblockd/legacy/v1/abandon_test.go

@@ -9,9 +9,8 @@ import (
 	"github.com/stretchr/testify/require"
 )
 
-// TestVolumeAbandonSubject_BothForms pins the abandon route the same way the
-// fenced route is pinned: node-addressed, with an empty-node fallback for a
-// single-node daemon.
+// TestVolumeAbandonSubject_BothForms pins the abandon route: node-addressed,
+// with an empty-node fallback for a single-node daemon.
 func TestVolumeAbandonSubject_BothForms(t *testing.T) {
 	assert.Equal(t, "ebs.node-a.abandon", viperblocklegacyv1.VolumeAbandonSubject("node-a"))
 	assert.Equal(t, "ebs.abandon", viperblocklegacyv1.VolumeAbandonSubject(""))

contracts/viperblockd/legacy/v1/fence.go

@@ -1,29 +0,0 @@
-package viperblocklegacyv1
-
-// VolumeFencedSubject is where a fenced volume is announced so the daemon on
-// this node can stop the guest that was using it. Node-addressed: the guest is
-// local, and every other node's daemon has nothing to do about it.
-func VolumeFencedSubject(node string) string {
-	if node == "" {
-		return "ebs.fenced"
-	}
-	return "ebs." + node + ".fenced"
-}
-
-// VolumeFencedEvent announces that a node tore down a volume's export because
-// the lease moved. Winner is who holds it now, for the operator reading the
-// guest's stop reason.
-type VolumeFencedEvent struct {
-	Volume string `json:"volume"`
-	Node   string `json:"node"`
-	Winner string `json:"winner"`
-	Reason string `json:"reason"`
-}
-
-// VolumeFencedAck is the daemon's answer to a VolumeFencedEvent. Any answer
-// ends the sender's retries; Instance names the guest stopped, if one was, and
-// Error says why the daemon could not act on the event.
-type VolumeFencedAck struct {
-	Instance string `json:"instance,omitempty"`
-	Error    string `json:"error,omitempty"`
-}

contracts/viperblockd/legacy/v1/fence_test.go

@@ -1,54 +0,0 @@
-package viperblocklegacyv1_test
-
-import (
-	"encoding/json"
-	"testing"
-
-	viperblocklegacyv1 "github.com/mulgadc/spinifex/contracts/viperblockd/legacy/v1"
-	"github.com/stretchr/testify/assert"
-	"github.com/stretchr/testify/require"
-)
-
-// TestVolumeFencedSubject_IsNodeAddressed pins the routing. A fenced guest is on
-// the node that lost the volume, so a subject without the node in it would reach
-// daemons with nothing to stop and, on a queue group, miss the one that has.
-func TestVolumeFencedSubject_IsNodeAddressed(t *testing.T) {
-	assert.Equal(t, "ebs.node-a.fenced", viperblocklegacyv1.VolumeFencedSubject("node-a"))
-	assert.Equal(t, "ebs.fenced", viperblocklegacyv1.VolumeFencedSubject(""),
-		"a single-node daemon has no node name, and still has to hear its own fences")
-}
-
-// TestVolumeFencedEvent_JSON pins the exact field spelling of the fenced
-// event, published by viperblockd and consumed by the daemon.
-func TestVolumeFencedEvent_JSON(t *testing.T) {
-	golden := `{"volume":"vol-1","node":"node-a","winner":"node-b","reason":"volume lease moved to node-b while this node had the volume open"}`
-	event := viperblocklegacyv1.VolumeFencedEvent{
-		Volume: "vol-1",
-		Node:   "node-a",
-		Winner: "node-b",
-		Reason: "volume lease moved to node-b while this node had the volume open",
-	}
-
-	marshaled, err := json.Marshal(event)
-	require.NoError(t, err)
-	assert.JSONEq(t, golden, string(marshaled))
-
-	var decoded viperblocklegacyv1.VolumeFencedEvent
-	require.NoError(t, json.Unmarshal([]byte(golden), &decoded))
-	assert.Equal(t, event, decoded)
-}
-
-// TestVolumeFencedAck_JSON pins the daemon's answer, which is what ends
-// viperblockd's retries of a fence notice.
-func TestVolumeFencedAck_JSON(t *testing.T) {
-	golden := `{"instance":"i-1","error":"bad event"}`
-	ack := viperblocklegacyv1.VolumeFencedAck{Instance: "i-1", Error: "bad event"}
-
-	marshaled, err := json.Marshal(ack)
-	require.NoError(t, err)
-	assert.JSONEq(t, golden, string(marshaled))
-
-	empty, err := json.Marshal(viperblocklegacyv1.VolumeFencedAck{})
-	require.NoError(t, err)
-	assert.JSONEq(t, `{}`, string(empty), "an ack for a volume with no local guest carries nothing")
-}

docs/COMMANDS.md

@@ -19,7 +19,7 @@ Every service is `PartOf=spinifex.target` and ordered by its own `After=`/`Wants
 | `spinifex-predastore.service` | `spx service predastore start` | S3-compatible object storage, serving that host's nodes and the S3 gate among them |
 | `spinifex-northstar.service` | `spx service northstar start` | Authoritative DNS for `*.spx3.net`, recursive via upstream forwarders, syncing zones from its S3 bucket |
 | `spinifex-viperblock.service` | `spx service viperblock start` | Block storage with the NBD plugin. `KillMode=process`: nbdkit serving a guest survives a restart of this unit |
-| `spinifex-daemon.service` | `spx service spinifex start` | VM orchestration, NATS subscriptions, health endpoint. `KillMode=process`: guest QEMU survives a restart and the new daemon reattaches. `SIGHUP` reloads config |
+| `spinifex-daemon.service` | `spx service spinifex start` | VM orchestration, NATS subscriptions, health endpoint. Two `ExecStartPre` steps run first: `spx admin hostnet check`, and `spx admin node-facts collect`, which runs as root and whose failure does not stop the daemon. `KillMode=process`: guest QEMU survives a restart and the new daemon reattaches. `SIGHUP` reloads config |
 | `spinifex-awsgw.service` | `spx service awsgw start` | AWS-compatible gateway: SigV4 auth, IAM policy enforcement, TLS. Also serves the Private Admin API |
 | `spinifex-vpcd.service` | `spx service vpcd start` | VPC daemon: translates `vpc.*` NATS events to OVN logical switches, ports and routers |
 | `spinifex-ui.service` | `spx service spinifex-ui start` | Web UI serving the React frontend |
@@ -95,7 +95,7 @@ Operational commands for inspecting cluster state. These fan out NATS requests t
 
 | Command | Flags | Description |
 |---------|-------|-------------|
-| `spx admin init` | `--nodes`, `--node`, `--bind`, `--advertise` (external IP off-host clients dial; auto-detected from the default route), `--port`, `--region`, `--az`, `--cluster-name`, `--cluster-bind`, `--cluster-routes`, `--predastore-nodes`, `--predastore-compaction-interval` (seconds; test clusters set a short one), `--services`, `--formation-timeout`, `--token-ttl`, `--force`, `--discard-jetstream` (with `--nodes` >= 2, remove this node's existing JetStream streams once every node has joined; without it a multi-node init over a store that holds streams is refused, and it never proceeds while NATS is running), `--skip-host-dns`, `--email` (operator address for update and security notices), `--no-telemetry`, `--gpu-passthrough`, `--ipsec` (default true; encrypt intra-AZ Geneve via OVN native IPsec cluster-wide — disable only for a trusted single-rack lab; on a host whose IPsec stack is libreswan it defaults to false and `--ipsec=true` is refused), `--host` (leader to join; falls back to multicast discovery), `--external-mode` (`pool` \| `nat` routed NAT for non-bridgeable uplinks, single-node; pair with `setup-ovn.sh --nat-uplink`; note: system instances (ECS/EKS/load-balancer agents) are unsupported in nat v1 — they require `pool` mode), `--external-iface` (WAN NIC for br-external, auto-detected), `--external-source` (`dhcp` \| `static`), `--external-pool` (`start-end`), `--external-prefix-len` (default 24), `--external-gateway` (auto-detected), `--external-dns` (comma-separated IPv4 addresses, at most 3; the DNS servers written to the external pool's `dns_servers` and handed to guests, verbatim; when omitted they are detected from the host resolver at init time, so set it to give every node of a cluster the same value), `--external-bind-bridge` (default `br-wan` when `--external-source=dhcp`), `--gateway-ip` (OVN gateway router's external IP for SNAT; defaults to the pool's range start, required for `nat` mode without DHCP) | Generates root IAM credentials (AKIA-prefixed access key + secret) → creates master.key (AES-256, 32 bytes, 0600) → writes bootstrap.json (consumed on first start) → generates CA + server TLS certificates → generates join token (written to `join-token` file, displayed in join command) → creates NATS config with auth token → writes spinifex.toml, awsgw.toml, predastore.toml → configures AWS CLI `spx` profile → creates directory structure under `~/spinifex/` → points the host resolver at this node's local northstar so the Spinifex zones (ELB/EKS names) resolve from the node itself (opt out with `--skip-host-dns`). `--force` on an already-initialized node is idempotent for crypto: it preserves the existing keys (master.key, predastore/viperblock encryption keys), system + admin credentials, and CA, and refreshes only the config files and the CA-signed server certificate (so a changed bind IP / SANs is picked up without breaking already-joined nodes or CA-baked AMIs). A genuine clean slate requires removing the data dirs first. |
+| `spx admin init` | `--nodes`, `--node`, `--bind`, `--advertise` (external IP off-host clients dial; auto-detected from the default route), `--port`, `--region`, `--az`, `--cluster-name`, `--cluster-bind`, `--cluster-routes`, `--predastore-nodes`, `--predastore-compaction-interval` (seconds; test clusters set a short one), `--services`, `--formation-timeout`, `--token-ttl`, `--force`, `--discard-jetstream` (with `--nodes` >= 2, remove this node's existing JetStream streams once every node has joined; without it a multi-node init over a store that holds streams is refused, and it never proceeds while NATS is running), `--skip-host-dns`, `--email` (operator address for update and security notices), `--no-telemetry`, `--gpu-passthrough`, `--ipsec` (default true; encrypt intra-AZ Geneve via OVN native IPsec cluster-wide — disable only for a trusted single-rack lab; on a host whose IPsec stack is libreswan it defaults to false and `--ipsec=true` is refused), `--external-mode` (`pool` \| `nat` routed NAT for non-bridgeable uplinks, single-node; pair with `setup-ovn.sh --nat-uplink`; note: system instances (ECS/EKS/load-balancer agents) are unsupported in nat v1 — they require `pool` mode), `--external-iface` (WAN NIC for br-external, auto-detected), `--external-source` (`dhcp` \| `static`), `--external-pool` (`start-end`), `--external-prefix-len` (default 24), `--external-gateway` (auto-detected), `--external-dns` (comma-separated IPv4 addresses, at most 3; the DNS servers written to the external pool's `dns_servers` and handed to guests, verbatim; when omitted they are detected from the host resolver at init time, so set it to give every node of a cluster the same value), `--external-bind-bridge` (default `br-wan` when `--external-source=dhcp`), `--gateway-ip` (OVN gateway router's external IP for SNAT; defaults to the pool's range start, required for `nat` mode without DHCP) | Generates root IAM credentials (AKIA-prefixed access key + secret) → creates master.key (AES-256, 32 bytes, 0600) → writes bootstrap.json (consumed on first start) → generates CA + server TLS certificates → generates join token (written to `join-token` file, displayed in join command) → creates NATS config with auth token → writes spinifex.toml, awsgw.toml, predastore.toml → configures AWS CLI `spx` profile → creates directory structure under `~/spinifex/` → points the host resolver at this node's local northstar so the Spinifex zones (ELB/EKS names) resolve from the node itself (opt out with `--skip-host-dns`). `--force` on an already-initialized node is idempotent for crypto: it preserves the existing keys (master.key, predastore/viperblock encryption keys), system + admin credentials, and CA, and refreshes only the config files and the CA-signed server certificate (so a changed bind IP / SANs is picked up without breaking already-joined nodes or CA-baked AMIs). A genuine clean slate requires removing the data dirs first. |
 | `spx admin join` | `--host` (required), `--node` (required), `--token` (required), `--bind`, `--advertise` (defaults to `--bind`, or the auto-detected WAN IP when `--bind` is 0.0.0.0), `--port`, `--region`, `--az`, `--cluster-bind`, `--cluster-routes`, `--data-dir`, `--services`, `--skip-host-dns`, `--email`, `--no-telemetry`, `--predastore-compaction-interval`, `--join-timeout` (how long to keep retrying while the formation server is unreachable, default 20m), `--force` (join even though this node is already initialized or holds JetStream streams, **discarding its own CA, master key and JetStream store**; the store is removed only after formation succeeds, and the join is refused while NATS is running even with `--force`) | Connects to leader node with join token (Authorization: Bearer header) → retrieves cluster configuration → configures local node to join cluster and participate in distributed operations → points the host resolver at this node's local northstar so the Spinifex zones resolve from the node itself (opt out with `--skip-host-dns`) |
 
 ### Version
@@ -129,13 +129,15 @@ Operational commands for inspecting cluster state. These fan out NATS requests t
 |---------|-------|-------------|
 | `spx admin preflight` | `--quiet` (print only non-OK rows) | Reports whether this host has the privileged helper scripts and sudoers grants the installed build depends on. `spx version` knows the build but not the host: `setup.sh` and `setup-ovn.sh` install those assets, and a binary-swap deploy (`make deploy`, or a manual copy) skips them. It also reports any running service still executing a replaced binary — overwriting `/usr/local/bin/spx` while `spinifex.target` is up does not move services onto it, so each keeps executing the replaced inode until it restarts and a node can serve a mixture of two builds indefinitely. Checks compare content, not existence, so a present-but-stale asset is caught. On x86_64 it also checks that the QEMU binary the daemon will run lists the `microvm` machine type and that the microvm firmware `/usr/share/spinifex/microvm/qboot.rom` is installed, since load balancers boot on both. It runs the host network check `spinifex-daemon` and `spinifex-vpcd` run before they start, and fails while firewalld or ufw is active, since the Spinifex nftables policy is the only supported node firewall. It reports a Warning, without failing, when systemd-resolved is active but `/etc/resolv.conf` is not its stub, since lookups then bypass the route that sends Spinifex names to northstar. Exits non-zero if any asset is Missing, Stale, Ungranted, Unsupported or Failed. |
 | `spx admin node js-probe` | `--timeout` (default: 10s) | Round-trips a canary key through the cluster-state KV bucket and exits non-zero if the write or the read-back fails. JetStream can latch a write error per message block and keep serving reads — nats-server's filestore records the failure on the block and returns it on every later write, and nothing clears it short of a restart. The HTTP health endpoint reports the server-wide JetStream state, which that latch never touches, so only an actual write observes it. |
+| `spx admin node-facts collect` | — | Runs `dmidecode`, `ethtool`, `udevadm`, `networkctl lldp` and `systemd-detect-virt`, and writes their raw output to `/etc/spinifex/node-facts/`, readable by the spinifex group. `spinifex-daemon.service` runs it as a root `ExecStartPre` because the daemon itself is unprivileged; the daemon parses the files when the operator console asks. A probe that fails writes no new file, so a file from an earlier successful run can remain and the console reads it as available; a source with no file is shown as unavailable. |
 | `spx admin banner` | `--boot-check` | Writes the node information banner to `/etc/issue` (shown before login on the physical/serial console) and `/etc/motd` (shown after SSH login). With `--boot-check` it also detects a management IP that has changed since last boot and updates `/etc/spinifex/node.conf` accordingly. |
 
 ### Storage Inspection
 
 | Command | Flags | Description |
 |---------|-------|-------------|
-| `spx admin volumes orphans` | — | Lists volumes the storage provider holds that the control plane has no record of. Their blocks consume space but they have no API handle, so they can be neither described nor deleted through the EC2 API. **Reports only, never deletes**: the evidence that an orphan is still wanted is the record that went missing, so removal is an operator decision taken per volume. |
+| `spx admin volumes orphans` | — | Lists volumes the storage provider holds that the control plane has no record of. Their blocks consume space but they have no API handle, so they can be neither described nor deleted through the EC2 API. **Reports only, never deletes**: the evidence that an orphan is still wanted is the record that went missing, so removal is an operator decision taken per volume. It also lists UEFI variable stores nothing owns: a `uefi/i-*/` prefix with no instance record and no store record, a `uefi/ami-*/` prefix with no AMI document, and a store record that is not `deleting` and has no instance record. These are reported and never deleted, because an absent instance record is not proof of a terminate. |
+| `spx admin volumes obligations` | — | Lists the open unpublish obligations, oldest first: a volume export that a force detach, or a terminate whose unmount did not complete, left on the node that held the attachment. Each row names the volume, the node, the instance and the age, plus attempts and the last error once the node has tried. The node settles each one once it answers; an obligation stays open while its node is down and closes without the node only once the node has been removed from the cluster. **Reports only.** |
 | `spx admin volumes unsealed` | — | Lists volumes a node has open, or had open, without a confirmed seal to the backend — the node named holds writes predastore may not have. Starting such a volume elsewhere is refused while that node is alive (its daemon heartbeat is under 60 seconds old, or cannot be read) and is otherwise allowed, in which case it opens from the last checkpoint that did reach the backend. This is the list of volumes where that would cost something. |
 | `spx admin storage status` | `--endpoint` (default: this node's AWS gateway), `--region` (default: this node's region), `--ca-bundle` (default: this node's CA) | Sends the signed `GetStorageStatus` action to the AWS gateway with credentials from the standard AWS chain (e.g. `AWS_PROFILE=spinifex`) and prints what the console's Service Metrics page shows: predastore encoding, meta and blob node counts, cluster health, a meta node table (health, raft state, leader, term, commit and applied index) and the blob nodes. Cluster health is `Degraded` when any meta node is unreachable or the nodes reporting raft state do not agree on one leader. A meta node read only through its host's `/readyz` shows healthy with blank raft fields. |
 
@@ -155,8 +157,9 @@ Operational commands for inspecting cluster state. These fan out NATS requests t
 | `spx admin account list --remote` | `--endpoint` (default: this node's AWS gateway), `--region` (default: this node's region), `--ca-bundle` (default: this node's CA) | Same listing read through `POST /admin/ListAccounts` over SigV4 instead of NATS, so an off-cluster operator or the load-test harness sees its tenants without SSH access to a node. Credentials come from the standard AWS credential chain: on a cluster node that means `AWS_PROFILE=spinifex` (there is no default profile, so without it the command fails to find credentials), or equivalent credentials; the caller must be an IAM user in the super-admin account holding `spinifex:ListAccounts`. |
 | `spx admin account quota get <account-id>` | `--endpoint` (default: this node's AWS gateway), `--region` (default: this node's region), `--ca-bundle` (default: this node's CA) | Reads the limits in force for one account through `POST /admin/GetAccountQuota` over SigV4, and reports for each dimension whether it came from the `[quota]` block in `awsgw.toml` (`config`) or from an override stored against the account (`override`). Credentials come from the standard AWS chain; the caller must be an IAM user in the super-admin account holding `spinifex:GetAccountQuota`. |
 | `spx admin account quota set <account-id>` | `--vcpus`, `--vpcs`, `--subnets`, `--eips`, `--volumes`, `--volumes-gib`, `--rds-instances`, `--load-balancers`, `--clear`, plus `--endpoint`, `--region`, `--ca-bundle` | Overrides individual limits for one account through `POST /admin/PutAccountQuota`, leaving every dimension not named inheriting the configured baseline. `-1` uncaps a dimension; `0` is a real limit that denies every request on it. `--clear` removes the whole override record and cannot be combined with a limit flag. The account must already exist. Prints the resulting limits, so a caller never has to guess what it applied. The caller must hold `spinifex:PutAccountQuota`. |
-| `spx admin account delete <account-id>` | `--dry-run`, `--force`, `--yes` (requires `--name`), `--name` (account name, must match the stored record), `--endpoint`, `--region`, `--ca-bundle`, `--client-token` | Always runs through `POST /admin/DeleteAccount` over SigV4 and follows `DescribeAccountDeletion` to completion; the CLI never tears an account down in its own process, so a CLI delete and an API delete cannot race. Credentials come from the standard AWS chain; the caller must be an IAM user in the super-admin account holding `spinifex:DeleteAccount` and `spinifex:DescribeAccountDeletion`. Prints the inventory it is about to remove → confirms the account name (prompted unless `--name` is given) → marks the account `TERMINATING`, which blocks its credentials cluster-wide → deletes its resources in dependency order (compute, attachments, storage, network, platform, identity, account), waiting for each stage to actually drain before starting the next → seals the account's quota override record so no later write can recreate it, which needs every gateway node the cluster configures to answer a read-only probe first (otherwise nothing is sealed and the account stays `TERMINATING`) → releases the name reservation and quota counter → deletes the account record. Anything that will not delete is reported as **stuck** with the reason, and the account stays `TERMINATING` so the residue keeps an owner. `000000000000` and `000000000001` are refused. `--dry-run` lists and changes nothing. The compute stage reaps ECS clusters, EKS nodegroups and clusters, and RDS instances before it reaps plain instances, because their backing VMs belong to the system account and a tenant-scoped instance listing cannot see them. The storage stage reaps the account's S3 buckets: uploads in flight are aborted, then the objects are removed a page at a time, then the bucket. Only buckets the account owns are ever listed, so the shared platform buckets are out of reach rather than filtered out. Platform-managed serving endpoints are reaped ahead of instances, because only their own delete releases the ENI and attached data volume alongside the VM, and spot instance requests are cancelled first so a live request cannot launch a replacement for something already terminated. The network stage reaps load balancers, then target groups, then NAT gateways and egress-only internet gateways before the subnets, security groups and addresses they hold. The platform stage reaps ACM certificates. |
+| `spx admin account delete <account-id>` | `--dry-run`, `--force`, `--non-finalizing` (hidden), `--yes` (requires `--name`), `--name` (account name, must match the stored record), `--endpoint`, `--region`, `--ca-bundle`, `--client-token` | Always runs through `POST /admin/DeleteAccount` over SigV4 and follows `DescribeAccountDeletion` to completion; the CLI never tears an account down in its own process, so a CLI delete and an API delete cannot race. Credentials come from the standard AWS chain; the caller must be an IAM user in the super-admin account holding `spinifex:DeleteAccount` and `spinifex:DescribeAccountDeletion`. Prints the inventory it is about to remove → confirms the account name (prompted unless `--name` is given) → marks the account `TERMINATING`, which blocks its credentials cluster-wide → deletes its resources in dependency order (compute, attachments, storage, network, platform, identity, account), waiting for each stage to actually drain before starting the next → seals the account's quota override record so no later write can recreate it, which needs every gateway node the cluster configures to answer a read-only probe first (otherwise nothing is sealed and the account stays `TERMINATING`) → releases the name reservation and quota counter → deletes the account record. Anything that will not delete is reported as **stuck** with the reason, and the account stays `TERMINATING` so the residue keeps an owner. `000000000000` and `000000000001` are refused. `--dry-run` lists and changes nothing. The compute stage reaps ECS clusters, EKS nodegroups and clusters, and RDS instances before it reaps plain instances, because their backing VMs belong to the system account and a tenant-scoped instance listing cannot see them. The storage stage reaps the account's S3 buckets: uploads in flight are aborted, then the objects are removed a page at a time, then the bucket. Only buckets the account owns are ever listed, so the shared platform buckets are out of reach rather than filtered out. Platform-managed serving endpoints are reaped ahead of instances, because only their own delete releases the ENI and attached data volume alongside the VM, and spot instance requests are cancelled first so a live request cannot launch a replacement for something already terminated. The network stage reaps load balancers, then target groups, then NAT gateways and egress-only internet gateways before the subnets, security groups and addresses they hold. The platform stage reaps ACM certificates. |
 | `spx admin account delete --force` | as above | Escalates only for resources the ordinary API refuses: it clears a volume attachment in the control plane without the guest's cooperation (`ec2.ForceDetachVolume`), hard-destroys after a graceful attempt times out, and treats an already-missing resource as deleted. It never reorders the stages and never widens scope beyond the account. Use it once the guest is being destroyed — this is the escape from a volume stuck attached to an instance that will not terminate, where neither can otherwise be deleted. |
+| `spx admin account delete --non-finalizing` | as above | An exercise run, not a finalising delete, but the cleanup is real. Marks the account `TERMINATING`, deletes the resources of each EC2 platform kind the evaluation found quiesced (for example key pairs, empty placement groups and orphaned tags), re-reads the owner inventory against that cleanup, and ends `EXERCISED` with the account still `TERMINATING`. No other stage runs and the account record, name reservation and quota are left in place. It refuses `--force` and `--dry-run`. |
 | `spx admin account delete --remote` | as above | Accepted for existing scripts and ignored: every delete already goes through the admin API. |
 | `spx admin principal create <name>` | `--grant` (repeatable or comma-separated; default: every admin method — `CreateAccount`, `DeleteAccount`, `DescribeAccountDeletion`, `GetAccountQuota`, `ListAccounts`, `PutAccountQuota`) | Creates an IAM user in the super-admin account with an inline `spx-admin-methods` policy allowing the requested methods, revokes any access key left from a previous run, and mints one new key. Each method is named rather than granted as `spinifex:*`, so a later addition to the admin surface is not authorised by an existing key. The principal's grants are exactly that one policy: re-running with fewer `--grant` values narrows it, and any other inline policy on the user is removed, so a rename cannot leave a wider grant behind. The secret is printed once; store it in a secret manager or an AWS profile, never in a config file. `admin` is refused as a name — it is the account's own administrator and rotating its key would revoke the bootstrap credential. |
 | `spx admin principal list` | — | Lists every principal in the super-admin account with its access-key count and what it may call. An attached policy is reported by name (`AdministratorAccess (attached)`) rather than expanded, which is how an unscoped principal is spotted. |

docs/DESIGN.md

@@ -65,9 +65,10 @@ The EC2 registration (`spinifex/domains/ec2/awsapi`, `NewRegistration`) parses t
 The gateway communicates with daemons via NATS request/response. Most calls go through `natsmsg.NATSRequest`, which marshals the input, attaches the account ID as a NATS header, and unmarshals the typed response:
 
 ```go
-func (s *NATSInstanceService) RunInstances(ctx context.Context, input *ec2.RunInstancesInput, accountID string) (*ec2.Reservation, error) {
-    topic := fmt.Sprintf("ec2.RunInstances.%s", aws.StringValue(input.InstanceType))
-    return natsmsg.NATSRequest[ec2.Reservation](ctx, s.natsConn, topic, input, 5*time.Minute, accountID)
+func runIntoReservation(ctx context.Context, input *ec2.RunInstancesInput, natsConn *nats.Conn, accountID, crID string) (*ec2.Reservation, error) {
+    subject := "ec2.RunInstances.cr." + crID
+    reservation, err := natsmsg.NATSRequest[ec2.Reservation](ctx, natsConn, subject, input, 5*time.Minute, accountID)
+    // ...
 }
 ```
 
@@ -111,14 +112,16 @@ When a daemon handles `RunInstances`:
 
 1. **Resource Check**: Validates CPU/memory availability via `ResourceManager`
 2. **Volume Generation**: Creates boot and EFI volumes via Viperblock
-3. **Volume Mount**: Sends a mount request to the colocated Viperblock worker on `ebs.{node}.mount`, receives NBD URI
+   A UEFI instance whose record carries a variable store pointer has no EFI volume: its `vars.fd` is fetched from the object under `uefi/` in the configured Predastore bucket that the pointer names, recorded in the `spinifex-uefi-stores` JetStream bucket (or reused when `seeded.json` on this node already names the same object) after a firmware architecture and VARS size check, and a record with both or neither is refused.
+   RunInstances still creates the EFI volume, nothing writes `vars.fd` back to the object store yet, and a terminated instance's store is removed by a cluster-wide reaper.
+3. **Volume Mount**: Publishes each volume through the EBS provider contract (`PublishVolume` on `ebs.provider.v1.{node}.mount`), receives its export
 4. **QEMU Launch**: Builds and executes QEMU command with NBD-backed disks
 5. **QMP Monitoring**: Establishes QEMU Machine Protocol connection for VM management
 6. **Command subscription**: Subscribes to `ec2.cmd.{instanceID}` for subsequent commands on this VM
 
 ```go
-// Volume mounting via NATS (runtime/daemon/vm_adapters.go)
-reply, err := ebsRequestWithTrace(ctx, a.nc, instance.AccountID, viperblocklegacyv1.MountSubject(a.node), ebsMountRequest, 30*time.Second)
+// Volume mounting through the EBS provider contract (runtime/daemon/vm_adapters.go)
+published, err := a.provider.PublishVolume(ebsprovider.WithAccountID(ctx, accountID), ebsprovider.PublishVolumeRequest{Versioned: ebsprovider.NewLatestVersioned(), VolumeID: req.Name, NodeID: a.node})
 
 // QEMU launch with NBD-backed drives in args (runtime/compute/vm/vm.go)
 cmd := exec.Command(qemuArchitecture, args...) // qemu-system-x86_64 or qemu-system-aarch64
@@ -224,6 +227,9 @@ It may keep writing for 40 seconds after the last write the store acknowledged,
 A slow acknowledgement therefore shortens the remaining validity rather than extending it, and an acquisition acknowledged after its validity has already run out is refused as retryable.
 
 For a holder that cannot renew, a timer is armed for the validity deadline, leaving a nominal 5 seconds before the entry can expire; when it fires, the export's nbdkit is fenced with SIGKILL (`volume_fence.go`).
+The kill comes first: the read that names the winner (bounded at 2 seconds) and the notice to the daemon follow it, so a hung lease store cannot hold the writer alive.
+The notice is a request that viperblockd repeats every 2 seconds until the local daemon replies with `VolumeFencedAck`, because the fence often runs while the node's NATS server is restarting.
+On the notice the daemon moves the guest using the volume to `error` with state reason `Server.VolumeFenced` and SIGKILLs its QEMU immediately rather than shutting it down gracefully, because there is no disk left to flush to; it only stops the guest process and removes its taps, and never terminates the instance or deletes a volume.
 The fence runs on its own goroutine, so Go scheduling, that goroutine and process termination all delay the kill, and enough delay can use up the whole margin; a kill held in uninterruptible sleep, which the fence waits up to 30 seconds for, is the extreme case.
 Once a lease is lost or has lapsed it stays lost: a renewal that succeeds afterwards is ignored, and another open on the same node cannot share it while the fence runs.
 

docs/compute/launching-instances/README.md

@@ -23,6 +23,7 @@ resources:
 - [Overview](#overview)
 - [Launch](#launch)
 - [Manage](#manage)
+- [Stopping and Data Volumes](#stopping-and-data-volumes)
 - [Modify Instance Attributes](#modify-instance-attributes)
 - [Instance Metadata Options](#instance-metadata-options)
 - [Spot Instances](#spot-instances)
@@ -73,6 +74,31 @@ aws ec2 terminate-instances --instance-ids $INSTANCE_ID
 aws ec2 reboot-instances --instance-ids $INSTANCE_ID
 ```
 
+## Stopping and Data Volumes
+
+A guest that powers itself off (`poweroff`, `shutdown -h`) goes `stopping` then `stopped`, the same as `stop-instances`.
+It is not restarted automatically.
+`InstanceInitiatedShutdownBehavior=terminate` is not honoured yet: a guest shutdown always stops the instance.
+
+Data volumes stay `in-use` and attached to a stopped instance, and `start-instances` mounts them again.
+While the instance is stopped, `attach-volume` of one of its data volumes to another instance returns `VolumeInUse`, and `delete-volume` is refused.
+
+`detach-volume` works on a stopped instance, with or without `--instance-id`.
+The volume becomes `available` and the next start does not attach it.
+The root volume of a stopped instance can be detached too, as on AWS.
+The internal EFI variable store is not a customer volume: it never appears in `describe-volumes` or `BlockDeviceMappings`, and detaching it is refused with `OperationNotPermitted`.
+The instance then cannot start: `start-instances` returns `InvalidParameterValue` with "Instance does not have a volume attached at root (/dev/sda1)".
+
+`attach-volume` works on a stopped instance too.
+The volume must be `available` and in the instance's availability zone, otherwise the call returns `VolumeInUse` or `InvalidVolume.ZoneMismatch`.
+The attachment is recorded on the stopped instance and the volume becomes `in-use` straight away.
+Attaching at the root device name (for example `/dev/sda1`) makes the volume the root again, and the next `start-instances` boots from it.
+
+An instance stopped by an earlier build had its data volumes released to `available` while its record still listed them.
+On start, a listed data volume that was deleted or is now attached to another instance is dropped with a warning in the daemon log, and the start continues.
+A listed volume that is still `available` and unattached is attached again.
+A missing root volume fails the start with the same `InvalidParameterValue` answer.
+
 ## Modify Instance Attributes
 
 Change instance type, user data, or termination protection. Instance type and user data require the instance to be **stopped** first.
@@ -135,6 +161,30 @@ aws ec2 get-console-output --instance-id $INSTANCE_ID \
   --query 'Output' --output text | base64 -d
 ```
 
+Like AWS, the call returns the most recent 64 KiB of the console.
+While an instance is shutting down after a terminate request, the call reads the renamed file described below.
+Once the instance has left the node's running set, which includes a stopped or terminated instance, the node no longer answers and the call returns `InvalidInstanceID.NotFound`.
+AWS still lists a terminated instance for a while after termination, so a caller may expect this call to work then; it does not here, a known gap.
+Whether the API should answer for a stopped instance is not settled here; only the file on the host is kept.
+
+### Where console logs live and how big they get
+
+The serial console of each instance is written on its host to `/run/spinifex/console-<instance-id>.log`.
+`/run` is memory-backed, so these files are cleared when the host reboots, including any left by older releases.
+The daemon keeps them bounded.
+A file over 4 MiB is cut back in place to its newest 1 MiB, and the guest keeps writing while it happens.
+The console is opened in append mode, so a reboot or auto-restart of the guest adds to the same file instead of starting over.
+A live `console-<instance-id>.log` is deleted only when it has not changed for an hour and no QEMU process on the host has it open, so a stopped instance keeps its console for an hour and a quiet running guest keeps it indefinitely.
+If the daemon cannot read the open files of every QEMU process on the host, it deletes nothing on that pass and logs a warning.
+This also clears the file of an instance that was stopped and then terminated, or that was recovered on another node, within about an hour of its last write.
+When an instance is terminated, including a failed first launch and the system instances behind load balancers, queues and EKS control planes, its file is renamed to `console-<instance-id>.log.terminated`.
+The daemon removes a `.terminated` file an hour after its last write.
+
+Known limits:
+
+- The log shipper follows only `console-*.log`, so it stops following a file when it is renamed on terminate, and the last few seconds of output that it had not yet read can be lost.
+- The instance password retrieval reads only the live file name, so it does not find a console renamed after terminate.
+
 ## Instance Types
 
 List instance types available on the current host. The catalog is generated from the host CPU (Intel, AMD, or ARM) and includes burstable (t-family), general purpose (m-family), compute optimised (c-family), and memory optimised (r-family) types.

docs/compute/vpc-networking/README.md

@@ -315,7 +315,7 @@ prefix_len = 24
 
 ### Host access to instances (jumpbox pattern)
 
-The spinifex host automatically reaches every instance's **private IP**: IGW attach installs a host route into OVN (`<vpc-cidr> via <gateway-transit-ip> dev spx-nat-host`) and exempts the transit net from SNAT, so replies to host-initiated connections come back un-NATted. No per-instance setup.
+The spinifex host automatically reaches every instance's **private IP**: each node converges a host route into OVN (`<vpc-cidr> via <gateway-transit-ip> dev spx-nat-host`) for every attached VPC, and the IGW attach exempts the transit net from SNAT, so replies to host-initiated connections come back un-NATted. No per-instance setup. This only exists in routed NAT mode; other modes install no host routes. Every node runs this loop itself, on each intent change, on a nudge from the local IGW hooks and every 30 seconds, so a node that missed the attach event still gets the route. A route is removed on every node when no attached VPC holds its CIDR any more, such as after a detach. When several attached VPCs share one CIDR, the bootstrap VPC wins, then the lowest VPC ID, so detaching one gateway repoints the route to another attached VPC instead of removing it.
 
 Security groups still apply and the default SG is closed to the host, same as AWS — open SSH/ICMP from the transit net first:
 

docs/containers/eks/README.md

@@ -59,6 +59,9 @@ Authentication uses **EKS access entries** (the API authentication mode) — IAM
 
 - **Access entries are the only auth path.** `authentication_mode` accepts `API` (the default when unset) and `API_AND_CONFIG_MAP`; bare `CONFIG_MAP` is rejected. Spinifex has no `aws-auth` ConfigMap path, so `API_AND_CONFIG_MAP` behaves as `API`: grant access with access entries.
 - **The cluster security group is auto-managed.** As on EKS, Spinifex creates one `eks-cluster-sg-<name>-<id>` in your VPC, tagged `kubernetes.io/cluster/<name>=owned` and `aws:eks:cluster-name`, allowing all traffic from itself. It is attached to the workers and the private-endpoint network interface, and `DescribeCluster` reports it as `clusterSecurityGroupId`. A launch template that names security groups replaces it on the workers, as on AWS. Clusters created before this change keep their `-control-plane-sg` and `-nodegroup-sg` until you recreate them. Deleting a cluster deletes the cluster security group and removes only the rules on your other security groups that reference it. `vpc_config.security_group_ids` is ignored.
+- **Node group deletes are asynchronous.** `DeleteNodegroup` returns at once with the group `DELETING`, and `DescribeNodegroup` answers `ResourceNotFoundException` once the workers and the group's instance profile are gone. Deleting a cluster or creating a group of the same name while a group is `DELETING` returns `ResourceInUseException`, as does a second delete of a `DELETING` group.
+- **If a node group delete cannot finish,** the group moves to `DELETE_FAILED` after 30 minutes, with a `NODE_TERMINATION_FAILURE` health issue naming the workers that are still running. Run `DeleteNodegroup` again to resume it.
+- **Each node group owns an EKS-managed instance profile.** Spinifex creates it, tags it `spinifex:eks-cluster`, `spinifex:eks-nodegroup` and `spinifex:eks-profile-id`, and puts the node role in it. It is removed with the group, so you can delete the node role afterwards with no extra steps. Spinifex never deletes an instance profile it did not create, or one whose tags, name or role no longer match what it recorded; it leaves the profile and logs a warning with its name.
 - **The worker image is fixed.** Node groups always boot Spinifex's `eks-node` image. `ami_type` is recorded but does not select the image, and the image must be registered on the cluster before you create a node group.
 - **Workers launch with IMDSv2 required and an IMDS hop limit of 2.**
   This matches AWS managed node groups and lets a pod without IRSA reach the node role through the SDK default credential chain.
@@ -377,6 +380,15 @@ Control plane readiness and kubectl authentication above lists which clusters re
 
 **`create-cluster` rejected with an authentication-mode error.** `CONFIG_MAP` is the one mode Spinifex rejects, because it has no `aws-auth` ConfigMap path. Set `authentication_mode` to `API` or `API_AND_CONFIG_MAP`, or leave it unset, and grant access with access entries.
 
+**`delete-role` fails with `DeleteConflict` after deleting a node group.** Node groups created before per-group instance profiles left an instance profile named after the role, which Spinifex no longer removes because it cannot prove it created it. Remove it yourself:
+
+```bash
+aws iam remove-role-from-instance-profile --instance-profile-name <role-name> --role-name <role-name>
+aws iam delete-instance-profile --instance-profile-name <role-name>
+```
+
+**A node group is `DELETE_FAILED`.** `aws eks describe-nodegroup` lists the workers that did not stop under `health.issues`. Once they are gone, or after you fix the cause, run `aws eks delete-nodegroup` again.
+
 **Addon install fails.** Only addons bundled into the registered `eks-node` image install successfully. If you pin an addon version in Terraform, use the AWS form from `DescribeAddonVersions` (see Add-on versions above), or leave `addon_version` unset so Spinifex selects the catalog default.
 
 ## Reference Architectures

docs/coverage/ec2/README.md

@@ -43,6 +43,21 @@ A missing `Name`, `BlockDeviceMappings` or root snapshot answers `MissingParamet
 Registering with `--ena-support`, `--no-ena-support` and neither, copying the image, and three refusals were run on a three-node cluster with the AWS CLI.
 Launching an instance from an `EnaSupport` image was not run, so the launch path is unchanged by construction and not by test.
 
+### CreateSnapshot Error Mapping
+
+`CreateSnapshot` drains and snapshots the volume inside the EBS provider.
+A provider that cannot snapshot a volume because it is in use, as the `qemunbd` provider cannot for an attached volume, answers `ErrVolumeInUse`, which becomes `IncorrectState` with the message `The volume 'vol-...' is in use. Detach it before creating a snapshot.`
+The `viperblockd` provider drains a mounted export and snapshots it, so an attached volume succeeds there.
+A provider that cannot answer now, for example because another node holds the volume, answers `ServiceUnavailable` so the caller retries.
+Any other provider failure is `ServerInternal`.
+The snapshot taken for `CreateImage` maps these errors the same way.
+
+### EBS Encryption by Default Follows the Platform
+
+Whether new volumes are encrypted is decided by the platform's encryption key, not by the account.
+`GetEbsEncryptionByDefault` and `EnableEbsEncryptionByDefault` report `true` when the platform has an encryption key configured and `false` otherwise, and `EnableEbsEncryptionByDefault` does not change that.
+`DisableEbsEncryptionByDefault` is refused with `OperationNotPermitted` and the message `EBS encryption by default cannot be disabled: this platform encrypts every volume` while the platform encrypts, and answers `false` where it does not.
+
 ### Operations
 
 | Operation |

docs/package-boundary/eks-cluster-lifecycle.md

@@ -34,7 +34,7 @@ Node group workers and CSI-provisioned EBS volumes are never removed by a cluste
 
 ## Gaps
 
-1. Four `eks.*` handlers still hold their request for work that grows with the resource, and the subject table names them as not conformant: `CreateCluster` (reclaiming a `FAILED` cluster of the same name tears its infra down inline), `DeleteNodegroup` (one awaited terminate per worker), `UpdateNodegroupConfig` (one awaited launch or terminate per worker added or removed) and `RestoreSnapshot` (launches a control-plane VM and fences the old one inline).
+1. Three `eks.*` handlers still hold their request for work that grows with the resource, and the subject table names them as not conformant: `CreateCluster` (reclaiming a `FAILED` cluster of the same name tears its infra down inline), `UpdateNodegroupConfig` (one awaited launch or terminate per worker added or removed) and `RestoreSnapshot` (launches a control-plane VM and fences the old one inline).
 2. AWS's answer to a repeat delete on a `DELETING` cluster has not been observed; `DELETING` from the record is the chosen behaviour.
 3. Control-plane, load balancer and VPC resources in the system account are keyed by the bare cluster name, so a teardown by name can reach another tenant's same-named cluster.
 4. The teardown deletes load balancers the in-cluster controller created in the customer account, which AWS leaves behind.

docs/security/network-connections/README.md

@@ -99,7 +99,7 @@ behavior in other words.
 | 4432 | Formation server | HTTPS | Cluster (bootstrap only) | Cluster join coordination; active only while a join token is valid. Binds the node's `--bind` (lan) address. See *Formation port lifecycle* below. | Short-lived bearer token + TLS¹ |
 | 4222 | spinifex-nats (client) | NATS + TLS | Cluster | Internal service bus for EC2/EBS/VPC/S3 handlers | Token + mutual TLS (cluster CA) |
 | 4248 | spinifex-nats (cluster) | NATS + TLS | Cluster | Inter-node NATS federation | Token + mutual TLS (cluster CA) |
-| 5001 | iperf2 (admin network test) | TCP | Cluster, any plane | Throughput test between two nodes, started only by the super-admin account. Listens on one plane address for the length of one test, then exits. Runs with no capabilities. | None — peer-scoped by firewall, and the client may only target its own plane subnet |
+| 5001 | iperf2 (admin network test) | TCP | Cluster, any plane | Throughput test between two nodes, started only by the super-admin account. Listens on one plane address for the length of one test, then exits. Runs with no capabilities. | None — peer-scoped by firewall, and the client is refused unless the target address is the one the named target node reports on that plane |
 | 5300 | northstar | DNS (UDP + TCP) | Cluster | Forward target for every node's per-instance DNS shim, dialled cross-node. Binds the wildcard by design. | None — restrict by firewall |
 | 6660 | predastore (blob node) | QUIC / UDP | Cluster | Erasure-coded object shard transport between hosts. Multi-node clusters only — see *Predastore ports* below. | TLS 1.3, server certificate verified against the cluster CA |
 | 7660 | predastore (meta node) | QUIC / UDP | Cluster | Raft consensus over global state — buckets and the object index — between hosts. Multi-node clusters only. | TLS 1.3, server certificate verified against the cluster CA |

docs/testing/e2e-coverage.md

@@ -110,7 +110,7 @@ Three entry points, in the order they run:
 
 ### `iam` — everything IAM and STS
 
-Policy evaluation against live requests rather than unit-level: a deny applying to the next request without a new credential, resource-scoped grants naming one key pair, `aws:SourceIp` seeing the real client address, assumed-role sessions carrying role policies, and revocation stopping one key alone.
+Policy evaluation against live requests rather than unit-level: a deny applying to the next request without a new credential, resource-scoped grants naming one key pair, a user rename needing `iam:UpdateUser` on both the old and the new name, `aws:SourceIp` seeing the real client address, assumed-role sessions carrying role policies, and revocation stopping one key alone.
 
 ### `cert` — the cluster CA and TLS identity
 

docs/vultr-integration/README.md

@@ -265,7 +265,12 @@ If this returns a list of available instance types, your installation is working
 
 Spinifex is running, but it holds nothing yet: no machine images, no networks, no instances.
 
-Continue to [Setting Up Your Cluster](/docs/setting-up-your-cluster) to import an AMI, create an SSH key pair, create a VPC with a public subnet, and launch your first instance.
+Log in to node 1, and follow [Setting Up Your Cluster](/docs/setting-up-your-cluster) there to import an AMI, create an SSH key pair, create a VPC with a public subnet, and launch your first instance:
+
+```bash
+./deploy-vultr.sh ssh
+```
+
 An instance you launch into a public subnet gets a Vultr reserved IP, which you can see in the Vultr portal under **Network → Reserved IPs**.
 
 ## 8. Test resiliency

internal/awsmodel/ec2error_test.go

@@ -77,8 +77,8 @@ func TestEC2ErrorCatalogMetadataAndClasses(t *testing.T) {
 	if catalog.metadata.VerifiedOn == "" {
 		t.Fatal("verifiedOn is empty")
 	}
-	if len(catalog.codes) != 115 {
-		t.Fatalf("catalog code count = %d, want 115", len(catalog.codes))
+	if len(catalog.codes) != 116 {
+		t.Fatalf("catalog code count = %d, want 116", len(catalog.codes))
 	}
 	if catalog.codes["InvalidParameterValue"] != ec2ErrorClassClient {
 		t.Fatal("InvalidParameterValue is not classified as client")

internal/awsmodel/model.go

@@ -92,6 +92,7 @@ type Shape struct {
 	Min             *float64
 	Max             *float64
 	Pattern         string
+	Documentation   string
 	LocationName    string
 	TimestampFormat string
 	XMLNamespace    XMLNamespace

internal/awsmodel/models/ec2/error-codes.json

@@ -1,6 +1,6 @@
 {
   "source": "https://docs.aws.amazon.com/ec2/latest/devguide/errors-overview.html",
-  "verifiedOn": "2026-08-05",
+  "verifiedOn": "2026-11-04",
   "commonClient": [
     "AuthFailure",
     "Blocked",
@@ -95,6 +95,7 @@
     "InvalidVolumeID.Malformed",
     "InvalidVpcID.NotFound",
     "InvalidVpcRange",
+    "NatGatewayNotFound",
     "OperationNotPermitted",
     "ReservationCapacityExceeded",
     "Resource.AlreadyAssociated",

internal/awsmodel/smithy.go

@@ -443,6 +443,9 @@ func (t *smithyTranslator) applyShapeTraits(shape *Shape, traits smithyTraits) e
 	if shape.Pattern, err = traits.string("smithy.api#pattern"); err != nil {
 		return err
 	}
+	if shape.Documentation, err = traits.string("smithy.api#documentation"); err != nil {
+		return err
+	}
 	if shape.LocationName, err = traits.string("smithy.api#xmlName"); err != nil {
 		return err
 	}

internal/ecsgw/client.go

@@ -7,9 +7,10 @@ package ecsgw
 
 import (
 	"bytes"
-	"context"
+	"crypto/sha256"
 	"crypto/tls"
 	"crypto/x509"
+	"encoding/hex"
 	"fmt"
 	"io"
 	"net/http"
@@ -17,10 +18,8 @@ import (
 	"strings"
 	"time"
 
-	"github.com/aws/aws-sdk-go/aws/credentials"
-	v4 "github.com/aws/aws-sdk-go/aws/signer/v4"
-
 	"github.com/mulgadc/bluebottle/pkg/tlsconfig"
+	"github.com/mulgadc/spinifex/internal/gwsign"
 )
 
 // targetPrefix is the X-Amz-Target service prefix the gateway's ECS dispatch
@@ -28,31 +27,26 @@ import (
 // suffix); this one mirrors the AWS JSON 1.1 convention.
 const targetPrefix = "AmazonEC2ContainerServiceV20141113."
 
-// CredentialsFunc yields the current SigV4 credentials. It is called once per
-// request so a provider backed by IMDS can rotate instance-role creds (and their
-// session token) transparently. Must be safe for concurrent use.
-type CredentialsFunc func(ctx context.Context) (accessKey, secretKey, sessionToken string, err error)
-
 // Client posts SigV4-signed AWS JSON 1.1 requests to the gateway for service
 // "ecs". One client is reused for register/heartbeat/state/poll.
 type Client struct {
 	baseURL    string
-	creds      CredentialsFunc
+	signer     *gwsign.Signer
 	region     string
 	httpClient *http.Client
 }
 
 // New builds a client. caPath optionally pins the gateway TLS CA; empty relies on
-// the system trust store. creds supplies the per-call instance-role credentials.
+// the system trust store. signer supplies the per-call instance-role credentials.
 // region defaults to us-east-1 when empty (SigV4 requires a non-empty region).
 // timeout bounds a single call — callers needing a long-poll pass a larger value
 // than the default register/heartbeat timeout.
-func New(baseURL, caPath string, creds CredentialsFunc, region string, timeout time.Duration) (*Client, error) {
+func New(baseURL, caPath string, signer *gwsign.Signer, region string, timeout time.Duration) (*Client, error) {
 	if baseURL == "" {
 		return nil, fmt.Errorf("ecsgw: baseURL is required")
 	}
-	if creds == nil {
-		return nil, fmt.Errorf("ecsgw: credentials func is required")
+	if signer == nil {
+		return nil, fmt.Errorf("ecsgw: signer is required")
 	}
 	if region == "" {
 		region = "us-east-1"
@@ -79,7 +73,7 @@ func New(baseURL, caPath string, creds CredentialsFunc, region string, timeout t
 
 	return &Client{
 		baseURL: strings.TrimRight(baseURL, "/"),
-		creds:   creds,
+		signer:  signer,
 		region:  region,
 		httpClient: &http.Client{
 			Timeout:   timeout,
@@ -90,7 +84,7 @@ func New(baseURL, caPath string, creds CredentialsFunc, region string, timeout t
 
 // Call SigV4-signs and POSTs body under X-Amz-Target "<prefix><action>" to the
 // gateway root, returning the 2xx response body. Credentials are fetched per call
-// and the v4 signer emits X-Amz-Security-Token for instance-role (temporary)
+// and the signer emits X-Amz-Security-Token for instance-role (temporary)
 // creds, so the gateway's ASIA path verifies them. No retry; callers wrap.
 func (c *Client) Call(action string, body []byte) ([]byte, error) {
 	req, err := http.NewRequest(http.MethodPost, c.baseURL+"/", bytes.NewReader(body))
@@ -100,12 +94,8 @@ func (c *Client) Call(action string, body []byte) ([]byte, error) {
 	req.Header.Set("Content-Type", "application/x-amz-json-1.1")
 	req.Header.Set("X-Amz-Target", targetPrefix+action)
 
-	akid, secret, token, err := c.creds(req.Context())
-	if err != nil {
-		return nil, fmt.Errorf("retrieve credentials: %w", err)
-	}
-	signer := v4.NewSigner(credentials.NewStaticCredentials(akid, secret, token))
-	if _, err := signer.Sign(req, bytes.NewReader(body), "ecs", c.region, time.Now()); err != nil {
+	sum := sha256.Sum256(body)
+	if err := c.signer.Sign(req, hex.EncodeToString(sum[:]), "ecs", c.region); err != nil {
 		return nil, fmt.Errorf("sign request: %w", err)
 	}
 

internal/ecsgw/client_test.go

@@ -1,24 +1,29 @@
 package ecsgw
 
 import (
-	"context"
 	"io"
 	"net/http"
 	"net/http/httptest"
 	"strings"
 	"testing"
 	"time"
+
+	"github.com/aws/aws-sdk-go-v2/credentials"
+	"github.com/mulgadc/bluebottle/pkg/sigv4"
+	"github.com/mulgadc/spinifex/internal/gwsign"
 )
 
-// testCreds returns a fixed temporary credential set (with a session token, as
-// IMDS instance-role creds carry).
-func testCreds(context.Context) (string, string, string, error) {
-	return "ASIATEST", "secret", "session-token-xyz", nil
+const testSecret = "secret"
+
+// testSigner signs with a fixed temporary credential set (with a session token,
+// as IMDS instance-role creds carry).
+func testSigner() *gwsign.Signer {
+	return gwsign.New(credentials.NewStaticCredentialsProvider("ASIATEST", testSecret, "session-token-xyz"))
 }
 
 func newTestClient(t *testing.T, srv *httptest.Server) *Client {
 	t.Helper()
-	c, err := New(srv.URL, "", testCreds, "ap-southeast-2", time.Second)
+	c, err := New(srv.URL, "", testSigner(), "ap-southeast-2", time.Second)
 	if err != nil {
 		t.Fatalf("New: %v", err)
 	}
@@ -30,7 +35,14 @@ func newTestClient(t *testing.T, srv *httptest.Server) *Client {
 // 2xx body.
 func TestClient_CallSignsAndTargets(t *testing.T) {
 	var gotTarget, gotAuth, gotBody, gotToken string
+	var verifyErr error
 	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		// sigv4.Parse reads and hashes the body itself, then restores it.
+		sr, err := sigv4.Parse(r)
+		if err == nil {
+			_, err = sr.Verify(testSecret, "ap-southeast-2", "ecs")
+		}
+		verifyErr = err
 		gotTarget = r.Header.Get("X-Amz-Target")
 		gotAuth = r.Header.Get("Authorization")
 		gotToken = r.Header.Get("X-Amz-Security-Token")
@@ -45,6 +57,9 @@ func TestClient_CallSignsAndTargets(t *testing.T) {
 	if err != nil {
 		t.Fatalf("Call: %v", err)
 	}
+	if verifyErr != nil {
+		t.Fatalf("gateway-side SigV4 verify failed: %v", verifyErr)
+	}
 	if string(out) != `{"ok":true}` {
 		t.Errorf("body = %q", out)
 	}
@@ -78,10 +93,10 @@ func TestClient_CallErrorStatus(t *testing.T) {
 }
 
 func TestNew_Validation(t *testing.T) {
-	if _, err := New("", "", testCreds, "r", 0); err == nil {
+	if _, err := New("", "", testSigner(), "r", 0); err == nil {
 		t.Error("want error for empty baseURL")
 	}
 	if _, err := New("https://gw:9999", "", nil, "r", 0); err == nil {
-		t.Error("want error for nil credentials func")
+		t.Error("want error for nil signer")
 	}
 }

internal/gwsign/signer.go

@@ -15,8 +15,8 @@ import (
 	v4 "github.com/aws/aws-sdk-go-v2/aws/signer/v4"
 	awsconfig "github.com/aws/aws-sdk-go-v2/config"
 	"github.com/aws/aws-sdk-go-v2/credentials"
-	"github.com/aws/aws-sdk-go-v2/credentials/ec2rolecreds"
 	"github.com/aws/aws-sdk-go-v2/feature/ec2/imds"
+	"github.com/mulgadc/spinifex/internal/imdscreds"
 )
 
 // Signer holds a credential provider and signs requests in place.
@@ -24,6 +24,12 @@ type Signer struct {
 	provider aws.CredentialsProvider
 }
 
+// New returns a Signer over provider. Wrap a provider that fetches remotely in
+// aws.NewCredentialsCache, since Sign retrieves on every call.
+func New(provider aws.CredentialsProvider) *Signer {
+	return &Signer{provider: provider}
+}
+
 // NewStatic returns a Signer using fixed long-lived keys (no session token).
 func NewStatic(accessKey, secretKey string) *Signer {
 	return &Signer{provider: credentials.NewStaticCredentialsProvider(accessKey, secretKey, "")}
@@ -46,12 +52,15 @@ func NewIMDS(ctx context.Context, region string) (*Signer, error) {
 
 // NewInstanceRole returns a Signer whose credentials come only from the
 // instance metadata service: no environment, profile or static-key fallback.
+// A failed refresh of expired credentials is an error, not the stale key.
 func NewInstanceRole() *Signer {
-	return newInstanceRole(imds.New(imds.Options{}))
+	return newInstanceRole("")
 }
 
-func newInstanceRole(client *imds.Client) *Signer {
-	return &Signer{provider: aws.NewCredentialsCache(ec2rolecreds.New(func(o *ec2rolecreds.Options) { o.Client = client }))}
+// newInstanceRole reads from the IMDS root base; "" means the default endpoint.
+// optFns adjust the SDK IMDS client, e.g. its retryer; none keeps the defaults.
+func newInstanceRole(base string, optFns ...func(*imds.Options)) *Signer {
+	return &Signer{provider: aws.NewCredentialsCache(imdscreds.NewProvider(nil, base, optFns...))}
 }
 
 // EnsureCredentials retrieves once from the provider, surfacing the error so a

internal/gwsign/signer_test.go

@@ -125,27 +125,41 @@ func TestNewIMDS_BuildsProvider(t *testing.T) {
 	}
 }
 
+// imdsStub serves one instance-role credential over IMDSv2. It echoes the token
+// TTL header as real IMDS does, so the SDK client caches the token rather than
+// re-fetching it, through the retry backoff, on every later call.
+func imdsStub(t *testing.T, accessKey, expiration string) *httptest.Server {
+	t.Helper()
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		switch r.URL.Path {
+		case "/latest/api/token":
+			w.Header().Set("X-Aws-Ec2-Metadata-Token-Ttl-Seconds", "21600")
+			_, _ = w.Write([]byte("token"))
+		case "/latest/meta-data/iam/security-credentials/":
+			_, _ = w.Write([]byte("sqsWorkerRole"))
+		case "/latest/meta-data/iam/security-credentials/sqsWorkerRole":
+			_, _ = w.Write([]byte(`{"Code":"Success","AccessKeyId":"` + accessKey + `","SecretAccessKey":"s","Token":"t","Expiration":"` + expiration + `"}`))
+		default:
+			http.NotFound(w, r)
+		}
+	}))
+	t.Cleanup(srv.Close)
+	return srv
+}
+
+// singleAttempt disables IMDS client retries so an unreachable endpoint fails
+// on the first attempt instead of after the SDK's 1s-per-retry backoff.
+func singleAttempt(o *imds.Options) { o.Retryer = aws.NopRetryer{} }
+
 // The instance-role signer ignores environment keys and signs only with the
 // credentials the metadata service issues.
 func TestNewInstanceRole_UsesOnlyIMDS(t *testing.T) {
 	t.Setenv("AWS_ACCESS_KEY_ID", testAccessKey)
 	t.Setenv("AWS_SECRET_ACCESS_KEY", testSecretKey)
 	const imdsKey = "ASIAIMDSONLYEXAMPLE1"
-	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
-		switch r.URL.Path {
-		case "/latest/api/token":
-			_, _ = w.Write([]byte("token"))
-		case "/latest/meta-data/iam/security-credentials/":
-			_, _ = w.Write([]byte("sqsWorkerRole"))
-		case "/latest/meta-data/iam/security-credentials/sqsWorkerRole":
-			_, _ = w.Write([]byte(`{"Code":"Success","AccessKeyId":"` + imdsKey + `","SecretAccessKey":"s","Token":"t","Expiration":"2099-01-01T00:00:00Z"}`))
-		default:
-			http.NotFound(w, r)
-		}
-	}))
-	t.Cleanup(srv.Close)
+	srv := imdsStub(t, imdsKey, "2099-01-01T00:00:00Z")
 
-	s := newInstanceRole(imds.New(imds.Options{Endpoint: srv.URL}))
+	s := newInstanceRole(srv.URL, singleAttempt)
 	req := httptest.NewRequest(http.MethodPost, "https://gw.example/", nil)
 	if err := s.Sign(req, emptySHA, testService, testRegion); err != nil {
 		t.Fatal(err)
@@ -155,10 +169,31 @@ func TestNewInstanceRole_UsesOnlyIMDS(t *testing.T) {
 	}
 
 	srv.Close()
-	cold := newInstanceRole(imds.New(imds.Options{Endpoint: srv.URL}))
+	cold := newInstanceRole(srv.URL, singleAttempt)
 	if err := cold.EnsureCredentials(context.Background()); err == nil {
 		t.Fatal("environment keys were used when IMDS was unreachable")
 	}
 }
 
+// Once the cached instance-role key has expired, a refresh that cannot reach
+// IMDS must fail the sign rather than reuse the expired key with a new expiry.
+func TestNewInstanceRole_FailedRefreshErrorsInsteadOfStaleKey(t *testing.T) {
+	const staleKey = "ASIASTALEEXAMPLE0001"
+	srv := imdsStub(t, staleKey, "2000-01-01T00:00:00Z")
+	s := newInstanceRole(srv.URL, singleAttempt)
+	if err := s.EnsureCredentials(t.Context()); err != nil {
+		t.Fatalf("first retrieve: %v", err)
+	}
+
+	srv.Close()
+	req := httptest.NewRequest(http.MethodPost, "https://gw.example/", nil)
+	err := s.Sign(req, emptySHA, testService, testRegion)
+	if err == nil {
+		t.Fatalf("Sign after failed refresh = nil, signed with %q", req.Header.Get("Authorization"))
+	}
+	if auth := req.Header.Get("Authorization"); auth != "" {
+		t.Fatalf("request carries Authorization %q after a failed refresh", auth)
+	}
+}
+
 const emptySHA = "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855"

internal/imdscreds/imdscreds.go

@@ -21,38 +21,25 @@ import (
 	"github.com/aws/aws-sdk-go-v2/feature/ec2/imds"
 )
 
-// Credentials is a resolved instance-role credential set with parsed expiry.
-type Credentials struct {
-	AccessKeyID     string
-	SecretAccessKey string
-	SessionToken    string
-	Expiration      time.Time
-}
-
 // Fetch resolves the instance role then its credentials via the SDK's IMDSv2
 // client and the ec2rolecreds provider. base is the IMDS root, e.g.
 // http://169.254.169.254/latest.
-func Fetch(client *http.Client, base string) (Credentials, error) {
+func Fetch(client *http.Client, base string) (aws.Credentials, error) {
 	creds, err := NewProvider(client, base).Retrieve(context.Background())
 	if err != nil {
-		return Credentials{}, fmt.Errorf("fetch IMDS credentials: %w", err)
+		return aws.Credentials{}, fmt.Errorf("fetch IMDS credentials: %w", err)
 	}
 	if creds.AccessKeyID == "" || creds.SecretAccessKey == "" {
-		return Credentials{}, fmt.Errorf("IMDS credentials missing access/secret key")
+		return aws.Credentials{}, fmt.Errorf("IMDS credentials missing access/secret key")
 	}
-	return Credentials{
-		AccessKeyID:     creds.AccessKeyID,
-		SecretAccessKey: creds.SecretAccessKey,
-		SessionToken:    creds.SessionToken,
-		Expiration:      creds.Expires,
-	}, nil
+	return creds, nil
 }
 
 // NewProvider builds an ec2rolecreds provider against base, an IMDS root URL
-// such as http://169.254.169.254/latest. A nil client uses the SDK IMDS
-// client's default HTTP client.
-func NewProvider(client *http.Client, base string) aws.CredentialsProvider {
-	imdsClient := NewClient(client, base)
+// such as http://169.254.169.254/latest. client and optFns are passed to
+// NewClient.
+func NewProvider(client *http.Client, base string, optFns ...func(*imds.Options)) aws.CredentialsProvider {
+	imdsClient := NewClient(client, base, optFns...)
 	return noExtendOnFailure{provider: ec2rolecreds.New(func(o *ec2rolecreds.Options) {
 		o.Client = imdsClient
 	})}
@@ -84,16 +71,15 @@ func (p noExtendOnFailure) AdjustExpiresBy(creds aws.Credentials, dur time.Durat
 	return p.provider.AdjustExpiresBy(creds, dur)
 }
 
-// NewClient builds an SDK IMDS client against base, an IMDS root URL such as
-// http://169.254.169.254/latest. A trailing /latest is stripped since the
-// client appends its own; both forms of base keep working. A nil client uses
-// the SDK IMDS client's default HTTP client.
-func NewClient(client *http.Client, base string) *imds.Client {
+// NewClient builds an SDK IMDS client against base, an IMDS root URL with or
+// without a trailing /latest. A nil client uses the SDK's default HTTP client;
+// optFns adjust the options after the defaults here, as imds.New applies them.
+func NewClient(client *http.Client, base string, optFns ...func(*imds.Options)) *imds.Client {
 	opts := imds.Options{Endpoint: rootEndpoint(base)}
 	if client != nil {
 		opts.HTTPClient = client
 	}
-	return imds.New(opts)
+	return imds.New(opts, optFns...)
 }
 
 // rootEndpoint strips a trailing /latest from base so it can be used as the

internal/imdscreds/imdscreds_test.go

@@ -1,9 +1,11 @@
 package imdscreds
 
 import (
+	"context"
 	"encoding/json"
 	"net/http"
 	"net/http/httptest"
+	"sync/atomic"
 	"testing"
 	"time"
 
@@ -60,8 +62,8 @@ func TestFetch_OK(t *testing.T) {
 	if got.AccessKeyID != "AKIATEST" || got.SecretAccessKey != "secrettest" || got.SessionToken != "sessiontoken" {
 		t.Errorf("creds mismatch: %+v", got)
 	}
-	if !got.Expiration.Equal(exp) {
-		t.Errorf("Expiration = %v, want %v", got.Expiration, exp)
+	if !got.Expires.Equal(exp) {
+		t.Errorf("Expires = %v, want %v", got.Expires, exp)
 	}
 }
 
@@ -132,3 +134,46 @@ func TestNewProvider_WithholdsExtendOnFailure(t *testing.T) {
 		t.Error("provider dropped AdjustExpiresBy, which a caller using an ExpiryWindow relies on")
 	}
 }
+
+// Behind aws.CredentialsCache, as the ecs-agent uses it, a refresh that fails
+// against an expired credential must surface as an error rather than hand back
+// the dead key with a pushed-out expiry.
+func TestNewProvider_CachedFailedRefreshErrors(t *testing.T) {
+	var hits atomic.Int32
+	mux := http.NewServeMux()
+	mux.HandleFunc("/latest/api/token", func(w http.ResponseWriter, r *http.Request) {
+		w.Header().Set("X-Aws-Ec2-Metadata-Token-Ttl-Seconds", "21600")
+		_, _ = w.Write([]byte("v2-token"))
+	})
+	mux.HandleFunc("/latest/meta-data/iam/security-credentials/", func(w http.ResponseWriter, r *http.Request) {
+		if r.URL.Path == "/latest/meta-data/iam/security-credentials/" {
+			_, _ = w.Write([]byte("node-role"))
+			return
+		}
+		if hits.Add(1) > 1 {
+			w.WriteHeader(http.StatusInternalServerError)
+			return
+		}
+		_ = json.NewEncoder(w).Encode(map[string]string{
+			"Code":            "Success",
+			"AccessKeyId":     "AKIA",
+			"SecretAccessKey": "secret",
+			"Expiration":      time.Now().Add(-time.Minute).UTC().Format(time.RFC3339),
+		})
+	})
+	srv := httptest.NewServer(mux)
+	t.Cleanup(srv.Close)
+
+	cache := aws.NewCredentialsCache(NewProvider(srv.Client(), srv.URL+"/latest"))
+	first, err := cache.Retrieve(context.Background())
+	if err != nil {
+		t.Fatalf("first Retrieve: %v", err)
+	}
+	got, err := cache.Retrieve(context.Background())
+	if err == nil {
+		t.Fatalf("refresh failure was reported as success: %+v", got)
+	}
+	if got.Expires.After(first.Expires) {
+		t.Errorf("expiry extended on refresh failure: %v -> %v", first.Expires, got.Expires)
+	}
+}

internal/stsauth/stsauth.go

@@ -8,8 +8,8 @@ package stsauth
 import (
 	"fmt"
 	"net/http"
-	"time"
 
+	awsv2 "github.com/aws/aws-sdk-go-v2/aws"
 	"github.com/aws/aws-sdk-go/aws"
 	awscreds "github.com/aws/aws-sdk-go/aws/credentials"
 	"github.com/aws/aws-sdk-go/aws/session"
@@ -20,24 +20,15 @@ import (
 // scope and the gateway does not enforce a specific value.
 const defaultRegion = "us-east-1"
 
-// Credentials is the decoded result of AssumeRole: a SigV4 credential set bound
-// to the assumed role, with its expiry.
-type Credentials struct {
-	AccessKeyID     string
-	SecretAccessKey string
-	SessionToken    string
-	Expiration      time.Time
-}
-
 // AssumeRole builds an aws-sdk STS client pointed at the gateway and calls
 // AssumeRole with the supplied static (instance-role) credentials, returning the
 // assumed-role credential set. httpClient must trust the gateway CA.
-func AssumeRole(region, gatewayURL string, httpClient *http.Client, akid, secret, sessionToken, roleARN, sessionName string) (Credentials, error) {
+func AssumeRole(region, gatewayURL string, httpClient *http.Client, akid, secret, sessionToken, roleARN, sessionName string) (awsv2.Credentials, error) {
 	if region == "" {
 		region = defaultRegion
 	}
 	if roleARN == "" || sessionName == "" {
-		return Credentials{}, fmt.Errorf("stsauth: roleARN and sessionName are required")
+		return awsv2.Credentials{}, fmt.Errorf("stsauth: roleARN and sessionName are required")
 	}
 	sess, err := session.NewSession(&aws.Config{
 		Region:      aws.String(region),
@@ -47,7 +38,7 @@ func AssumeRole(region, gatewayURL string, httpClient *http.Client, akid, secret
 		DisableSSL:  aws.Bool(false),
 	})
 	if err != nil {
-		return Credentials{}, fmt.Errorf("new session: %w", err)
+		return awsv2.Credentials{}, fmt.Errorf("new session: %w", err)
 	}
 
 	out, err := sts.New(sess).AssumeRole(&sts.AssumeRoleInput{
@@ -55,16 +46,17 @@ func AssumeRole(region, gatewayURL string, httpClient *http.Client, akid, secret
 		RoleSessionName: aws.String(sessionName),
 	})
 	if err != nil {
-		return Credentials{}, fmt.Errorf("AssumeRole: %w", err)
+		return awsv2.Credentials{}, fmt.Errorf("AssumeRole: %w", err)
 	}
 	if out.Credentials == nil {
-		return Credentials{}, fmt.Errorf("AssumeRole returned no credentials")
+		return awsv2.Credentials{}, fmt.Errorf("AssumeRole returned no credentials")
 	}
 	c := out.Credentials
-	return Credentials{
+	return awsv2.Credentials{
 		AccessKeyID:     aws.StringValue(c.AccessKeyId),
 		SecretAccessKey: aws.StringValue(c.SecretAccessKey),
 		SessionToken:    aws.StringValue(c.SessionToken),
-		Expiration:      aws.TimeValue(c.Expiration),
+		CanExpire:       c.Expiration != nil,
+		Expires:         aws.TimeValue(c.Expiration),
 	}, nil
 }

internal/stsauth/stsauth_test.go

@@ -41,8 +41,8 @@ func TestAssumeRole_DecodesCredentials(t *testing.T) {
 		t.Errorf("bad creds: %+v", creds)
 	}
 	want, _ := time.Parse(time.RFC3339, "2031-01-02T03:04:05Z")
-	if !creds.Expiration.Equal(want) {
-		t.Errorf("Expiration = %v, want %v", creds.Expiration, want)
+	if !creds.Expires.Equal(want) {
+		t.Errorf("Expires = %v, want %v", creds.Expires, want)
 	}
 	// The request must carry the AssumeRole action and the role ARN we asked for.
 	if !strings.Contains(gotBody, "Action=AssumeRole") {

internal/testkit/awserror.go

@@ -0,0 +1,29 @@
+package testutil
+
+import (
+	"testing"
+
+	awserrors "github.com/mulgadc/spinifex/spinifex/foundation/aws/errors"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+// RequireAWSError asserts err carries the AWS code and exact client-facing
+// message; an empty message asserts the bare code with no message attached.
+func RequireAWSError(t *testing.T, err error, code, message string) {
+	t.Helper()
+	require.Error(t, err)
+	got, msg, ok := awserrors.ResolveErrorDetail(err)
+	require.True(t, ok, "error %v carries no AWS code", err)
+	assert.Equal(t, code, got)
+	assert.Equal(t, message, msg)
+}
+
+// RequireAWSCode asserts err carries the AWS code, whatever its message.
+func RequireAWSCode(t *testing.T, err error, code string) {
+	t.Helper()
+	require.Error(t, err)
+	got, ok := awserrors.ResolveErrorCode(err)
+	require.True(t, ok, "error %v carries no AWS code", err)
+	assert.Equal(t, code, got)
+}

internal/testkit/drainsocket.go

@@ -5,6 +5,7 @@ import (
 	"os"
 	"path/filepath"
 	"sync"
+	"sync/atomic"
 	"testing"
 
 	"github.com/stretchr/testify/require"
@@ -28,10 +29,11 @@ func SocketTempDir(t *testing.T) string {
 // StartDrainSocket stands in for the NBD plugin's drain socket, at the path the
 // plugin serves it on the node hosting volumeID ({dataDir}/viperblock/{volumeID}
 // /snapshot.sock). Every connection is answered with ack ("OK\n" for a drain
-// that reached S3, "ERR\n" for one that did not) and closed.
-func StartDrainSocket(t *testing.T, dataDir, volumeID, ack string) {
+// that reached S3, "ERR\n" for one that did not) and closed. connections
+// reports how many drains have been asked for so far.
+func StartDrainSocket(t *testing.T, dataDir, volumeID, ack string) (connections func() int) {
 	t.Helper()
-	serveDrainSocket(t, dataDir, volumeID, ack, nil, nil)
+	return serveDrainSocket(t, dataDir, volumeID, ack, nil, nil)
 }
 
 // StartBlockingDrainSocket stands in for a plugin whose flush is still running:
@@ -43,7 +45,7 @@ func StartBlockingDrainSocket(t *testing.T, dataDir, volumeID, ack string) (acce
 
 	acceptedCh := make(chan struct{})
 	releaseCh := make(chan struct{})
-	serveDrainSocket(t, dataDir, volumeID, ack, acceptedCh, releaseCh)
+	_ = serveDrainSocket(t, dataDir, volumeID, ack, acceptedCh, releaseCh)
 
 	var once sync.Once
 	t.Cleanup(func() { once.Do(func() { close(releaseCh) }) })
@@ -52,8 +54,9 @@ func StartBlockingDrainSocket(t *testing.T, dataDir, volumeID, ack string) (acce
 }
 
 // serveDrainSocket answers every connection with ack, optionally closing
-// accepted on the first one and waiting for release before it replies.
-func serveDrainSocket(t *testing.T, dataDir, volumeID, ack string, accepted chan struct{}, release <-chan struct{}) {
+// accepted on the first one and waiting for release before it replies. The
+// returned func counts the connections accepted so far.
+func serveDrainSocket(t *testing.T, dataDir, volumeID, ack string, accepted chan struct{}, release <-chan struct{}) func() int {
 	t.Helper()
 
 	dir := filepath.Join(dataDir, "viperblock", volumeID)
@@ -63,6 +66,7 @@ func serveDrainSocket(t *testing.T, dataDir, volumeID, ack string, accepted chan
 	require.NoError(t, err)
 	t.Cleanup(func() { _ = ln.Close() })
 
+	var count atomic.Int64
 	go func() {
 		var once sync.Once
 		for {
@@ -70,6 +74,7 @@ func serveDrainSocket(t *testing.T, dataDir, volumeID, ack string, accepted chan
 			if err != nil {
 				return // listener closed by the cleanup above
 			}
+			count.Add(1)
 			if accepted != nil {
 				once.Do(func() { close(accepted) })
 			}
@@ -80,4 +85,5 @@ func serveDrainSocket(t *testing.T, dataDir, volumeID, ack string, accepted chan
 			_ = conn.Close()
 		}
 	}()
+	return func() int { return int(count.Load()) }
 }

internal/testkit/nats.go

@@ -27,8 +27,7 @@ func StartTestNATS(t testing.TB) (*server.Server, *nats.Conn) {
 	}
 	ns, err := server.NewServer(opts)
 	require.NoError(t, err)
-	go ns.Start()
-	waitReadyForConnections(t, ns)
+	startServer(t, ns)
 	t.Cleanup(func() { ns.Shutdown() })
 
 	nc, err := nats.Connect(ns.ClientURL())
@@ -37,17 +36,14 @@ func StartTestNATS(t testing.TB) (*server.Server, *nats.Conn) {
 	return ns, nc
 }
 
-// waitReadyForConnections spins until the server accepts connections. Calling
-// ReadyForConnections with a longer timeout makes it sleep 25ms per attempt,
-// which every embedded server in the suite would otherwise pay.
-func waitReadyForConnections(t testing.TB, ns *server.Server) {
+// startServer runs Start in this goroutine: it returns once JetStream is
+// enabled and the client listener is bound, so the readiness check passes on
+// its first look instead of spinning or sleeping while the server comes up.
+func startServer(t testing.TB, ns *server.Server) {
 	t.Helper()
-	deadline := time.Now().Add(5 * time.Second)
-	for !ns.ReadyForConnections(time.Millisecond) {
-		if time.Now().After(deadline) {
-			t.Fatal("embedded NATS server not ready for connections within 5s")
-		}
-		time.Sleep(200 * time.Microsecond)
+	ns.Start()
+	if !ns.ReadyForConnections(5 * time.Second) {
+		t.Fatal("embedded NATS server not ready for connections within 5s")
 	}
 }
 
@@ -66,8 +62,7 @@ func StartTestJetStream(t *testing.T) (*server.Server, *nats.Conn, jetstream.Jet
 	}
 	ns, err := server.NewServer(opts)
 	require.NoError(t, err)
-	go ns.Start()
-	waitReadyForConnections(t, ns)
+	startServer(t, ns)
 	t.Cleanup(func() { ns.Shutdown() })
 
 	nc, err := nats.Connect(ns.ClientURL())
@@ -167,11 +162,11 @@ func registerVpcdStub(t *testing.T, nc *nats.Conn, defaultPayload []byte) {
 	})
 }
 
-// SeedKV creates a KV bucket and populates it with the given entries.
+// SeedKV creates or updates a KV bucket and populates it with the given entries.
 // Returns the KV handle for further use.
 func SeedKV(t *testing.T, js jetstream.JetStream, bucket string, entries map[string][]byte) jetstream.KeyValue {
 	t.Helper()
-	kv, err := js.CreateKeyValue(t.Context(), jetstream.KeyValueConfig{Bucket: bucket, History: 1})
+	kv, err := js.CreateOrUpdateKeyValue(t.Context(), jetstream.KeyValueConfig{Bucket: bucket, History: 1})
 	require.NoError(t, err)
 	for k, v := range entries {
 		_, err := kv.Put(t.Context(), k, v)

internal/testkit/nats_test.go

@@ -0,0 +1,19 @@
+package testutil
+
+import (
+	"testing"
+
+	"github.com/nats-io/nats.go/jetstream"
+	"github.com/stretchr/testify/require"
+)
+
+func TestSeedKV_ExistingBucket(t *testing.T) {
+	_, _, js := StartTestJetStream(t)
+	_, err := js.CreateKeyValue(t.Context(), jetstream.KeyValueConfig{Bucket: "seeded", History: 5})
+	require.NoError(t, err)
+
+	kv := SeedKV(t, js, "seeded", map[string][]byte{"k": []byte("v")})
+	entry, err := kv.Get(t.Context(), "k")
+	require.NoError(t, err)
+	require.Equal(t, []byte("v"), entry.Value())
+}

scripts/README.md

@@ -98,7 +98,10 @@ Both ship to `/usr/local/share/spinifex/` on every node, so they are available o
 | `build-system-image.sh` | Builds a minimal system AMI (Alpine or Ubuntu) from a manifest. All customisation happens inside the libguestfs appliance, so the build touches no host block device, no host mount, and needs no sudo. `--import` also registers it as an AMI. |
 | `publish-system-image.sh` | Converts a built raw image to compressed qcow2, writes the `.sha256` sidecar the catalog verifier expects, and uploads both to the R2 bucket behind `iso.mulgadc.com/system-ami/`. |
 | `build-microvm-image.sh` | Builds the minimal Alpine kernel + initramfs used by the QEMU microvm machine type and the lb-agent, carrying haproxy (ALB) and nginx-stream (NLB) data planes. |
-| `build-qboot.sh` | Builds qboot, the firmware the daemon passes to every microvm with `-bios`, from a pinned upstream commit in a digest-pinned Debian container. Fails unless the output matches the recorded sha256, which is also Debian 13's `qboot.rom`. Writes `qboot.rom` and its GPL-2.0 licence beside the microvm kernel, and the matching source tarball, which releases publish, under `build/qboot/`. `make build-microvm-image` runs it. |
+| `build-lambda-runtime-microvm-image.sh` | Builds the Lambda runtime microVM kernel and initramfs: a guest running the Lambda guest agent (`bin/lambda-agent` by default) with virtio-net for its host control connection. Takes an optional agent binary and output directory (default `build/microvm/lambda-runtime`). Not shipped. |
+| `build-sqs-store-microvm-image.sh` | Builds the SQS microVM kernel and initramfs: one guest running the dedicated SQS `nats-server` and the SQS worker as separate processes, with ext4 support and a clean power-down. The guest never formats a disk. Output goes to `build/microvm/sqs-worker`. |
+| `lib/microvm-image.sh` | Shared steps for the two builders above, sourced rather than run: tool checks, the Alpine rootfs, and module stripping that keeps `virtio*`, `*fw_cfg*` and the modules the caller names with their dependencies. |
+| `build-qboot.sh` | Builds qboot, the firmware the daemon passes to every microvm with `-bios`, from a pinned upstream commit in a digest-pinned Debian container. Fails unless the output matches the recorded sha256, which is also Debian 13's `qboot.rom`. A host whose `/usr/share/qemu/qboot.rom` (`qemu-system-data`) has that hash installs it and skips the container; any other host compiles. Writes `qboot.rom` and its GPL-2.0 licence beside the microvm kernel, and the matching source tarball, which releases publish, under `build/qboot/`. `make build-microvm-image` runs it. |
 | `mkosi-build.sh` | Runs an mkosi system-image build inside the pinned builder container, so a runner needs only Docker and the toolchain cannot drift. `--image` names an output and expands to the right profile list. |
 | `mkosi-publish.sh` | Publishes an mkosi-built image to R2 from inside the same pinned container — qcow2 conversion, `.sha256` sidecar, and upload. |
 | `mkosi-common.sh` | Shared plumbing sourced by both mkosi scripts. Builds the toolchain image once, with the builder account matched to the invoking user so bind-mounted output is not written back as root. |

scripts/build-lambda-runtime-microvm-image.sh

@@ -3,24 +3,28 @@
 # + initramfs: a guest running the Lambda guest agent, with virtio-net for its
 # host control connection. It is not shipped.
 #
+# Usage: build-lambda-runtime-microvm-image.sh [agent-binary [output-dir]]
+#
 # Inputs:
 #   $REPO_ROOT/build/microvm/lambda-runtime/{init.sh,inittab,run-agent}
-#   $REPO_ROOT/bin/lambda-agent  (make bin/lambda-agent)
+#   agent binary, default $REPO_ROOT/bin/lambda-agent (make bin/lambda-agent)
 #
-# Outputs:
-#   $REPO_ROOT/build/microvm/lambda-runtime/vmlinuz
-#   $REPO_ROOT/build/microvm/lambda-runtime/initramfs.cpio.gz
+# Outputs, in output-dir (default $REPO_ROOT/build/microvm/lambda-runtime):
+#   vmlinuz
+#   initramfs.cpio.gz
 #
 # Rootfs strategy and module stripping: scripts/lib/microvm-image.sh.
 # shellcheck disable=SC2034 # variables are consumed by the sourced library
 set -euo pipefail
 
 REPO_ROOT="$(cd "$(dirname "$0")/.." && pwd)"
-BUILD_DIR="$REPO_ROOT/build/microvm/lambda-runtime"
-INIT_SH="$BUILD_DIR/init.sh"
-INITTAB="$BUILD_DIR/inittab"
-RUN_AGENT="$BUILD_DIR/run-agent"
-LAMBDA_AGENT_BIN="$REPO_ROOT/bin/lambda-agent"
+INPUT_DIR="$REPO_ROOT/build/microvm/lambda-runtime"
+INIT_SH="$INPUT_DIR/init.sh"
+INITTAB="$INPUT_DIR/inittab"
+RUN_AGENT="$INPUT_DIR/run-agent"
+LAMBDA_AGENT_BIN="${1:-$REPO_ROOT/bin/lambda-agent}"
+# The library stages beside BUILD_DIR, so it names the output, not the inputs.
+BUILD_DIR="${2:-$INPUT_DIR}"
 TAG="[build-lambda-runtime-microvm-image]"
 
 ALPINE_VERSION="${ALPINE_VERSION:-3.24}"
@@ -38,7 +42,8 @@ MICROVM_GATE_MODULES="qemu_fw_cfg virtio_net"
 source "$REPO_ROOT/scripts/lib/microvm-image.sh"
 
 echo "$TAG repo root: $REPO_ROOT"
-echo "$TAG build dir: $BUILD_DIR"
+echo "$TAG input dir: $INPUT_DIR"
+echo "$TAG output dir: $BUILD_DIR"
 
 microvm_check_tools
 
@@ -57,6 +62,9 @@ if ! file -b "$LAMBDA_AGENT_BIN" | grep -q "statically linked"; then
     exit 1
 fi
 
+# Absolute, because packing the initramfs runs from inside the chroot.
+mkdir -p "$BUILD_DIR"
+BUILD_DIR=$(cd "$BUILD_DIR" && pwd)
 microvm_init_dirs
 microvm_make_rootfs
 microvm_empty_dev

scripts/build-lambda-runtime-microvm-image_test.sh

@@ -0,0 +1,81 @@
+#!/bin/sh
+# build-lambda-runtime-microvm-image_test.sh — a build into a chosen output
+# directory stages and publishes there, and leaves the default input directory
+# untouched.
+#
+# The script runs in a fixture repo against the real library, with only the
+# rootfs, kernel and packing steps replaced so no image is downloaded.
+set -eu
+
+SCRIPTS_DIR=$(CDPATH='' cd -- "$(dirname -- "$0")" && pwd)
+TMP=$(mktemp -d)
+trap 'rm -rf "$TMP"' EXIT
+
+fails=0
+pass() { echo "  ok   - $1"; }
+fail() { echo "  FAIL - $1"; fails=$((fails + 1)); }
+
+echo "-- build-lambda-runtime-microvm-image.sh output directory"
+
+REPO="$TMP/repo"
+IN="$REPO/build/microvm/lambda-runtime"
+mkdir -p "$REPO/scripts/lib" "$IN" "$REPO/bin" "$TMP/bin"
+cp "$SCRIPTS_DIR/build-lambda-runtime-microvm-image.sh" "$REPO/scripts/"
+for f in init.sh inittab run-agent; do printf '%s\n' "$f" > "$IN/$f"; done
+printf 'default image\n' > "$IN/vmlinuz"
+printf '#!/bin/sh\n' > "$REPO/bin/agent-testlink"
+chmod +x "$REPO/bin/agent-testlink"
+printf '#!/bin/sh\necho "ELF 64-bit, statically linked"\n' > "$TMP/bin/file"
+chmod +x "$TMP/bin/file"
+BEFORE=$(cd "$IN" && ls -A | cksum)
+
+# The real library, with the steps that need a rootfs or kernel replaced.
+# Packing runs from inside the chroot as the real step does, and records where
+# the stage is while it still exists.
+cat > "$REPO/scripts/lib/microvm-image.sh" <<EOF
+. "$SCRIPTS_DIR/lib/microvm-image.sh"
+microvm_check_tools() { :; }
+microvm_make_rootfs() { mkdir -p "\$CHROOT_DIR/etc"; }
+microvm_empty_dev() { :; }
+microvm_locate_kernel() { :; }
+microvm_strip_modules() { :; }
+microvm_strip_rootfs() { :; }
+microvm_stage_kernel() { printf 'kernel\n' > "\$STAGE_DIR/vmlinuz"; }
+microvm_pack_initramfs() {
+    (cd "\$CHROOT_DIR" && printf 'initramfs\n' | gzip > "\$STAGE_DIR/initramfs.cpio.gz")
+    dirname "\$STAGE_DIR" > "$TMP/stage-parent"
+    ls -A "$IN" > "$TMP/input-during"
+}
+EOF
+
+# The output is given relative to the repo root, as make passes it.
+OUT="$REPO/build/microvm/lambda-testlink-runtime"
+if ! (cd "$REPO" && PATH="$TMP/bin:$PATH" bash scripts/build-lambda-runtime-microvm-image.sh bin/agent-testlink build/microvm/lambda-testlink-runtime) > "$TMP/log" 2>&1; then
+    fail "build into output dir: $(tr '\n' ' ' < "$TMP/log")"
+else
+    if [ "$(cat "$TMP/stage-parent")" = "$OUT" ]; then
+        pass "stage created in the output directory"
+    else
+        fail "stage created in $(cat "$TMP/stage-parent"), want $OUT"
+    fi
+    if grep -q '^\.stage\.' "$TMP/input-during"; then
+        fail "stage present in the default input directory during the build"
+    else
+        pass "no stage in the default input directory during the build"
+    fi
+    if [ "$(cd "$IN" && ls -A | cksum)" = "$BEFORE" ] && grep -qx 'default image' "$IN/vmlinuz"; then
+        pass "default input directory unchanged"
+    else
+        fail "default input directory changed: $(cd "$IN" && ls -A | tr '\n' ' ')"
+    fi
+    if grep -qx kernel "$OUT/vmlinuz" && gzip -t "$OUT/initramfs.cpio.gz" && [ "$(ls -A "$OUT" | wc -l)" -eq 2 ]; then
+        pass "output directory holds exactly the published pair"
+    else
+        fail "output directory: $(ls -A "$OUT" | tr '\n' ' ')"
+    fi
+fi
+
+if [ "$fails" -ne 0 ]; then
+    echo "$fails failure(s)"
+    exit 1
+fi

scripts/build-qboot.sh

@@ -9,7 +9,8 @@
 #
 # The build runs in a digest-pinned Debian container, so its output is
 # reproducible and is checked against a recorded hash. That hash is also the
-# sha256 of Debian 13's /usr/share/qemu/qboot.rom, which is built from the same commit.
+# sha256 of Debian 13's /usr/share/qemu/qboot.rom, which is built from the same
+# commit, so a host whose copy matches installs it and skips the container.
 set -euo pipefail
 
 REPO_ROOT="$(cd "$(dirname "$0")/.." && pwd)"
@@ -20,18 +21,13 @@ QBOOT_REPO="${QBOOT_REPO:-https://gitlab.com/qemu-project/qboot.git}"
 QBOOT_COMMIT="8ca302e86d685fa05b16e2b208888243da319941"
 QBOOT_SHA256="5c9c4b12c4438e3cf11d087263afe6519c8c05dd5896727b8b169b8f5dbf9877"
 BUILD_IMAGE="debian:trixie@sha256:4c25275446ca371766b8b093c696782eabce06263fccacb8d5d813b3dc44b308"
+# Debian 13's qemu-system-data ships this exact ROM, so a host that has it
+# skips the container. QBOOT_HOST_ROM moves the path for tests.
+QBOOT_HOST_ROM="${QBOOT_HOST_ROM:-/usr/share/qemu/qboot.rom}"
 
 log() { echo "[build-qboot] $*"; }
 fail() { echo "ERROR: $*" >&2; exit 1; }
 
-CONTAINER_TOOL=""
-for tool in docker podman; do
-    if command -v "$tool" >/dev/null 2>&1 && "$tool" info >/dev/null 2>&1; then
-        CONTAINER_TOOL=$tool
-        break
-    fi
-done
-[ -n "$CONTAINER_TOOL" ] || fail "qboot builds in a container; install docker or podman"
 command -v git >/dev/null 2>&1 || fail "required tool not found: git"
 
 WORK=$(mktemp -d)
@@ -48,19 +44,42 @@ mkdir -p "$SRC_DIR"
 src_tarball="$SRC_DIR/qboot-${QBOOT_COMMIT}.tar.gz"
 git -C "$WORK/src" archive --format=tar.gz --prefix="qboot-${QBOOT_COMMIT}/" -o "$src_tarball" "$QBOOT_COMMIT"
 
-log "building in $BUILD_IMAGE with $CONTAINER_TOOL"
-# The firmware comes back on stdout, so nothing the container writes is left
-# root-owned on the host.
-"$CONTAINER_TOOL" run --rm --platform linux/amd64 -v "$WORK/src:/src:ro" \
-    -e DEBIAN_FRONTEND=noninteractive "$BUILD_IMAGE" sh -euc '
-        {
-            apt-get update -qq
-            apt-get install -y -qq --no-install-recommends gcc libc6-dev meson ninja-build binutils
-            meson setup /tmp/build /src
-            ninja -C /tmp/build
-        } >/dev/null
-        cat /tmp/build/bios.bin
-    ' > "$WORK/qboot.rom"
+build_in_container() {
+    local CONTAINER_TOOL=""
+    for tool in docker podman; do
+        if command -v "$tool" >/dev/null 2>&1 && "$tool" info >/dev/null 2>&1; then
+            CONTAINER_TOOL=$tool
+            break
+        fi
+    done
+    [ -n "$CONTAINER_TOOL" ] || fail "qboot builds in a container; install docker or podman"
+    log "building in $BUILD_IMAGE with $CONTAINER_TOOL"
+    # The firmware comes back on stdout, so nothing the container writes is left
+    # root-owned on the host.
+    "$CONTAINER_TOOL" run --rm --platform linux/amd64 -v "$WORK/src:/src:ro" \
+        -e DEBIAN_FRONTEND=noninteractive "$BUILD_IMAGE" sh -euc '
+            {
+                apt-get update -qq
+                apt-get install -y -qq --no-install-recommends gcc libc6-dev meson ninja-build binutils
+                meson setup /tmp/build /src
+                ninja -C /tmp/build
+            } >/dev/null
+            cat /tmp/build/bios.bin
+        ' > "$WORK/qboot.rom"
+}
+
+host_sha=""
+[ -r "$QBOOT_HOST_ROM" ] && host_sha=$(sha256sum "$QBOOT_HOST_ROM" | cut -d' ' -f1)
+if [ "$host_sha" = "$QBOOT_SHA256" ]; then
+    log "using host ROM $QBOOT_HOST_ROM (sha256 matches)"
+    cp "$QBOOT_HOST_ROM" "$WORK/qboot.rom"
+elif [ -n "$host_sha" ]; then
+    log "host ROM $QBOOT_HOST_ROM has sha256 $host_sha, not $QBOOT_SHA256; building"
+    build_in_container
+else
+    log "no host ROM at $QBOOT_HOST_ROM; building"
+    build_in_container
+fi
 
 got_sha=$(sha256sum "$WORK/qboot.rom" | cut -d' ' -f1)
 if [ "$got_sha" != "$QBOOT_SHA256" ]; then

scripts/build-qboot_test.sh

@@ -0,0 +1,88 @@
+#!/bin/sh
+# build-qboot_test.sh — a host ROM matching the recorded hash is installed
+# without the container; any other host ROM, or none, builds in it, and a
+# container ROM with the wrong hash still fails.
+set -eu
+
+SCRIPTS_DIR=$(CDPATH='' cd -- "$(dirname -- "$0")" && pwd)
+TMP=$(mktemp -d)
+trap 'rm -rf "$TMP"' EXIT
+
+fails=0
+pass() { echo "  ok   - $1"; }
+fail() { echo "  FAIL - $1"; fails=$((fails + 1)); }
+
+echo "-- build-qboot.sh host ROM"
+
+COMMIT=$(sed -n 's/^QBOOT_COMMIT="\(.*\)"$/\1/p' "$SCRIPTS_DIR/build-qboot.sh")
+printf 'pinned rom\n' > "$TMP/good.rom"
+printf 'other rom\n' > "$TMP/other.rom"
+GOOD_SHA=$(sha256sum "$TMP/good.rom" | cut -d' ' -f1)
+
+# git and docker stubs. git fakes a checkout of the pinned commit; docker
+# prints $STUB_ROM as the firmware and records that it ran.
+mkdir -p "$TMP/bin"
+cat > "$TMP/bin/git" <<STUB
+#!/bin/sh
+case "\$1" in
+init) mkdir -p "\$3"; printf 'GPL-2.0\n' > "\$3/LICENSE"; exit 0 ;;
+esac
+shift 2
+case "\$1" in
+rev-parse) echo "$COMMIT" ;;
+archive) while [ "\$1" != "-o" ]; do shift; done; printf 'source\n' > "\$2" ;;
+esac
+STUB
+cat > "$TMP/bin/docker" <<'STUB'
+#!/bin/sh
+[ "$1" = info ] && exit 0
+echo ran >> "$DOCKER_LOG"
+cat "$STUB_ROM"
+STUB
+chmod +x "$TMP/bin/git" "$TMP/bin/docker"
+
+# run <name> <host-rom> <container-rom>: build in a fresh fixture repo.
+run() {
+    repo="$TMP/$1"
+    mkdir -p "$repo/scripts"
+    sed "s/^QBOOT_SHA256=.*/QBOOT_SHA256=\"$GOOD_SHA\"/" "$SCRIPTS_DIR/build-qboot.sh" > "$repo/scripts/build-qboot.sh"
+    : > "$TMP/$1.docker"
+    PATH="$TMP/bin:$PATH" QBOOT_HOST_ROM="$2" STUB_ROM="$3" DOCKER_LOG="$TMP/$1.docker" \
+        bash "$repo/scripts/build-qboot.sh" > "$TMP/$1.log" 2>&1
+}
+
+# outputs <name>: the ROM is the pinned one and the licence and source exist.
+outputs() {
+    cmp -s "$TMP/$1/build/microvm/qboot.rom" "$TMP/good.rom" &&
+        [ -s "$TMP/$1/build/microvm/qboot.LICENSE" ] &&
+        [ -s "$TMP/$1/build/qboot/qboot-$COMMIT.tar.gz" ]
+}
+
+if run match "$TMP/good.rom" "$TMP/other.rom" && outputs match && [ ! -s "$TMP/match.docker" ]; then
+    pass "matching host ROM installed, container not run"
+else
+    fail "matching host ROM: $(tr '\n' ' ' < "$TMP/match.log")"
+fi
+
+if run mismatch "$TMP/other.rom" "$TMP/good.rom" && outputs mismatch && [ -s "$TMP/mismatch.docker" ] &&
+    grep -q 'has sha256' "$TMP/mismatch.log"; then
+    pass "mismatched host ROM builds in the container"
+else
+    fail "mismatched host ROM: $(tr '\n' ' ' < "$TMP/mismatch.log")"
+fi
+
+if run missing "$TMP/absent.rom" "$TMP/good.rom" && outputs missing && [ -s "$TMP/missing.docker" ]; then
+    pass "missing host ROM builds in the container"
+else
+    fail "missing host ROM: $(tr '\n' ' ' < "$TMP/missing.log")"
+fi
+
+if run badbuild "$TMP/absent.rom" "$TMP/other.rom"; then
+    fail "container ROM with the wrong hash was accepted"
+elif [ -e "$TMP/badbuild/build/microvm/qboot.rom" ]; then
+    fail "container ROM with the wrong hash was installed"
+else
+    pass "container ROM with the wrong hash fails and installs nothing"
+fi
+
+[ "$fails" -eq 0 ] || { echo "$fails failure(s)"; exit 1; }

scripts/images/ecs-agent-gpu/ecs-agent.service

@@ -9,8 +9,7 @@ Wants=network-online.target
 
 [Service]
 Type=simple
-# Static config (gateway URL, CA, region, cluster, seeded IAM creds,
-# containerd socket) comes from /etc/spinifex-ecs/agent.env, dropped by
+# Static config (gateway URL, CA, region, cluster, containerd socket) comes from /etc/spinifex-ecs/agent.env, dropped by
 # cloud-init at launch. Absent file is tolerated — the agent falls back to
 # its built-in defaults.
 EnvironmentFile=-/etc/spinifex-ecs/agent.env

scripts/images/ecs-agent/ecs-agent.initd

@@ -3,8 +3,7 @@
 # IMDSv2, registers + heartbeats + polls task assignments through the AWS gateway
 # over TLS+SigV4 (no NATS), and pulls task images through containerd.
 #
-# Static config (gateway URL, CA, region, cluster, seeded IAM creds, containerd
-# socket) comes from /etc/spinifex-ecs/agent.env, dropped by cloud-init at launch.
+# Static config (gateway URL, CA, region, cluster, containerd socket) comes from /etc/spinifex-ecs/agent.env, dropped by cloud-init at launch.
 
 description="Spinifex ECS container-instance agent"
 command="/usr/local/bin/ecs-agent"
@@ -28,7 +27,7 @@ start_pre() {
         # shellcheck disable=SC1091
         . /etc/spinifex-ecs/agent.env
         export ECS_GATEWAY_URL ECS_GATEWAY_CA ECS_REGION ECS_CLUSTER \
-            ECS_ACCESS_KEY ECS_SECRET_KEY ECS_IMDS_BASE ECS_CONTAINERD_SOCKET \
+            ECS_IMDS_BASE ECS_CONTAINERD_SOCKET \
             ECS_HEARTBEAT_INTERVAL ECS_POLL_INTERVAL
     fi
 }

scripts/images/eks-node/eks-node-role.sh

@@ -159,6 +159,7 @@ case "${ROLE}" in
         svc_start mulga-eks-state-report
         svc_start mulga-eks-addon-sync
         svc_start mulga-eks-node-gc
+        svc_optional mulga-eks-log-forward
         svc_optional eks-pod-identity-webhook
         ;;
     server-join)
@@ -178,12 +179,14 @@ case "${ROLE}" in
         svc_start konnectivity-server
         svc_start mulga-eks-state-report
         svc_start mulga-eks-node-gc
+        svc_optional mulga-eks-log-forward
         svc_optional eks-pod-identity-webhook
         ;;
     agent)
         log "configuring agent role"
         svc_enable k3s-agent
         svc_start k3s-agent
+        svc_optional mulga-eks-log-forward
         ;;
     *)
         die "SPINIFEX_K3S_ROLE missing or invalid: '${ROLE}'"

scripts/images/eks-node/eks-node-role_test.sh

@@ -146,6 +146,15 @@ for INIT in openrc systemd; do
     )"
     check "server(${INIT}): enables node-gc" "yes" "$(present "${CALLS}" "$(enable_cmd "${INIT}" mulga-eks-node-gc)")"
     check "server(${INIT}): starts node-gc" "yes" "$(present "${CALLS}" "$(start_cmd "${INIT}" mulga-eks-node-gc)")"
+    check "server(${INIT}): enables log-forward" "yes" "$(present "${CALLS}" "$(enable_cmd "${INIT}" mulga-eks-log-forward)")"
+    check "server(${INIT}): starts log-forward" "yes" "$(present "${CALLS}" "$(start_cmd "${INIT}" mulga-eks-log-forward)")"
+    check "server(${INIT}): starts log-forward before the webhook" "yes" "$(
+        awk -v f="$(start_cmd "${INIT}" mulga-eks-log-forward)" -v w="$(start_cmd "${INIT}" eks-pod-identity-webhook)" '
+            index($0, f) == 1 { fn = NR }
+            index($0, w) == 1 { wn = NR }
+            END { print (fn > 0 && wn > fn) ? "yes" : "no" }
+        ' "${CALLS}"
+    )"
     check "server(${INIT}): no k3s-agent" "no" "$(present "${CALLS}" "$(enable_cmd "${INIT}" k3s-agent)")"
     check "server(${INIT}): self-disables" "yes" "$(present "${CALLS}" "$(disable_cmd "${INIT}" eks-node-role)")"
     check "server(${INIT}): pod-identity-webhook is the last call" "$(start_cmd "${INIT}" eks-pod-identity-webhook)" "$(grep -v "$(disable_cmd "${INIT}" eks-node-role)" "${CALLS}" | tail -n 1)"
@@ -196,6 +205,15 @@ for INIT in openrc systemd; do
     # published the cluster-identical artifacts, and a join re-publish would
     # only race the bootstrap bus. This is deliberate, not an oversight.
     check "server-join(${INIT}): pod-identity-webhook is the last call" "$(start_cmd "${INIT}" eks-pod-identity-webhook)" "$(grep -v "$(disable_cmd "${INIT}" eks-node-role)" "${CALLS}" | tail -n 1)"
+    check "server-join(${INIT}): enables log-forward" "yes" "$(present "${CALLS}" "$(enable_cmd "${INIT}" mulga-eks-log-forward)")"
+    check "server-join(${INIT}): starts log-forward" "yes" "$(present "${CALLS}" "$(start_cmd "${INIT}" mulga-eks-log-forward)")"
+    check "server-join(${INIT}): starts log-forward before the webhook" "yes" "$(
+        awk -v f="$(start_cmd "${INIT}" mulga-eks-log-forward)" -v w="$(start_cmd "${INIT}" eks-pod-identity-webhook)" '
+            index($0, f) == 1 { fn = NR }
+            index($0, w) == 1 { wn = NR }
+            END { print (fn > 0 && wn > fn) ? "yes" : "no" }
+        ' "${CALLS}"
+    )"
     check "server-join(${INIT}): no k3s-first-boot" "yes" "$(absent_all "${CALLS}" "k3s-first-boot")"
     check "server-join(${INIT}): no k3s-agent" "yes" "$(absent_all "${CALLS}" "$(enable_cmd "${INIT}" k3s-agent)")"
     check "server-join(${INIT}): self-disables" "yes" "$(present "${CALLS}" "$(disable_cmd "${INIT}" eks-node-role)")"
@@ -207,8 +225,39 @@ for INIT in openrc systemd; do
     check "agent(${INIT}): exit 0" 0 "${rc}"
     check "agent(${INIT}): role file" "agent" "$(cat "${ROLE_FILE}")"
     check "agent(${INIT}): enables k3s-agent" "yes" "$(present "${CALLS}" "$(enable_cmd "${INIT}" k3s-agent)")"
+    check "agent(${INIT}): enables log-forward" "yes" "$(present "${CALLS}" "$(enable_cmd "${INIT}" mulga-eks-log-forward)")"
+    check "agent(${INIT}): starts log-forward" "yes" "$(present "${CALLS}" "$(start_cmd "${INIT}" mulga-eks-log-forward)")"
+    check "agent(${INIT}): log-forward is enabled after k3s-agent" "yes" "$(
+        awk -v k="$(enable_cmd "${INIT}" k3s-agent)" -v f="$(enable_cmd "${INIT}" mulga-eks-log-forward)" '
+            index($0, k) == 1 { kn = NR }
+            index($0, f) == 1 { fn = NR }
+            END { print (kn > 0 && fn > kn) ? "yes" : "no" }
+        ' "${CALLS}"
+    )"
+
     check "agent(${INIT}): no server services" "yes" "$(absent_all "${CALLS}" "eks-token-webhook" "eks-pod-identity-webhook" "k3s-first-boot" "mulga-eks-node-gc" "$(enable_cmd "${INIT}" k3s)")"
 
+    # A broken log forwarder must never stop a node joining.
+    for FAILING in enable start; do
+        run_case "${INIT}" agent_logfwd_${FAILING}_fail
+        printf 'SPINIFEX_K3S_ROLE=agent\n' > "${EKS_NODE_ENVFILE}"
+        if [ "${FAILING}" = enable ]; then
+            STUB_FAIL="$(enable_cmd "${INIT}" mulga-eks-log-forward)"
+            export STUB_FAIL
+            STUB_FAIL="${STUB_FAIL#rc-update add }"; STUB_FAIL="${STUB_FAIL#systemctl enable }"
+        else
+            STUB_FAIL="$(start_cmd "${INIT}" mulga-eks-log-forward)"
+            export STUB_FAIL
+            STUB_FAIL="${STUB_FAIL#rc-service }"; STUB_FAIL="${STUB_FAIL#systemctl start --no-block }"
+        fi
+        invoke
+        unset STUB_FAIL
+        check "agent(${INIT}): log-forward ${FAILING} failure still exits 0" 0 "${rc}"
+        check "agent(${INIT}): log-forward ${FAILING} failure still enables k3s-agent" "yes" "$(present "${CALLS}" "$(enable_cmd "${INIT}" k3s-agent)")"
+        check "agent(${INIT}): log-forward ${FAILING} failure still starts k3s-agent" "yes" "$(present "${CALLS}" "$(start_cmd "${INIT}" k3s-agent)")"
+        check "agent(${INIT}): log-forward ${FAILING} failure still self-disables" "yes" "$(present "${CALLS}" "$(disable_cmd "${INIT}" eks-node-role)")"
+    done
+
     # --- agent (inferred from agent.env, no explicit role) ---
     run_case "${INIT}" agent_inferred
     printf 'K3S_URL=https://nlb:443\nK3S_TOKEN=abc\n' > "${EKS_NODE_AGENT_ENVFILE}"

scripts/images/eks-node/manifest.conf

@@ -45,6 +45,8 @@ scripts/images/eks-node/mulga-eks-addon-sync.initd:/etc/init.d/mulga-eks-addon-s
 scripts/images/eks-node/mulga-eks-addon-sync.sh:/usr/local/sbin/mulga-eks-addon-sync \
 scripts/images/eks-node/mulga-eks-node-gc.initd:/etc/init.d/mulga-eks-node-gc \
 scripts/images/eks-node/mulga-eks-node-gc.sh:/usr/local/sbin/mulga-eks-node-gc \
+scripts/images/eks-node/mulga-eks-log-forward.initd:/etc/init.d/mulga-eks-log-forward \
+scripts/images/eks-node/mulga-eks-log-forward.sh:/usr/local/sbin/mulga-eks-log-forward \
 scripts/images/eks-node/konnectivity-server.initd:/etc/init.d/konnectivity-server \
 scripts/images/eks-node/konnectivity-server-prestart.sh:/usr/local/sbin/konnectivity-server-prestart \
 scripts/images/eks-node/konnectivity/konnectivity-agent.yaml:/usr/share/spinifex-eks/konnectivity/konnectivity-agent.yaml \

scripts/images/eks-node/mulga-eks-log-forward.initd

@@ -0,0 +1,30 @@
+#!/sbin/openrc-run
+# mulga-eks-log-forward — copies k3s warnings and errors to the serial console
+# so they reach the Elasticsearch sink. Best effort: nothing depends on it and
+# a failure here never affects the node.
+
+description="Forward k3s warnings and errors to the serial console"
+command="/bin/sh"
+command_background="yes"
+supervisor="supervise-daemon"
+respawn_delay=5
+pidfile="/run/mulga-eks-log-forward.pid"
+output_log="/dev/null"
+error_log="/dev/null"
+
+depend() {
+    after k3s k3s-agent
+}
+
+start_pre() {
+    role=$(cat /etc/spinifex-eks/role 2>/dev/null || true)
+    if [ "${role}" = "agent" ]; then
+        log=/var/log/k3s-agent.log
+        FORWARD_PREFIX="[k3s-agent]"
+    else
+        log=/var/log/k3s.log
+        FORWARD_PREFIX="[k3s]"
+    fi
+    export FORWARD_PREFIX
+    command_args="-c 'tail -n +1 -F ${log} 2>/dev/null | /usr/local/sbin/mulga-eks-log-forward'"
+}

scripts/images/eks-node/mulga-eks-log-forward.sh

@@ -0,0 +1,90 @@
+#!/bin/sh
+# mulga-eks-log-forward — copy the warnings and errors of a k3s log to the
+# serial console, where the host picks them up for the Elasticsearch sink.
+#
+# Reads the tailed log on stdin. Lines that matter are redacted first, then
+# cut to FORWARD_MAX_BYTES, then rate limited: at most FORWARD_RATE_PER_MIN a
+# minute reach the console and the rest are reported as one "N lines dropped"
+# line. Every console write has a timeout, so a stuck console loses lines
+# instead of stalling the tail or k3s.
+#
+# Env: FORWARD_CONSOLE (default /dev/console), FORWARD_PREFIX, FORWARD_MAX_BYTES
+# (512), FORWARD_RATE_PER_MIN (60), FORWARD_WRITE_TIMEOUT seconds (2) and
+# FORWARD_CLOCK_CMD (default: seconds of uptime, read without forking).
+
+CONSOLE=${FORWARD_CONSOLE:-/dev/console}
+PREFIX=${FORWARD_PREFIX:-[k3s]}
+MAX_BYTES=${FORWARD_MAX_BYTES:-512}
+RATE=${FORWARD_RATE_PER_MIN:-60}
+WRITE_TIMEOUT=${FORWARD_WRITE_TIMEOUT:-2}
+CLOCK_CMD=${FORWARD_CLOCK_CMD:-}
+
+now() {
+    if [ -n "${CLOCK_CMD}" ]; then
+        sh -c "${CLOCK_CMD}"
+    else
+        read -r up _ < /proc/uptime && echo "${up%%.*}"
+    fi
+}
+
+# emit TEXT: one console line, abandoned after WRITE_TIMEOUT. Returns non-zero
+# when the console did not take it.
+emit() {
+    timeout "${WRITE_TIMEOUT}" sh -c 'printf "%s\n" "$1" >> "$2"' sh "$1" "${CONSOLE}" 2>/dev/null
+}
+
+# Filter, redact and truncate in one awk pass so lines that do not matter cost
+# no fork. Redaction runs before the cut so a secret straddling the limit
+# cannot leave a readable prefix behind.
+select_lines() {
+    LC_ALL=C awk -v max="${MAX_BYTES}" '
+    {
+        line = $0
+        gsub(/\r/, "", line)
+        keep = 0
+        if (line ~ /level=(warning|error)/) keep = 1
+        else if (line ~ /^[WE][0-9][0-9][0-9][0-9] /) keep = 1
+        else if (line ~ /NodeNotReady|PLEG|[Ll]ease|[Nn]ode status/) keep = 1
+        if (!keep) next
+
+        gsub(/[Bb]earer +[A-Za-z0-9._~+\/=-]+/, "Bearer [REDACTED]", line)
+        gsub(/K10[0-9a-f]+(::[A-Za-z0-9-]+:[^ "\047]+)?/, "[REDACTED]", line)
+        gsub(/eyJ[A-Za-z0-9_.-]+/, "[REDACTED]", line)
+        gsub(/(AKIA|ASIA)[A-Z0-9]+/, "[REDACTED]", line)
+        gsub(/X-Amz-Signature=[^& "\047]*/, "X-Amz-Signature=[REDACTED]", line)
+        gsub(/X-Amz-Credential=[^& "\047]*/, "X-Amz-Credential=[REDACTED]", line)
+        gsub(/X-Amz-Security-Token=[^& "\047]*/, "X-Amz-Security-Token=[REDACTED]", line)
+        gsub(/[^[:print:]\t]/, " ", line)
+        print substr(line, 1, max)
+    }'
+}
+
+win_start=0
+sent=0
+dropped=0
+
+flush_dropped() {
+    [ "${dropped}" -gt 0 ] || return 0
+    emit "${PREFIX} ${dropped} lines dropped"
+    dropped=0
+}
+
+select_lines | {
+    while IFS= read -r line; do
+        t=$(now) || t=0
+        if [ "${win_start}" -eq 0 ] || [ $((t - win_start)) -ge 60 ]; then
+            flush_dropped
+            win_start=${t}
+            sent=0
+            [ "${win_start}" -eq 0 ] && win_start=1
+        fi
+        if [ "${sent}" -ge "${RATE}" ]; then
+            dropped=$((dropped + 1))
+            continue
+        fi
+        sent=$((sent + 1))
+        emit "${PREFIX} ${line}" || dropped=$((dropped + 1))
+    done
+    flush_dropped
+}
+exit 0

scripts/images/eks-node/mulga-eks-log-forward_test.sh

@@ -0,0 +1,190 @@
+#!/bin/sh
+# Self-contained POSIX test for mulga-eks-log-forward.sh: the line filter,
+# redaction, truncation, rate cap and blocked-console behaviour. The script
+# reads the tailed log on stdin and writes to FORWARD_CONSOLE, which the test
+# points at a temp file (or a reader-less FIFO), so no root or tty is needed.
+#
+# Run: sh scripts/images/eks-node/mulga-eks-log-forward_test.sh
+set -eu
+
+SCRIPT_DIR=$(cd "$(dirname "$0")" && pwd)
+SCRIPT="${SCRIPT_DIR}/mulga-eks-log-forward.sh"
+WORK=$(mktemp -d)
+trap 'rm -rf "${WORK}"' EXIT
+
+PASS=0
+FAIL=0
+check() { # check <desc> <expected> <actual>
+    if [ "$2" = "$3" ]; then
+        PASS=$((PASS + 1))
+    else
+        FAIL=$((FAIL + 1))
+        echo "FAIL: $1"
+        echo "  expected: [$2]"
+        echo "  actual:   [$3]"
+    fi
+}
+
+# forward <name>: feeds stdin to the script with a per-case console file and a
+# fixed clock, leaving the forwarded text in ${OUT}.
+forward() {
+    CASE="${WORK}/$1"
+    mkdir -p "${CASE}"
+    OUT="${CASE}/console"
+    : > "${OUT}"
+    FORWARD_CONSOLE="${OUT}" FORWARD_PREFIX="[k3s-agent]" \
+        FORWARD_CLOCK_CMD="echo 1000" sh "${SCRIPT}" 2>/dev/null || true
+}
+forwarded() { grep -c . "${OUT}" || true; }
+
+# --- filter: lines that must pass ---
+forward filter_pass <<'EOF'
+time="2026-11-03T03:00:00Z" level=warning msg="slow response from apiserver"
+time="2026-11-03T03:00:01Z" level=error msg="failed to renew"
+W1103 03:00:02.123456    1234 reflector.go:561] watch of *v1.Pod ended
+E1103 03:00:03.123456    1234 kubelet.go:2400] node not found
+I1103 03:00:04.123456    1234 kubelet_node_status.go:73] Node status update failed
+I1103 03:00:05.123456    1234 controller.go:193] Failed to update lease
+I1103 03:00:06.123456    1234 kubelet.go:2200] PLEG is not healthy
+time="2026-11-03T03:00:07Z" level=info msg="Event: NodeNotReady kubelet"
+EOF
+check "filter: all eight interesting lines pass" "8" "$(forwarded)"
+check "filter: lines carry the prefix" "8" "$(grep -c '^\[k3s-agent\] ' "${OUT}" || true)"
+
+# --- filter: lines that must be dropped ---
+forward filter_drop <<'EOF'
+time="2026-11-03T03:00:00Z" level=info msg="Starting k3s agent"
+time="2026-11-03T03:00:01Z" level=debug msg="tick"
+I1103 03:00:02.123456    1234 server.go:100] Version: v1.32.5
+I1103 03:00:03.123456    1234 kubelet.go:1000] SyncLoop (UPDATE, "api")
+EOF
+check "filter: info and debug lines are dropped" "0" "$(forwarded)"
+
+# --- redaction: one case per secret shape ---
+redact() { # redact <name> <line> <secret-that-must-not-appear>
+    forward "redact_$1" <<EOF
+time="2026-11-03T03:00:00Z" level=error msg="$2"
+EOF
+    check "redact(${1}): line still forwarded" "1" "$(forwarded)"
+    if grep -qF "$3" "${OUT}"; then
+        check "redact(${1}): secret removed" "absent" "present"
+    else
+        check "redact(${1}): secret removed" "absent" "absent"
+    fi
+    check "redact(${1}): marker present" "1" "$(grep -c 'REDACTED' "${OUT}" || true)"
+}
+redact bearer "request failed Authorization: Bearer abc123SECRETvalue456" "abc123SECRETvalue456"
+redact k3s_token "join failed token K1009f3c2d4e5a6b7c8d9e0f1a2b3c4d5e6f7a8b9c0d1e2f3a4b5c6d7e8f9a0b1c2::server:s3cr3tPassw0rd" "s3cr3tPassw0rd"
+redact jwt "bad token eyJhbGciOiJIUzI1NiJ9.eyJzdWIiOiIxMjM0NTY3ODkwIn0.c2lnbmF0dXJlLXNlY3JldA" "eyJhbGciOiJIUzI1NiJ9"
+redact akia "credentials AKIAIOSFODNN7EXAMPLE rejected" "AKIAIOSFODNN7EXAMPLE"
+redact asia "credentials ASIAIOSFODNN7EXAMPLE rejected" "ASIAIOSFODNN7EXAMPLE"
+redact amz_signature "GET /obj?X-Amz-Signature=0123456789abcdef0123456789abcdef&X-Amz-Expires=60 failed" "0123456789abcdef0123456789abcdef"
+redact amz_credential "GET /obj?X-Amz-Credential=ZZZsecretscope%2F20261103%2Fap%2Fs3%2Faws4_request failed" "ZZZsecretscope"
+
+# --- truncation: text over 512 bytes is cut ---
+LONG=$(awk 'BEGIN { for (i = 0; i < 2000; i++) printf "a"; }')
+forward truncate <<EOF
+time="2026-11-03T03:00:00Z" level=error msg="${LONG}"
+EOF
+check "truncate: still forwarded" "1" "$(forwarded)"
+BODY=$(sed 's/^\[k3s-agent\] //' "${OUT}" | head -n 1)
+check "truncate: body is exactly 512 bytes" "512" "$(printf '%s' "${BODY}" | wc -c | tr -d ' ')"
+check "truncate: keeps the start of the line" "time=\"2026-11-03T03:00:00Z\"" "$(printf '%s' "${BODY}" | cut -c1-27)"
+
+# --- redaction runs before the cut: a secret over the limit leaves no prefix ---
+# The pad puts the 512-byte cut right after the bare marker (AKIA, eyJ, K10).
+# A cut fragment that long matches no redaction pattern, so redacting after the
+# cut would leave it in the output.
+HEAD='time="2026-11-03T03:00:00Z" level=error msg="'
+straddle() { # straddle <name> <secret> <bytes of the secret kept by the cut>
+    PAD=$(awk -v n=$((512 - ${#HEAD} - 1 - $3)) 'BEGIN { for (i = 0; i < n; i++) printf "x"; }')
+    forward "straddle_$1" <<EOF
+${HEAD}${PAD} $2 tail"
+EOF
+    check "straddle(${1}): line forwarded" "1" "$(forwarded)"
+    check "straddle(${1}): cut keeps the marker length" "$3" "$(sed 's/^\[k3s-agent\] //' "${OUT}" | head -n 1 | awk -v p="${PAD} " '{ print length($0) - index($0, p) - length(p) + 1 }')"
+    if grep -qF "$(printf '%s' "$2" | cut -c1-"$3")" "${OUT}"; then
+        check "straddle(${1}): no fragment survives the cut" "absent" "present"
+    else
+        check "straddle(${1}): no fragment survives the cut" "absent" "absent"
+    fi
+}
+straddle jwt "eyJhbGciOiJIUzI1NiJ9.eyJzdWIiOiIxMjM0In0.c2ln" 3
+straddle akia "AKIAIOSFODNN7EXAMPLE" 4
+straddle k3s_token "K1009f3c2d4e5a6b7c8d9e0f1a2b3c4d5e6f7a8b9c0d1e2f3a4b5c6d7e8f9a0b1c2::server:s3cr3tPassw0rd" 3
+
+# --- rate cap: 60 a minute, then one summary of what was dropped ---
+gen() { # gen <count>
+    awk -v n="$1" 'BEGIN { for (i = 0; i < n; i++) printf "time=\"t\" level=error msg=\"loop %d\"\n", i }'
+}
+gen 100 > "${WORK}/burst"
+forward rate_cap < "${WORK}/burst"
+check "rate cap: 60 lines forwarded plus one summary" "61" "$(forwarded)"
+check "rate cap: one dropped summary" "1" "$(grep -c '40 lines dropped' "${OUT}" || true)"
+check "rate cap: first line kept" "1" "$(grep -c 'loop 0"' "${OUT}" || true)"
+check "rate cap: line past the cap is not forwarded" "0" "$(grep -c 'loop 60"' "${OUT}" || true)"
+
+# --- rate cap: a new minute starts a new window and reports the old drops ---
+# The clock command advances 60 s every 70 calls, so lines 0-69 share a window
+# (10 dropped) and lines 70-74 land in the next one.
+CASE="${WORK}/rate_window"
+mkdir -p "${CASE}"
+OUT="${CASE}/console"
+: > "${OUT}"
+echo 0 > "${CASE}/calls"
+cat > "${CASE}/clock" <<EOF
+#!/bin/sh
+n=\$(cat "${CASE}/calls")
+echo \$((n + 1)) > "${CASE}/calls"
+echo \$((1000 + n / 70 * 60))
+EOF
+chmod +x "${CASE}/clock"
+gen 75 | FORWARD_CONSOLE="${OUT}" FORWARD_PREFIX="[k3s-agent]" FORWARD_CLOCK_CMD="${CASE}/clock" sh "${SCRIPT}" 2>/dev/null || true
+check "window: 60 + 5 forwarded plus one summary" "66" "$(forwarded)"
+check "window: summary names the 10 dropped" "1" "$(grep -c '10 lines dropped' "${OUT}" || true)"
+check "window: new window forwards its lines" "1" "$(grep -c 'loop 74"' "${OUT}" || true)"
+
+# --- blocked console: lines are dropped, the forwarder neither stalls nor fails ---
+CASE="${WORK}/blocked"
+mkdir -p "${CASE}"
+mkfifo "${CASE}/console"
+gen 3 > "${CASE}/in"
+(
+    FORWARD_CONSOLE="${CASE}/console" FORWARD_PREFIX="[k3s-agent]" FORWARD_CLOCK_CMD="echo 1000" \
+        FORWARD_WRITE_TIMEOUT=1 sh "${SCRIPT}" < "${CASE}/in"
+) &
+PID=$!
+i=0
+while kill -0 "${PID}" 2>/dev/null && [ "${i}" -lt 20 ]; do
+    sleep 1
+    i=$((i + 1))
+done
+if kill -0 "${PID}" 2>/dev/null; then
+    kill "${PID}" 2>/dev/null || true
+    check "blocked console: forwarder finished" "finished" "stalled"
+else
+    check "blocked console: forwarder finished" "finished" "finished"
+fi
+rc=0
+wait "${PID}" || rc=$?
+check "blocked console: exit 0" "0" "${rc}"
+
+# --- the service is wired in, and nothing depends on it ---
+INITD="${SCRIPT_DIR}/mulga-eks-log-forward.initd"
+check "initd ships" "yes" "$([ -f "${INITD}" ] && echo yes || echo no)"
+check "manifest installs the initd" "1" "$(grep -c 'mulga-eks-log-forward.initd:/etc/init.d/mulga-eks-log-forward' "${SCRIPT_DIR}/manifest.conf" || true)"
+check "manifest installs the script" "1" "$(grep -c 'mulga-eks-log-forward.sh:/usr/local/sbin/mulga-eks-log-forward' "${SCRIPT_DIR}/manifest.conf" || true)"
+DEPS=0
+for f in "${SCRIPT_DIR}"/*.initd; do
+    [ "${f}" = "${INITD}" ] && continue
+    if grep -E '^[[:space:]]*(need|use|want)[[:space:]].*mulga-eks-log-forward' "${f}" >/dev/null 2>&1; then
+        DEPS=$((DEPS + 1))
+    fi
+done
+check "nothing depends on the forwarder" "0" "${DEPS}"
+if [ -f "${INITD}" ]; then
+    check "initd does not need anything it could block boot on" "0" "$(grep -cE '^[[:space:]]*need[[:space:]]' "${INITD}" || true)"
+fi
+
+echo "mulga-eks-log-forward_test: ${PASS} passed, ${FAIL} failed"
+[ "${FAIL}" -eq 0 ]

scripts/images/eks-node/setup.sh

@@ -45,10 +45,10 @@ ln -sf /usr/local/bin/k3s /usr/local/bin/ctr
 chmod 0755 /etc/init.d/eks-node-role /etc/init.d/k3s /etc/init.d/k3s-agent \
     /etc/init.d/eks-token-webhook /etc/init.d/k3s-first-boot /etc/init.d/mulga-eks-state-report \
     /etc/init.d/mulga-eks-addon-sync /etc/init.d/mulga-eks-node-gc /etc/init.d/konnectivity-server \
-    /etc/init.d/eks-pod-identity-webhook
+    /etc/init.d/eks-pod-identity-webhook /etc/init.d/mulga-eks-log-forward
 chmod 0755 /usr/local/sbin/eks-node-role /usr/local/sbin/k3s-first-boot \
     /usr/local/sbin/mulga-eks-state-report /usr/local/sbin/mulga-eks-addon-sync \
-    /usr/local/sbin/mulga-eks-node-gc
+    /usr/local/sbin/mulga-eks-node-gc /usr/local/sbin/mulga-eks-log-forward
 # start_pre helpers shared with the systemd wrappers that will eventually
 # replace these runscripts.
 chmod 0755 /usr/local/sbin/k3s-prestart /usr/local/sbin/k3s-agent-prestart \

scripts/install-node.sh

@@ -260,6 +260,14 @@ out() {
     ssh "${SSH_OPTS[@]}" "$SSH_USER@$host" "$@" 2>/dev/null || true
 }
 
+# remote_join_cmd prints the command that joins one node. The token is not in it: it arrives on
+# stdin and is read after sudo has logged its command, so sudo's log never holds it. The other
+# values are positional arguments of a literal script, never spliced into it.
+remote_join_cmd() {
+    local script='exec spx admin join --force --node "$1" --bind "$2" --cluster-bind "$2" --advertise "$3" --host "$4" --token "$(cat)" --region "$5" --az "$6"'
+    printf 'sudo bash -c %q _ %q %q %q %q %q %q' "$script" "$@"
+}
+
 SETUP_OVN="/usr/local/share/spinifex/setup-ovn.sh"
 
 # --- Preflight -------------------------------------------------------------
@@ -652,14 +660,18 @@ fi
 [ -n "$EXT_IFACE" ] && init_args+=(--external-iface="$EXT_IFACE")
 [ -n "$EMAIL" ] && init_args+=(--email="$EMAIL")
 
-INIT_LOG=/tmp/spx-init.log
+# init prints the join token, so its log is root-only and outside the world-writable /tmp.
+INIT_LOG=/var/log/spx-init.log
+
+# redact_tokens hides join tokens in text about to be shown to the operator.
+redact_tokens() { sed -E 's/spx_join_[A-Za-z0-9_-]+/spx_join_[redacted]/g'; }
 
 log "initializing ${HOSTS[0]} as ${NODE_NAMES[0]} (waiting for $((N - 1)) node(s) to join)"
 
 # init blocks in a formation server until every node has joined, so it runs
 # detached and the token is read back out of its output.
 on "${HOSTS[0]}" "sudo rm -f $INIT_LOG; \
-    sudo setsid sh -c 'spx admin init ${init_args[*]} >$INIT_LOG 2>&1' </dev/null >/dev/null 2>&1 & \
+    sudo setsid sh -c 'umask 077; spx admin init ${init_args[*]} >$INIT_LOG 2>&1' </dev/null >/dev/null 2>&1 & \
     exit 0" || fail "${HOSTS[0]}: could not start spx admin init"
 
 TOKEN=""
@@ -675,7 +687,7 @@ else
         elapsed=$((elapsed + 5))
     done
     if [ -z "$TOKEN" ]; then
-        out "${HOSTS[0]}" "sudo tail -40 $INIT_LOG" | sed 's/^/  | /' >&2
+        out "${HOSTS[0]}" "sudo tail -40 $INIT_LOG" | redact_tokens | sed 's/^/  | /' >&2
         fail "${HOSTS[0]}: spx admin init produced no join token within ${TOKEN_TIMEOUT}s"
     fi
     log "  join token issued"
@@ -691,11 +703,9 @@ fi
 # waiting on a shell that node 2 is still holding.
 JOIN_PIDS=()
 for i in $(seq 1 $((N - 1))); do
-    on "${HOSTS[$i]}" "sudo spx admin join --force \
-        --node ${NODE_NAMES[$i]} \
-        --bind ${LAN_IPS[$i]} --cluster-bind ${LAN_IPS[$i]} --advertise ${WAN_IPS[$i]} \
-        --host ${LAN_IPS[0]}:$PORT --token $TOKEN \
-        --region $REGION --az $AZ" &
+    # The token goes to the node on stdin, so it is in no ssh argument and not in sudo's log.
+    # spx admin join takes it only as --token, so it is still in that process's arguments.
+    on "${HOSTS[$i]}" "$(remote_join_cmd "${NODE_NAMES[$i]}" "${LAN_IPS[$i]}" "${WAN_IPS[$i]}" "${LAN_IPS[0]}:$PORT" "$REGION" "$AZ")" <<<"$TOKEN" &
     JOIN_PIDS+=("$!")
 done
 
@@ -724,7 +734,7 @@ if ! $DRY_RUN; then
         elapsed=$((elapsed + 5))
     done
     if [ "$elapsed" -ge "$FORMATION_TIMEOUT" ]; then
-        out "${HOSTS[0]}" "sudo tail -40 $INIT_LOG" | sed 's/^/  | /' >&2
+        out "${HOSTS[0]}" "sudo tail -40 $INIT_LOG" | redact_tokens | sed 's/^/  | /' >&2
         fail "${HOSTS[0]}: formation did not complete within ${FORMATION_TIMEOUT}s"
     fi
     log "  formation complete"

scripts/install-node_test.sh

@@ -0,0 +1,41 @@
+#!/usr/bin/env bash
+# Tests remote_join_cmd from install-node.sh: the join token must reach spx only on stdin, so
+# sudo's log (which holds the command it was given) never contains it, and the other values must
+# arrive as plain arguments even when they hold shell syntax.
+# preflight runs this with sh
+[ -n "${BASH_VERSION:-}" ] || exec bash "$0" "$@"
+set -euo pipefail
+here="$(cd "$(dirname "$0")" && pwd)"
+work="$(mktemp -d)"
+trap 'rm -rf "$work"' EXIT
+mkdir "$work/bin"
+cd "$work"
+
+cat > "$work/bin/sudo" <<'SH'
+#!/bin/sh
+echo "$*" >> "$FAKE_SUDO_LOG"
+exec "$@"
+SH
+cat > "$work/bin/spx" <<'SH'
+#!/bin/sh
+for a in "$@"; do printf '<%s>' "$a"; done > "$FAKE_SPX_ARGV"
+SH
+chmod +x "$work/bin/sudo" "$work/bin/spx"
+export FAKE_SUDO_LOG="$work/sudo.log" FAKE_SPX_ARGV="$work/spx.argv" PATH="$work/bin:$PATH"
+
+eval "$(sed -n '/^remote_join_cmd()/,/^}/p' "$here/install-node.sh")"
+
+fails=0
+check() { if [ "$2" = "$3" ]; then echo "ok   $1"; else echo "FAIL $1: want '$3' got '$2'"; fails=$((fails + 1)); fi; }
+
+TOKEN=spx_join_FAKETOKEN-abc_1
+cmd="$(remote_join_cmd node2 10.0.0.2 203.0.113.2 '10.0.0.1:4432;touch PWNED' us-east-1 'az $(touch PWNED2)')"
+bash -c "$cmd" <<<"$TOKEN"
+
+check "the token is not in the command" "$(printf '%s' "$cmd" | grep -c FAKETOKEN || true)" 0
+check "the token is not in sudo's log" "$(grep -c FAKETOKEN "$FAKE_SUDO_LOG" || true)" 0
+check "spx got the token" "$(grep -o '<--token><[^>]*>' "$FAKE_SPX_ARGV")" "<--token><$TOKEN>"
+check "spx got the other values as given" "$(cat "$FAKE_SPX_ARGV")" "<admin><join><--force><--node><node2><--bind><10.0.0.2><--cluster-bind><10.0.0.2><--advertise><203.0.113.2><--host><10.0.0.1:4432;touch PWNED><--token><$TOKEN><--region><us-east-1><--az><az \$(touch PWNED2)>"
+check "no value was run as a command" "$(find "$work" -maxdepth 1 -name 'PWNED*' | wc -l | tr -d ' ')" 0
+
+[ "$fails" -eq 0 ]

scripts/lib/lambda-testlink-build_test.sh

@@ -0,0 +1,111 @@
+#!/bin/sh
+# lambda-testlink-build_test.sh — the test-link make targets refuse a malformed
+# image digest and an unsupported host or GOARCH before touching any output.
+#
+# Every case is refused on the recipe's first lines, so nothing is compiled.
+set -eu
+
+LIB_DIR=$(CDPATH='' cd -- "$(dirname -- "$0")" && pwd)
+REPO_ROOT=$(dirname -- "$(dirname -- "$LIB_DIR")")
+TMP=$(mktemp -d)
+trap 'rm -rf "$TMP"' EXIT
+
+fails=0
+pass() { echo "  ok   - $1"; }
+fail() { echo "  FAIL - $1"; fails=$((fails + 1)); }
+
+echo "-- make build-lambda-testlink / build-lambda-testlink-image guards"
+
+PWN="$TMP/pwn"
+OUT="$TMP/out"
+HEX=0123456789abcdef0123456789abcdef0123456789abcdef0123456789abcdef
+mkdir -p "$OUT"
+printf 'old daemon\n' > "$OUT/spx"
+printf 'old initializer\n' > "$OUT/hardkill.so"
+PAIR=$(cat "$OUT/spx" "$OUT/hardkill.so" | cksum)
+
+# unchanged <name> — the existing pair is intact, no scratch was made and no
+# injected command ran.
+unchanged() {
+    if [ -e "$PWN" ]; then
+        fail "$1: injected command ran"
+        rm -f "$PWN"
+    elif [ -e "$OUT.tmp" ]; then
+        fail "$1: scratch directory created"
+    elif [ "$(cat "$OUT/spx" "$OUT/hardkill.so" | cksum)" != "$PAIR" ]; then
+        fail "$1: existing pair replaced"
+    else
+        pass "$1"
+    fi
+}
+
+# refused <name> <want> <log> <status> — make failed with the expected message.
+refused() {
+    if [ "$4" -eq 0 ]; then
+        fail "$1: make succeeded"
+    elif ! grep -q "$2" "$3"; then
+        fail "$1: wrong refusal: $(tr '\n' ' ' < "$3")"
+    else
+        unchanged "$1"
+    fi
+}
+
+testlink() {
+    make -s -C "$REPO_ROOT" --no-print-directory build-lambda-testlink LAMBDA_TESTLINK_DIR="$OUT" "$@" > "$TMP/log" 2>&1
+}
+
+DIGEST_MSG="LAMBDA_IMAGE_DIGEST must be 64 lowercase hex digits"
+NL='
+'
+for v in '' "\$(shell touch $PWN)" "\$\$(touch $PWN)" "\`touch $PWN\`" "$HEX; touch $PWN" \
+    "$HEX\" -X main.x=\"y" "$HEX -X main.x=y" "$HEX${NL}touch $PWN" \
+    0123456789ABCDEF0123456789ABCDEF0123456789ABCDEF0123456789ABCDEF "${HEX%?}" "${HEX}0"; do
+    # GNU make itself expands a command-line $(shell ...) while parsing, for
+    # any target, so that one literal is tested only from the environment.
+    case "$v" in
+    "\$(shell"*) ;;
+    *)
+        status=0
+        testlink LAMBDA_IMAGE_DIGEST="$v" || status=$?
+        refused "command-line digest [$v]" "$DIGEST_MSG" "$TMP/log" "$status"
+        ;;
+    esac
+    status=0
+    LAMBDA_IMAGE_DIGEST="$v" testlink || status=$?
+    refused "environment digest [$v]" "$DIGEST_MSG" "$TMP/log" "$status"
+done
+
+status=0
+env -u LAMBDA_IMAGE_DIGEST make -s -C "$REPO_ROOT" --no-print-directory build-lambda-testlink LAMBDA_TESTLINK_DIR="$OUT" > "$TMP/log" 2>&1 || status=$?
+refused "digest unset" "$DIGEST_MSG" "$TMP/log" "$status"
+
+ARCH_MSG="supports only an x86_64 host building linux/amd64"
+SHIM="$TMP/shim"
+mkdir -p "$SHIM"
+printf '#!/bin/sh\necho aarch64\n' > "$SHIM/uname"
+chmod +x "$SHIM/uname"
+
+status=0
+GOARCH=arm64 testlink LAMBDA_IMAGE_DIGEST="$HEX" || status=$?
+refused "GOARCH=arm64" "$ARCH_MSG" "$TMP/log" "$status"
+status=0
+PATH="$SHIM:$PATH" testlink LAMBDA_IMAGE_DIGEST="$HEX" || status=$?
+refused "aarch64 host" "$ARCH_MSG" "$TMP/log" "$status"
+
+# The image target refuses before building qboot or any agent binary.
+for c in "GOARCH=arm64" "PATH=$SHIM:$PATH"; do
+    status=0
+    env "$c" make -s -C "$REPO_ROOT" --no-print-directory build-lambda-testlink-image > "$TMP/log" 2>&1 || status=$?
+    if [ "$status" -eq 0 ]; then
+        fail "image target [${c%%=*}]: make succeeded"
+    elif ! grep -q "$ARCH_MSG" "$TMP/log" || grep -q "qboot\|go build" "$TMP/log"; then
+        fail "image target [${c%%=*}]: did not refuse first: $(tr '\n' ' ' < "$TMP/log")"
+    else
+        pass "image target [${c%%=*}] refused first"
+    fi
+done
+
+if [ "$fails" -ne 0 ]; then
+    echo "$fails failure(s)"
+    exit 1
+fi

scripts/terraform/latitude-spx/.gitignore

@@ -0,0 +1,10 @@
+.terraform/
+terraform.tfstate
+terraform.tfstate.*
+*.tfplan
+*.auto.tfvars
+crash.log
+terraform.tfvars
+*.pem
+*.log
+.terraform.lock.hcl

scripts/terraform/latitude-spx/README.md

@@ -0,0 +1,137 @@
+# latitude-spx
+
+Builds a Spinifex deployment on latitude.sh bare metal from your laptop with `deploy-latitude.sh`.
+The commands, flags, step markers and state layout follow `vultr-spx`.
+
+**Status: not yet run end to end.**
+The script was written from the plan doc findings (`docs/development/feature/latitude-bm.md`, L1 to L26) and from the provider schema.
+No latitude.sh API call or host run was made while building it, so treat the first `up` as the test.
+One node has been the only shape run live, and only in part (see Unverified). The three-node flow, including the floating pool, has not been run: Ben takes that over.
+
+## What it creates
+
+| Resource | Notes |
+|---|---|
+| Project | `spinifex-<name>`, unless `--project` names an existing one (then it is never deleted) |
+| SSH key | `spinifex-<name>`, account scoped |
+| User data | `spinifex-<name>`, host preparation only (no `{{ }}`, latitude.sh substitutes them) |
+| Servers | `m4-metal-medium`, site `DAL`, Ubuntu 26.04, hourly billing, hostnames `spx-<name>-nodeN` |
+| Elastic IP | one by default (`elastic_ip_count` in `variables.tf`, 1 to 5; set `TF_VAR_elastic_ip_count` to change it). With one node it is a static pool address and the count must stay 1, because a second pool is never tried when the first is full; with three nodes the addresses form the `latitude` floating pool, which Spinifex moves between servers. `nginx-webserver` needs one address and runs on one node; `nginx-alb` needs two public addresses and is a three-node workbook |
+| Virtual network | three nodes only: `spinifex-<name>`, one VLAN assignment per server |
+
+## Requirements
+
+OpenTofu (or Terraform) 1.5 or later, `aws` CLI, `python3`, `curl`, `ssh`, an SSH key pair.
+
+| Variable | When | Meaning |
+|---|---|---|
+| `LATITUDESH_AUTH_TOKEN` | always | the account API token; used by the provider and this script only, never sent to a node |
+| `LATITUDESH_NODE_TOKEN` | `--nodes 3` | a second, different, writable API key, copied to each node (`/etc/spinifex/latitude/api-token`, owned by `spinifex-daemon` with mode 0400, in a root-owned 0711 directory; the other services share the `spinifex` group, so no group may read it; ownership and mode are set on every `up`; a symlink or any other owner there is refused) so the daemon can move Elastic IPs |
+
+The script refuses a node key it can see is read-only (it lists `/auth/api_keys` and matches the key's last five characters). If the key is not among the account's keys, write authority is first proven by the first Elastic IP move.
+
+Exposure: a latitude.sh API key is team-wide. It is not scoped to Elastic IPs, so anyone who reads it from a node can do whatever the key's owner can in the team (the API exposes `read_only` and `allowed_ips` on a key, and no narrower scope). Create the node key with `allowed_ips` set to the nodes' public addresses, and revoke it after `destroy`.
+
+Nodes are reached as `ubuntu` with the deployment's SSH key, and remote commands run through passwordless `sudo`: latitude.sh puts the key on the image's login user and root login is not enabled.
+
+## Usage
+
+```
+export LATITUDESH_AUTH_TOKEN=...        # from the latitude.sh dashboard
+./deploy-latitude.sh up --name demo     # one node
+
+export LATITUDESH_NODE_TOKEN=...        # a second, writable key
+./deploy-latitude.sh up --name demo3 --nodes 3
+./deploy-latitude.sh status --name demo
+./deploy-latitude.sh ssh --name demo node1 'systemctl status spinifex.target'
+./deploy-latitude.sh credentials --name demo
+./deploy-latitude.sh stop node1 --name demo
+./deploy-latitude.sh start node1 --name demo
+./deploy-latitude.sh destroy --name demo
+```
+
+Options: `--name`, `--site` (alias `--region`), `--plan`, `--os`, `--project ID|SLUG`, `--distro TARBALL`, `--channel`, `--spx-region`, `--spx-az`, `--ssh-key`, `--image`, `--data-disk-id`, `--vpc-subnet` (the VLAN's /24, three nodes), `--nodes 1|3`, `--yes`.
+`./deploy-latitude.sh --help` lists them with defaults.
+
+Three nodes need `--distro` with a tarball built from this branch (`make distro-amd64`) until a release carries the latitude provider.
+A released spx rejects `source="latitude"` and its daemon restarts in a loop, so `up --nodes 3` without `--distro` stops in preflight.
+Before the pool is written, each node's `/usr/local/bin/spx` is checked for the `latitude_elastic_ip_ids` key name; spx has no validate subcommand, so this is a string search of the binary, not a run of its config check.
+
+## What `up` does
+
+`terraform` then, per node: `host` (wait for SSH as `ubuntu` and cloud-init, `/dev/kvm`), `disk`, `vlan` (three nodes), `install`, `form`, `start`; then `pool` (one node) or `cluster` (three nodes: `install-node.sh` forms the cluster); three nodes then install the node token on every `up` and add the `latitude` pool (`pool`); then `restart` (three nodes wait for the current daemon to log "floating pool ready" for the pool), `public`, `image` and the AWS profile.
+Each step writes `step.<name>[.nodeN]` in the state directory, and a rerun resumes at the first step without one.
+
+- **Data disk.** `host-data-disk.sh` picks the blank NVMe by content and by-id EUI, never by `nvmeXn1` (L1, L11).
+A candidate is at least 100 GB with no signature, partitions, holders or mounts, and not in the root filesystem's ancestry; with none it stops and lists the disks.
+With several (LON2 has two 1.7 TB and one spare 447 GB blank NVMe next to the root disk) it takes the largest, the lowest by-id path among equals, prints what it chose and the blank disks it left, and never touches the others (`--data-disk-id` chooses another).
+It mounts by UUID with `nofail` and the script records the UUID in `data_uuid.<node>`.
+- **Mount check.** Before anything writes under `/var/lib/spinifex`, and on every run, the mounted filesystem UUID must equal the recorded one.
+The units that write there also get a `RequiresMountsFor=/var/lib/spinifex` drop-in.
+The mount is by UUID in fstab, so renamed disks do not change it, and there is no check at service start.
+- **Firewall.** `setup.sh --firewall on` runs before Spinifex first starts.
+- **One node.** The Elastic IP is a static pool (`gateway_ip = "192.0.2.1"`, the node's next hop as `gateway`, no `prefix_len`) with the IMDS addresses moved off the ones latitude.sh routes.
+- **Three nodes.** The `vlan` step makes `eno2.<vid>` with `ip link` and persists it as systemd-networkd files for that link plus a drop-in on the `eno2` unit.
+It never runs `netplan apply`, refuses a subnet that overlaps the SSH peer or a route, and removes the VLAN again if the public interface's address, default route or the route to the SSH peer differ afterwards.
+`form_cluster` then puts the cluster and Geneve planes on that VLAN with `install-node.sh` (`--user ubuntu`); its output is not kept, as it can hold the join token.
+- **Floating pool (three nodes).** The `latitude` pool names the Elastic IP IDs and the project ID from the Terraform outputs, and each node's `provider_host_id` is its server ID. `spinifex.toml` is edited idempotently and read back. Failover moves an address with one `PATCH /elastic_ips/{id}` and waits for it to be active; latitude.sh has no unassign, so a node that loses an address is not told to release it.
+- **State** is in `~/.spinifex/latitude/<name>/`: Terraform state, `ids` (every `latitudesh_*` object in the state, merged on each apply and before each destroy), `step.*` markers, logs and the data disk UUIDs.
+- **AWS profile** `spinifex-latitude-<name>` is replaced on each `up`, with the CA at `~/.spinifex/latitude-<name>-ca.pem`.
+The secret key is written to `~/.aws/credentials` (0600, other profiles kept, the old file saved beside it) and never appears in a process argument.
+
+## Token handling
+
+`LATITUDESH_AUTH_TOKEN` is read from the environment by the provider and by the script's API calls.
+It is not a Terraform variable, is on no command line (curl reads it from stdin), is not sent to a node, and is written to no file.
+`LATITUDESH_NODE_TOKEN` is checked with curl reading it from stdin and reaches each node on stdin of the install script, so it is in no argument on either side.
+`TF_LOG` is unset by the script.
+The join token that `install-node.sh` creates is shared with Vultr and OCI: its init log is root-only, and `spx admin join` still takes it as `--token` on the joining node (tracked separately).
+
+## Recovery
+
+If SSH is lost once the firewall is armed, use the serial console.
+`up` enables the server's out-of-band connection with the deployment's SSH key and prints the `ssh` command it returns before it arms the firewall, falling back to the Out-of-Band tab in the dashboard.
+Logging in there needs an OS user password, which you set yourself while SSH still works.
+A fresh deployment that cannot be recovered is cheaper to `destroy` and `up` again.
+The VLAN step prints the console command first.
+It refuses a subnet that holds the SSH peer or overlaps an existing route, and removes the VLAN again if the route to the SSH peer or the public interface changes.
+
+## destroy
+
+`destroy` asks you to type the deployment name (`--yes` skips only that), then records every ID in the Terraform state and stops without destroying anything if it cannot.
+It runs `terraform destroy`, fetches every recorded ID by itself (2xx present, 404 gone, anything else fails and keeps the records), and deletes what remains of those.
+Servers, keys, user data, virtual networks and projects whose names match the deployment but that are not recorded are listed by ID and deleted only after you type `delete-untracked` for exactly those IDs, which `--yes` never supplies; an ID that shows up later needs its own confirmation, and without a terminal they are kept and `destroy` exits 1.
+Elastic IPs are touched only by recorded ID.
+Elastic IP release is asynchronous and only works once the address is active or in error, so the sweep retries.
+Because the Elastic IP list ignores pagination, `destroy` ends by asking you to check the dashboard for leftovers and type `checked`.
+An `up` that failed while creating can leave an object nothing recorded, such as an Elastic IP; that dashboard check is the only thing that finds it.
+
+## Unverified
+
+Nothing in the three-node flow has run against real servers. For one node, only the parts the person ran have been exercised, and the script has changed since.
+
+- The OS slug `ubuntu_26_04_x64_lts` (the `latitudesh_operating_system` data source fails the plan if it is wrong).
+- That the deploy key is on `ubuntu` and `sudo -n` works for it, and that `sudo` resets nothing the script needs (the SSH peer address is read before `sudo` for that reason).
+- The power action request (`POST /servers/{id}/actions` with `power_off` and `power_on`).
+- Resolving `--project` as a slug through `GET /projects/{slug}`, and the `in_stock` shape of `GET /plans` used by the stock check, which only warns when it cannot read it.
+- The `GET` and `DELETE` paths for `/user_data/{id}` and `/virtual_networks/{id}`.
+- That the Spinifex API certificate names the node's public address, which the profile check needs.
+- That the private NIC is `eno2`, its tagging and MTU (the script uses 1500), and that the networkd files persist the VLAN across a reboot.
+- `GET /auth/api_keys` and its `token_last_slice` and `read_only` fields, which the node key check relies on; and that a key with `allowed_ips` set to the nodes' addresses still works from them.
+- The Elastic IP move itself: `PATCH /elastic_ips/{id}` with `server_id`, its 202 then `active`, the error codes, and how long a move takes; the client waits up to its move timeout and makes one write per call.
+- That the daemon reads the node token and logs "floating pool ready", then fails over when a node stops (`stop node1`) and the address lands on another node.
+- `install-node.sh` as `--user ubuntu` on latitude.sh servers.
+
+## Not exercised
+
+Any real API call from the three-node flow, `destroy`, `host-data-disk.sh` on a host and the VLAN step on a server.
+Guest launch, public reachability of the Elastic IP, the Terraform workbooks and failover are not part of `up`; prove them by hand.
+`e2e-latitude.sh` and removing the old `install-node.sh` in this directory are not done.
+
+## Tests
+
+`sh node-scripts_test.sh` checks the node token install (install, unchanged, rotation, modes, an older install converging, symlinks and owners refused; run it as root too, with `sudo sh node-scripts_test.sh`, to check that only the daemon's user can read the file) and the readiness check (old success then current failure) with stand-in tools.
+`bash deploy-argv_test.sh` checks that neither API token, nor the five characters used to find the node key, is an argument of any external command in the node token paths, and that the SSH peer address comes from the same connection after sudo's environment reset.
+`scripts/install-node_test.sh` (part of `make preflight`) checks that the join token reaches `spx admin join` only on stdin, so sudo's log never holds it.
+The first two are not part of `make preflight`.
+The Go side: `cd spinifex && GOWORK=off go test ./providers/cloud/latitude/ ./domains/network/external/floating/ ./bootstrap/config/ ./runtime/daemon/`.

scripts/terraform/latitude-spx/cloud-init/user-data.yaml

@@ -0,0 +1,23 @@
+#cloud-config
+# Host preparation only. Spinifex is installed and formed over SSH by deploy-latitude.sh.
+# latitude.sh substitutes double-brace variables in user data, so this file holds none.
+write_files:
+  - path: /etc/cloud/cloud.cfg.d/99-spinifex-network.cfg
+    content: |
+      network: {config: disabled}
+  # The image's netplan lists eno2 with no address and without optional, so boot waits two minutes
+  # for it. --any returns once one link, eno1, is online; it changes the wait, not the network.
+  # UNVERIFIED: the binary path is the Ubuntu one and the drop-in first applies on the next boot.
+  - path: /etc/systemd/system/systemd-networkd-wait-online.service.d/10-spinifex-any.conf
+    content: |
+      [Service]
+      ExecStart=
+      ExecStart=/usr/lib/systemd/systemd-networkd-wait-online --any
+# runcmd carries on past a failing command, so each line stands alone.
+runcmd:
+  - [sh, -c, "command -v ufw >/dev/null && ufw --force disable || true"]
+  - [sh, -c, "systemctl disable --now unattended-upgrades.service apt-daily-upgrade.timer apt-daily.timer || true"]
+  # The image ships empty apt lists and no git, which the workbooks need.
+  - [sh, -c, "apt-get update -qq && DEBIAN_FRONTEND=noninteractive apt-get install -y -qq git || true"]
+  - [sh, -c, "systemctl daemon-reload || true"]
+  - [sh, -c, "touch /var/lib/cloud/spinifex-host-ready"]

scripts/terraform/latitude-spx/compute.tf

@@ -0,0 +1,70 @@
+locals {
+  hostname_prefix = "spx-${var.deployment_name}-node"
+  project         = var.project != "" ? var.project : latitudesh_project.deploy[0].slug
+  project_id      = var.project != "" ? var.project_id : latitudesh_project.deploy[0].id
+}
+
+# Both lookups fail the plan when a slug does not exist, which is the only check of the OS slug.
+data "latitudesh_operating_system" "os" {
+  slug = var.operating_system
+}
+
+data "latitudesh_plan" "plan" {
+  slug = var.plan
+}
+
+resource "latitudesh_project" "deploy" {
+  count             = var.project == "" ? 1 : 0
+  name              = "spinifex-${var.deployment_name}"
+  description       = "Spinifex deployment ${var.deployment_name}"
+  environment       = "Development"
+  provisioning_type = "on_demand"
+}
+
+# The key is team scoped. The server login user is ubuntu, not root.
+resource "latitudesh_ssh_key" "deploy" {
+  name       = "spinifex-${var.deployment_name}"
+  public_key = trimspace(file(pathexpand(var.ssh_public_key_path)))
+}
+
+# Host preparation only. Latitude substitutes double-brace variables inside user data, so the file
+# must hold none. Spinifex is installed over SSH by deploy-latitude.sh.
+resource "latitudesh_user_data" "host" {
+  description = "spinifex-${var.deployment_name}"
+  content     = filebase64("${path.module}/cloud-init/user-data.yaml")
+}
+
+# billing must be hourly: the default is monthly, which bills upfront with no refund.
+resource "latitudesh_server" "node" {
+  count            = var.node_count
+  project          = local.project
+  hostname         = "${local.hostname_prefix}${count.index + 1}"
+  site             = var.site
+  plan             = data.latitudesh_plan.plan.slug
+  operating_system = data.latitudesh_operating_system.os.slug
+  billing          = "hourly"
+  ssh_keys         = [latitudesh_ssh_key.deploy.id]
+  user_data        = latitudesh_user_data.host.id
+
+  timeouts = {
+    create = "60m"
+  }
+}
+
+# Elastic IPs are bought here and never created or deleted by Spinifex. Each starts on a server, spread
+# over the nodes, and server_id is ignored afterwards so a move by Spinifex is not undone by a later apply.
+resource "latitudesh_elastic_ip" "pool" {
+  count = var.elastic_ip_count
+  # The provider documents ID or slug but posts the value as project_id unresolved: a slug gave 404.
+  project   = local.project_id
+  server_id = latitudesh_server.node[count.index % var.node_count].id
+
+  lifecycle {
+    ignore_changes = [server_id]
+  }
+}
+
+moved {
+  from = latitudesh_elastic_ip.node1
+  to   = latitudesh_elastic_ip.pool[0]
+}

scripts/terraform/latitude-spx/deploy-argv_test.sh

@@ -0,0 +1,228 @@
+#!/usr/bin/env bash
+# Checks that neither API token is ever an argument of an external command in the node token
+# paths of deploy-latitude.sh. External commands are stand-ins that record their arguments and
+# their standard input; the test fails if a token is in the first and proves it is in the second.
+# preflight runs this with sh
+[ -n "${BASH_VERSION:-}" ] || exec bash "$0" "$@"
+set -euo pipefail
+exec </dev/null
+here="$(cd "$(dirname "$0")" && pwd)"
+work="$(mktemp -d)"
+trap 'rm -rf "$work"' EXIT
+mkdir "$work/bin" "$work/home"
+real_python="$(command -v python3)"
+
+AUTH=FAKEAUTHTOKEN9
+NODE=FAKENODETOKEN7
+
+# Every command the script runs outside bash records "name args" and, for the first line of
+# standard input, "name stdin: ...".
+for c in curl ssh scp aws tofu ssh-keygen; do
+    cat > "$work/bin/$c" <<'SH'
+#!/bin/sh
+echo "$(basename "$0") $*" >> "$FAKE_ARGV"
+if [ ! -t 0 ]; then
+    in="$(cat)"
+    [ -z "$in" ] || echo "$(basename "$0") stdin: $in" >> "$FAKE_STDIN"
+fi
+case "$(basename "$0")" in
+    curl) printf '{"data":[{"attributes":{"token_last_slice":"%s","read_only":false}}]}\n200' "$(printf '%s' "$FAKE_NODE" | tail -c 5)" ;;
+    ssh)
+        if [ -n "${FAKE_SSH_EVAL:-}" ]; then
+            # Behave like a login on a node: skip the options and host, run the rest with the
+            # connection variable of this connection, and let sudo reset the environment.
+            while [ "$#" -gt 0 ]; do case "$1" in ubuntu@*) shift; break ;; *) shift ;; esac; done
+            SSH_CONNECTION="$FAKE_SSH_CONNECTION" bash -c "$*"
+        else
+            echo installed
+        fi ;;
+esac
+SH
+    chmod +x "$work/bin/$c"
+done
+cat > "$work/bin/sudo" <<'SH'
+#!/bin/sh
+while [ "$#" -gt 0 ]; do case "$1" in -*) shift ;; *) break ;; esac; done
+exec env -i PATH="$PATH" "$@"
+SH
+chmod +x "$work/bin/sudo"
+cat > "$work/bin/python3" <<SH
+#!/bin/sh
+echo "python3 \$*" >> "\$FAKE_ARGV"
+exec "$real_python" "\$@"
+SH
+chmod +x "$work/bin/python3"
+
+export FAKE_ARGV="$work/argv" FAKE_STDIN="$work/stdin" FAKE_NODE="$NODE" PATH="$work/bin:$PATH" HOME="$work/home"
+: > "$FAKE_ARGV"
+: > "$FAKE_STDIN"
+export LATITUDESH_AUTH_TOKEN="$AUTH" LATITUDESH_NODE_TOKEN="$NODE"
+
+# shellcheck source=deploy-latitude.sh
+source "$here/deploy-latitude.sh"
+
+# shellcheck disable=SC2034 # read by the sourced functions
+NODES=3 IP=192.0.2.10 NODE_NAME=node1
+check_node_token >/dev/null
+install_key >/dev/null 2>&1
+
+fails=0
+for tok in "$AUTH" "$NODE" "${NODE: -5}"; do
+    if grep -q "$tok" "$FAKE_ARGV"; then
+        echo "FAIL token $tok is in a command's arguments:"
+        grep "$tok" "$FAKE_ARGV"
+        fails=$((fails + 1))
+    else
+        echo "ok   $tok is in no argument"
+    fi
+done
+for what in "curl stdin:.*Bearer $NODE" "ssh stdin: $NODE"; do
+    if grep -q "$what" "$FAKE_STDIN"; then
+        echo "ok   the check can see the token where it is sent ($what)"
+    else
+        echo "FAIL the token never reached stdin ($what): the test would pass without checking anything"
+        fails=$((fails + 1))
+    fi
+done
+
+# The peer address comes from the same connection that runs the command, whichever source
+# address that connection has, and survives sudo's environment reset.
+export FAKE_SSH_EVAL=1
+for src in 198.51.100.7 203.0.113.99; do
+    export FAKE_SSH_CONNECTION="$src 51000 192.0.2.10 22"
+    got="$(on 'printf %s "$SSH_PEER"')"
+    if [ "$got" = "$src" ]; then echo "ok   peer $src reaches the command"; else echo "FAIL peer for $src was '$got'"; fails=$((fails + 1)); fi
+done
+
+# A three-node run needs a build that knows the latitude source: a release would be refused by its
+# own config check. The guard needs --distro, and the node is checked before it gets the pool.
+check_result() { if [ "$2" = "$3" ]; then echo "ok   $1"; else echo "FAIL $1: want '$3' got '$2'"; fails=$((fails + 1)); fi; }
+check_result "three nodes without --distro stop" "$( (NODES=3 DISTRO="" require_branch_build) >/dev/null 2>&1 && echo ran || echo stopped)" stopped
+check_result "three nodes with --distro go on" "$( (NODES=3 DISTRO=/tmp/x.tar.gz require_branch_build) >/dev/null 2>&1 && echo ran || echo stopped)" ran
+check_result "one node without --distro goes on" "$( (NODES=1 DISTRO="" require_branch_build) >/dev/null 2>&1 && echo ran || echo stopped)" ran
+check_result "one node with one address goes on" "$( (NODES=1 TF_VAR_elastic_ip_count=1 require_one_address) >/dev/null 2>&1 && echo ran || echo stopped)" ran
+check_result "one node with no setting goes on" "$( (NODES=1 TF_VAR_elastic_ip_count="" require_one_address) >/dev/null 2>&1 && echo ran || echo stopped)" ran
+check_result "one node with two addresses stops" "$( (NODES=1 TF_VAR_elastic_ip_count=2 require_one_address) >/dev/null 2>&1 && echo ran || echo stopped)" stopped
+check_result "three nodes with three addresses go on" "$( (NODES=3 TF_VAR_elastic_ip_count=3 require_one_address) >/dev/null 2>&1 && echo ran || echo stopped)" ran
+printf 'binary with latitude_elastic_ip_ids inside\n' > "$work/spx-new"
+printf 'binary of a release\n' > "$work/spx-old"
+SPX_BIN="$work/spx-new"
+check_result "a build that knows latitude is accepted" "$( (node_knows_latitude) >/dev/null 2>&1 && echo ran || echo stopped)" ran
+# shellcheck disable=SC2034 # read by the sourced functions
+SPX_BIN="$work/spx-old"
+check_result "a release is refused" "$( (node_knows_latitude) >/dev/null 2>&1 && echo ran || echo stopped)" stopped
+# The profile spx writes under sudo is in the ubuntu user's home, so node-side reads name that
+# directory instead of relying on root's HOME.
+mkdir "$work/awsdir"
+printf '[other]\naws_access_key_id = X\n[spinifex]\naws_access_key_id = AK1\naws_secret_access_key = SK1\n[z]\nq = 1\n' > "$work/awsdir/credentials"
+node_ip() { printf 192.0.2.10; }
+# shellcheck disable=SC2034 # read by the sourced functions
+NODE_AWS_DIR="$work/awsdir"
+check_result "credentials come from the node's AWS dir" "$(cmd_credentials | tr '\n' ' ')" "[spinifex] aws_access_key_id = AK1 aws_secret_access_key = SK1 "
+
+# A failed API call logs its status and a documented code, never the response text, which can
+# echo the request's headers.
+mkdir "$work/bin2"
+cat > "$work/bin2/curl" <<'SH'
+#!/bin/sh
+cat > /dev/null
+printf '{"errors":[{"code":"%s","detail":"Authorization: Bearer %s"}]}\n401' "$FAKE_CODE" "$FAKE_LEAK"
+SH
+chmod +x "$work/bin2/curl"
+logged="$(FAKE_CODE=UNAUTHORIZED FAKE_LEAK="$AUTH" PATH="$work/bin2:$PATH" latitude GET /servers 2>&1 || true)"
+check_result "a failed call names its status and code" "$(grep -c 'returned 401 (UNAUTHORIZED)' <<<"$logged" || true)" 1
+check_result "a failed call never logs the token" "$(grep -c "$AUTH" <<<"$logged" || true)" 0
+logged="$(FAKE_CODE="$AUTH" FAKE_LEAK="$AUTH" PATH="$work/bin2:$PATH" latitude GET /servers 2>&1 || true)"
+check_result "an undocumented code is not logged" "$(grep -c "$AUTH" <<<"$logged" || true)" 0
+
+# A recorded ID that now names another deployment's object is not ours to delete.
+# shellcheck disable=SC2034 # read by the sourced functions
+NAME=demo
+latitude() { printf '%s' "$FAKE_LAT"; }
+orig_expected_project_id="$(declare -f expected_project_id)"
+expected_project_id() { echo proj_1; }
+srv() {
+    local frag=',"project":{"id":"proj_1"}'
+    [ "$#" -lt 2 ] || frag="$2"
+    printf '{"data":{"id":"sv_1","attributes":{"hostname":"%s"%s}}}' "$1" "$frag"
+}
+FAKE_LAT="$(srv spx-demo-node2)"
+check_result "a recorded server with our hostname is ours" "$(recorded_identity server sv_1 && echo ours || echo "rc=$?")" ours
+FAKE_LAT="$(srv spx-other-node1)"
+check_result "a recorded server with another hostname is not" "$(recorded_identity server sv_1 && echo ours || echo "rc=$?")" rc=1
+FAKE_LAT="$(srv spx-demo2-node1)"
+check_result "a hostname that only starts like ours is not" "$(recorded_identity server sv_1 && echo ours || echo "rc=$?")" rc=1
+FAKE_LAT="$(srv spx-demo-node1 ',"project":{"id":"proj_other"}')"
+check_result "our hostname in another project is not ours" "$(recorded_identity server sv_1 && echo ours || echo "rc=$?")" rc=1
+FAKE_LAT="$(srv spx-demo-node1 '')"
+check_result "a server with no project is unknown" "$(recorded_identity server sv_1 && echo ours || echo "rc=$?")" rc=2
+FAKE_LAT="$(srv spx-demo-node1 ',"project":{}')"
+check_result "a project with no id is unknown" "$(recorded_identity server sv_1 && echo ours || echo "rc=$?")" rc=2
+FAKE_LAT="$(srv spx-demo-node1 ',"project":null')"
+check_result "a null project is unknown" "$(recorded_identity server sv_1 && echo ours || echo "rc=$?")" rc=2
+expected_project_id() { return 1; }
+FAKE_LAT="$(srv spx-demo-node1)"
+check_result "no known project id is unknown" "$(recorded_identity server sv_1 && echo ours || echo "rc=$?")" rc=2
+expected_project_id() { echo proj_1; }
+FAKE_LAT='not json'
+check_result "an unreadable object is unknown" "$(recorded_identity server sv_1 && echo ours || echo "rc=$?")" rc=2
+KNOWN_IDS="$work/ids"
+printf 'server sv_1\n' > "$KNOWN_IDS"
+lat_exists() { return 0; }
+FAKE_LAT="$(srv spx-other-node1)"
+check_result "destroy stops before Terraform when a recorded ID was reused" "$( (check_recorded_identity) >/dev/null 2>&1 && echo went || echo stopped)" stopped
+FAKE_LAT="$(srv spx-demo-node1 ',"project":{"id":"proj_other"}')"
+check_result "destroy stops when the same hostname is in another project" "$( (check_recorded_identity) >/dev/null 2>&1 && echo went || echo stopped)" stopped
+FAKE_LAT="$(srv spx-demo-node1 '')"
+check_result "destroy stops when the project cannot be read" "$( (check_recorded_identity) >/dev/null 2>&1 && echo went || echo stopped)" stopped
+FAKE_LAT="$(srv spx-demo-node1)"
+check_result "destroy goes on when the recorded server is ours" "$( (check_recorded_identity) >/dev/null 2>&1 && echo went || echo stopped)" went
+leftovers_out="$( (lat_match() { return 0; }; FAKE_LAT="$(srv spx-other-node1)" leftovers) 2>/dev/null || true)"
+check_result "a reused ID is not listed as a leftover to delete" "$(grep -c sv_1 <<<"$leftovers_out" || true)" 0
+
+# After terraform destroy the state's project output is gone; the remembered ID still lets the
+# sweep delete a matching leftover and keep a foreign one.
+eval "$orig_expected_project_id"
+TF_BIN="$work/tofu"
+printf '#!/bin/sh\nif [ -f "%s/gone" ]; then exit 1; fi\nprintf proj_1\n' "$work" > "$TF_BIN"
+chmod +x "$TF_BIN"
+EXPECTED_PROJECT=""
+remember_project_id
+check_result "the project ID is remembered while the state holds it" "$EXPECTED_PROJECT" proj_1
+: > "$work/gone"
+check_result "the state no longer answers after destroy" "$("$TF_BIN" output >/dev/null 2>&1 && echo up || echo gone)" gone
+printf 'server sv_1\nserver sv_2\n' > "$KNOWN_IDS"
+lat_match() { return 1; }
+latitude() { if [ "$2" = /servers/sv_1 ]; then srv spx-demo-node1; else srv spx-other-node1 ',"project":{"id":"proj_other"}'; fi; }
+swept="$(leftovers 2>/dev/null || true)"
+check_result "after destroy a matching leftover is still listed" "$(grep -c 'server sv_1 ' <<<"$swept" || true)" 1
+check_result "after destroy a foreign one is kept out" "$(grep -c 'server sv_2 ' <<<"$swept" || true)" 0
+EXPECTED_PROJECT=""
+swept="$(leftovers 2>/dev/null || true)"
+check_result "without the remembered ID a leftover is unknown, not matched" "$(grep -c 'server sv_1 sv_1 (unknown)' <<<"$swept" || true)" 1
+expected_project_id() { echo proj_1; }
+
+# Cleanup never deletes an object whose ownership could not be read, recorded or not.
+deleted="$work/deleted"
+: > "$deleted"
+latitude() { printf '%s %s\n' "$1" "$2" >> "$deleted"; }
+printf 'server sv_1\nserver sv_2\n' > "$KNOWN_IDS"
+delete_left $'server sv_1 sv_1 (unknown)\nserver sv_2 spx-demo-node2\n' "" >/dev/null 2>&1
+check_result "an unreadable recorded object is not deleted" "$(grep -c sv_1 "$deleted" || true)" 0
+check_result "a verified recorded object is deleted" "$(grep -c 'DELETE .*sv_2' "$deleted" || true)" 1
+
+# Both deployment sizes write the metadata remap.
+cfg="$work/spinifex.toml"
+extract() { sed -n "/^$1()/,/^}/p" "$here/deploy-latitude.sh" | sed -n "/<<'PY'/,/^PY\$/p" | sed '1d;$d' | sed "s|/etc/spinifex/spinifex.toml|$cfg|"; }
+printf '[network]\nexternal_mode = "nat"\n\n[bootstrap]\ny = 1\n' > "$cfg"
+extract configure_pool | python3 -I - 203.0.113.1 203.0.113.254
+check_result "one node gets the metadata remap" "$(grep -c 'imds_host_meta_ip = "169.254.42.254"' "$cfg")" 1
+printf '[network]\nexternal_mode = "nat"\n\n[nodes.node1]\nx = 1\n\n[nodes.node2]\nx = 1\n\n[nodes.node3]\nx = 1\n\n[bootstrap]\ny = 1\n' > "$cfg"
+extract configure_latitude_pool | python3 -I - proj_1 eip_1,eip_2 node1=sv_1,node2=sv_2,node3=sv_3
+check_result "three nodes get the metadata remap" "$(grep -c 'imds_host_dns_ip  = "169.254.42.253"' "$cfg")" 1
+check_result "three nodes get the latitude pool" "$(grep -c 'source                  = "latitude"' "$cfg")" 1
+extract configure_latitude_pool | python3 -I - proj_1 eip_1,eip_2 node1=sv_1,node2=sv_2,node3=sv_3
+check_result "writing the config again changes nothing" "$(grep -c 'imds_host_meta_ip' "$cfg")" 1
+unset FAKE_SSH_EVAL
+
+[ "$fails" -eq 0 ]