diff --git a/Makefile b/Makefile
index 442a7dd..885f962 100644
--- a/Makefile
+++ b/Makefile
@@ -1,4 +1,4 @@
-.PHONY: help build run test lint verify verify-docs verify-deploy loadtest clean docker-up docker-down migrate-up migrate-down generate test-cover frontend-build qa-stats
+.PHONY: help build run test lint verify verify-docs verify-deploy loadtest acme-cert-manager-test acme-rfc-conformance-test clean docker-up docker-down migrate-up migrate-down generate test-cover frontend-build qa-stats
 
 # Default target - show help
 help:
@@ -168,6 +168,35 @@ loadtest:
 	@echo "==> results landed in deploy/test/loadtest/results/"
 	@if [ -f deploy/test/loadtest/results/summary.txt ]; then cat deploy/test/loadtest/results/summary.txt; fi
 
+# Phase 5 — kind-driven cert-manager integration test. Requires
+# `kind`, `kubectl`, `helm`, and a local Docker daemon. Sets
+# KIND_AVAILABLE=1 so the test runs (it skips cleanly when unset, which
+# is the CI default — kind is too heavy for per-PR CI). The test
+# brings up a fresh cluster, installs cert-manager 1.15, helm-installs
+# certctl-test, applies a ClusterIssuer + Certificate, and asserts the
+# Secret lands.
+acme-cert-manager-test:
+	@echo "==> running cert-manager integration test (requires kind/kubectl/helm)"
+	@KIND_AVAILABLE=1 go test -tags=integration -count=1 -timeout=15m \
+	  ./deploy/test/acme-integration/...
+
+# Phase 5 — RFC 8555 conformance against `lego` driving the certctl
+# server. Hermetic: brings up a single certctl-server via docker
+# compose, points lego at it, runs the conformance scenarios. Skips
+# when the operator hasn't built the test image (`make docker-build`
+# first).
+acme-rfc-conformance-test:
+	@echo "==> running RFC 8555 conformance via lego"
+	@if ! command -v lego >/dev/null 2>&1; then \
+	  echo "lego not installed — go install github.com/go-acme/lego/v4/cmd/lego@latest"; \
+	  exit 1; \
+	fi
+	@cd deploy/test/loadtest && docker compose up -d certctl postgres
+	@sleep 8
+	@CERTCTL_ACME_DIR=https://localhost:8443/acme/profile/prof-test/directory \
+	  bash deploy/test/acme-integration/conformance-lego.sh
+	@cd deploy/test/loadtest && docker compose down
+
 # Database targets (requires migrate tool)
 migrate-up:
 	@echo "Running migrations..."
diff --git a/cmd/server/main.go b/cmd/server/main.go
index 5bf7acc..cce8f2d 100644
--- a/cmd/server/main.go
+++ b/cmd/server/main.go
@@ -785,6 +785,25 @@ func main() {
 	// when absent, falls back to last-33%-of-validity).
 	acmeService.SetRevocationDelegate(revocationSvc)
 	acmeService.SetRenewalPolicyLookup(renewalPolicyRepo)
+	// Phase 5 — per-account rate limiter. In-memory token-buckets,
+	// shared across all entry points (CreateOrder / RotateAccountKey /
+	// RespondToChallenge). Restart wipes counters; orders/hour caps are
+	// eventual-consistency anyway. Persistent rate limiting is a
+	// follow-up if production telemetry shows abuse patterns we can't
+	// catch in a single restart cycle.
+	acmeRateLimiter := acmepkg.NewRateLimiter()
+	acmeService.SetRateLimiter(acmeRateLimiter)
+	// Phase 5 — ACME GC sweeper. Disabled when GCInterval <= 0; the
+	// scheduler.SetACMEGarbageCollector(nil) leg short-circuits in
+	// scheduler.Start (the loopCount + go-routine launch are gated on
+	// non-nil acmeGC). Wired here (not earlier with the other scheduler
+	// loops) because the GC service needs a fully-constructed acmeService.
+	if cfg.ACMEServer.Enabled && cfg.ACMEServer.GCInterval > 0 {
+		sched.SetACMEGarbageCollector(acmeService)
+		sched.SetACMEGCInterval(cfg.ACMEServer.GCInterval)
+		logger.Info("ACME GC scheduler enabled",
+			"interval", cfg.ACMEServer.GCInterval.String())
+	}
 	acmeHandler := handler.NewACMEHandler(acmeService)
 
 	// Build the API router with all handlers
diff --git a/deploy/test/acme-integration/cert-manager-install.sh b/deploy/test/acme-integration/cert-manager-install.sh
new file mode 100755
index 0000000..666249c
--- /dev/null
+++ b/deploy/test/acme-integration/cert-manager-install.sh
@@ -0,0 +1,24 @@
+#!/usr/bin/env bash
+#
+# Phase 5 — install cert-manager 1.15.0 into the kind cluster brought
+# up by kind-config.yaml. Idempotent: re-running waits for the
+# existing deployment to be Ready instead of reinstalling.
+#
+# Called from: deploy/test/acme-integration/certmanager_test.go
+# Standalone: bash deploy/test/acme-integration/cert-manager-install.sh
+set -euo pipefail
+
+CERT_MANAGER_VERSION="${CERT_MANAGER_VERSION:-v1.15.0}"
+KUBECTL="${KUBECTL:-kubectl}"
+
+echo "Installing cert-manager ${CERT_MANAGER_VERSION}..."
+${KUBECTL} apply -f \
+  "https://github.com/cert-manager/cert-manager/releases/download/${CERT_MANAGER_VERSION}/cert-manager.yaml"
+
+echo "Waiting for cert-manager controller to be Ready (timeout 5m)..."
+${KUBECTL} -n cert-manager wait --for=condition=Available --timeout=5m \
+  deployment/cert-manager \
+  deployment/cert-manager-cainjector \
+  deployment/cert-manager-webhook
+
+echo "cert-manager ${CERT_MANAGER_VERSION} ready."
diff --git a/deploy/test/acme-integration/certificate-test.yaml b/deploy/test/acme-integration/certificate-test.yaml
new file mode 100644
index 0000000..8ba154a
--- /dev/null
+++ b/deploy/test/acme-integration/certificate-test.yaml
@@ -0,0 +1,20 @@
+# Phase 5 — Certificate resource the integration test applies and
+# waits for. The certctl-test-trust ClusterIssuer (trust_authenticated
+# mode) issues the cert without any solver round-trip; the resulting
+# Secret 'test-com-tls' is asserted to carry tls.crt + tls.key.
+apiVersion: cert-manager.io/v1
+kind: Certificate
+metadata:
+  name: test-com
+  namespace: default
+spec:
+  secretName: test-com-tls
+  commonName: test.example.com
+  dnsNames:
+    - test.example.com
+    - www.test.example.com
+  issuerRef:
+    name: certctl-test-trust
+    kind: ClusterIssuer
+  duration: 720h     # 30d
+  renewBefore: 240h  # 10d
diff --git a/deploy/test/acme-integration/certmanager_test.go b/deploy/test/acme-integration/certmanager_test.go
new file mode 100644
index 0000000..1957869
--- /dev/null
+++ b/deploy/test/acme-integration/certmanager_test.go
@@ -0,0 +1,167 @@
+// Copyright (c) certctl
+// SPDX-License-Identifier: BSL-1.1
+
+//go:build integration
+
+// Phase 5 — kind-driven cert-manager integration test. Verifies the
+// certctl ACME server end-to-end against a real cert-manager 1.15+
+// deployment in a kind cluster. The test sequences:
+//
+//  1. Bring up the kind cluster (kind-config.yaml).
+//  2. Install cert-manager 1.15 (cert-manager-install.sh).
+//  3. Helm-install certctl-server with acmeServer.enabled=true.
+//  4. Apply the ClusterIssuer + Certificate.
+//  5. Wait for the Certificate to become Ready.
+//  6. Assert the Secret has tls.crt + tls.key.
+//
+// Gated behind KIND_AVAILABLE — CI doesn't run kind and skips this
+// cleanly. Operators run locally via `make acme-cert-manager-test`.
+
+package acmeintegration
+
+import (
+	"context"
+	"fmt"
+	"os"
+	"os/exec"
+	"strings"
+	"testing"
+	"time"
+)
+
+// kindAvailable returns true when the operator opted into the kind-
+// driven test path. CI default is opt-out (env unset → skip).
+func kindAvailable() bool {
+	return os.Getenv("KIND_AVAILABLE") != ""
+}
+
+// kindClusterName is the name passed to `kind create/delete cluster`.
+// Kept as a const so the test cleanup uses the exact same name as
+// setup (avoid orphan-cluster-after-flake).
+const kindClusterName = "certctl-acme-test"
+
+// TestCertManagerTrustAuthenticatedIssuance is the happy-path
+// integration: cert-manager submits a new-order against a profile in
+// trust_authenticated mode; certctl auto-resolves authzs (no solver
+// round-trip in this mode); cert-manager finalizes; the Secret lands.
+//
+// Runtime: ~6-8 minutes wall-clock on a workstation (most of which is
+// kind-create + cert-manager-controller-bootstrap, both cached on
+// re-runs after the first). Skips cleanly when KIND_AVAILABLE is
+// unset.
+func TestCertManagerTrustAuthenticatedIssuance(t *testing.T) {
+	if !kindAvailable() {
+		t.Skip("KIND_AVAILABLE unset — kind-driven cert-manager integration test skipped")
+	}
+	ctx := context.Background()
+
+	t.Log("creating kind cluster")
+	runCmd(t, ctx, "kind", "create", "cluster",
+		"--name", kindClusterName,
+		"--config", "kind-config.yaml")
+	t.Cleanup(func() {
+		// Best-effort cluster teardown — never fail the test on cleanup
+		// failure (operator can `kind delete cluster` manually).
+		_ = exec.Command("kind", "delete", "cluster", "--name", kindClusterName).Run()
+	})
+
+	t.Log("installing cert-manager")
+	runCmd(t, ctx, "bash", "cert-manager-install.sh")
+
+	// Step 3 — deploy certctl-server. The Helm chart at
+	// deploy/helm/certctl/ takes acmeServer.enabled=true; the operator
+	// is expected to have built + pushed (or kind-loaded) a `:test`
+	// image tag before the test runs. Document this in docs/acme-server.md.
+	t.Log("helm-installing certctl-test")
+	runCmd(t, ctx, "helm", "install", "certctl-test", "../../helm/certctl/",
+		"--set", "acmeServer.enabled=true",
+		"--set", "acmeServer.defaultProfileId=prof-test",
+		"--set", "image.tag=test",
+	)
+	waitForDeploymentReady(t, ctx, "default", "certctl-test", 3*time.Minute)
+
+	t.Log("applying ClusterIssuer + Certificate")
+	runCmd(t, ctx, "kubectl", "apply", "-f", "clusterissuer-trust-authenticated.yaml")
+	runCmd(t, ctx, "kubectl", "apply", "-f", "certificate-test.yaml")
+
+	t.Log("waiting for Certificate to become Ready")
+	waitForCertificateReady(t, ctx, "default", "test-com", 3*time.Minute)
+
+	t.Log("asserting Secret has tls.crt")
+	assertSecretHasCert(t, ctx, "default", "test-com-tls")
+
+	t.Log("happy-path issuance verified end-to-end")
+}
+
+// runCmd runs the command; failures fail the test immediately. We
+// stream combined stdout+stderr to t.Log on completion so the operator
+// can read the kubectl/kind output in CI logs (when run there with
+// KIND_AVAILABLE=1).
+func runCmd(t *testing.T, ctx context.Context, name string, args ...string) {
+	t.Helper()
+	cmd := exec.CommandContext(ctx, name, args...) //nolint:gosec // ARGS are test-controlled literals.
+	out, err := cmd.CombinedOutput()
+	if err != nil {
+		t.Fatalf("%s %s failed: %v\n%s", name, strings.Join(args, " "), err, out)
+	}
+	t.Logf("%s %s: %s", name, strings.Join(args, " "), strings.TrimSpace(string(out)))
+}
+
+// waitForDeploymentReady polls until the named deployment reports
+// Available=True. Wraps `kubectl wait` with a Go-level timeout so test
+// hangs are bounded.
+func waitForDeploymentReady(t *testing.T, ctx context.Context, namespace, name string, timeout time.Duration) {
+	t.Helper()
+	cctx, cancel := context.WithTimeout(ctx, timeout)
+	defer cancel()
+	cmd := exec.CommandContext(cctx, "kubectl", "-n", namespace, "wait",
+		"--for=condition=Available", fmt.Sprintf("--timeout=%ds", int(timeout.Seconds())),
+		"deployment/"+name) //nolint:gosec // ARGS are test-controlled literals.
+	out, err := cmd.CombinedOutput()
+	if err != nil {
+		t.Fatalf("deployment %s/%s did not become Ready in %v: %v\n%s",
+			namespace, name, timeout, err, out)
+	}
+}
+
+// waitForCertificateReady polls until the cert-manager Certificate
+// resource transitions to Ready=True. cert-manager's own
+// reconciliation loop is what advances the state; this just blocks
+// until the controller is happy.
+func waitForCertificateReady(t *testing.T, ctx context.Context, namespace, name string, timeout time.Duration) {
+	t.Helper()
+	cctx, cancel := context.WithTimeout(ctx, timeout)
+	defer cancel()
+	cmd := exec.CommandContext(cctx, "kubectl", "-n", namespace, "wait",
+		"--for=condition=Ready", fmt.Sprintf("--timeout=%ds", int(timeout.Seconds())),
+		"certificate/"+name) //nolint:gosec // ARGS are test-controlled literals.
+	out, err := cmd.CombinedOutput()
+	if err != nil {
+		// Dump the Certificate's events on failure so the operator
+		// can see exactly which reconciliation step failed.
+		describe := exec.Command("kubectl", "-n", namespace, "describe", "certificate", name)
+		describeOut, _ := describe.CombinedOutput()
+		t.Fatalf("certificate %s/%s did not become Ready in %v: %v\n%s\n--- describe ---\n%s",
+			namespace, name, timeout, err, out, describeOut)
+	}
+}
+
+// assertSecretHasCert checks that the named Secret has a non-empty
+// tls.crt entry. We don't validate the chain itself here — that's the
+// job of certctl's own integration test layer; this just confirms
+// cert-manager wrote something into the Secret on the
+// trust_authenticated happy-path.
+func assertSecretHasCert(t *testing.T, ctx context.Context, namespace, name string) {
+	t.Helper()
+	cctx, cancel := context.WithTimeout(ctx, 30*time.Second)
+	defer cancel()
+	cmd := exec.CommandContext(cctx, "kubectl", "-n", namespace, "get", "secret", name,
+		"-o", "jsonpath={.data.tls\\.crt}") //nolint:gosec // ARGS are test-controlled literals.
+	out, err := cmd.CombinedOutput()
+	if err != nil {
+		t.Fatalf("get secret %s/%s: %v\n%s", namespace, name, err, out)
+	}
+	if len(out) == 0 {
+		t.Fatalf("secret %s/%s has empty tls.crt", namespace, name)
+	}
+}
diff --git a/deploy/test/acme-integration/clusterissuer-challenge.yaml b/deploy/test/acme-integration/clusterissuer-challenge.yaml
new file mode 100644
index 0000000..7b7e748
--- /dev/null
+++ b/deploy/test/acme-integration/clusterissuer-challenge.yaml
@@ -0,0 +1,31 @@
+# Phase 5 — sample ClusterIssuer for the certctl challenge auth mode
+# (RFC 8555 §8 HTTP-01 / DNS-01 / TLS-ALPN-01). Use this for public-
+# trust-style deployments where per-identifier ownership proof is
+# required.
+#
+# Same bootstrap-root caBundle requirement as the trust_authenticated
+# variant — see clusterissuer-trust-authenticated.yaml comments.
+apiVersion: cert-manager.io/v1
+kind: ClusterIssuer
+metadata:
+  name: certctl-test-challenge
+spec:
+  acme:
+    email: test@example.com
+    # Point at a profile whose certificate_profiles.acme_auth_mode is
+    # set to 'challenge'. The certctl operator manages this column
+    # per-profile; see certctl/docs/acme-server.md "Per-profile auth
+    # mode" section.
+    server: https://certctl-test.default.svc.cluster.local:8443/acme/profile/prof-challenge/directory
+    caBundle: |
+      LS0tLS1CRUdJTiBDRVJUSUZJQ0FURS0tLS0tCi4uLgotLS0tLUVORCBDRVJUSUZJQ0FURS0tLS0tCg==
+    privateKeySecretRef:
+      name: certctl-test-challenge-account-key
+    solvers:
+      # HTTP-01 via the in-cluster ingress-nginx. The cert-manager
+      # http-solver pod publishes the key authorization at
+      # http://<identifier>/.well-known/acme-challenge/<token>; the
+      # certctl HTTP01Validator (Phase 3) fetches it.
+      - http01:
+          ingress:
+            class: nginx
diff --git a/deploy/test/acme-integration/clusterissuer-trust-authenticated.yaml b/deploy/test/acme-integration/clusterissuer-trust-authenticated.yaml
new file mode 100644
index 0000000..47437ed
--- /dev/null
+++ b/deploy/test/acme-integration/clusterissuer-trust-authenticated.yaml
@@ -0,0 +1,42 @@
+# Phase 5 — sample ClusterIssuer for the certctl trust_authenticated
+# auth mode (RFC 8555 §6 + certctl auth_mode=trust_authenticated, where
+# the JWS-authenticated ACME account is trusted to issue any identifier
+# the profile policy permits — no per-identifier ownership challenges).
+#
+# Use this as the starting template for any internal-PKI rollout.
+# Replace the caBundle placeholder with the base64-encoded PEM of the
+# certctl-server's self-signed bootstrap root, then `kubectl apply`.
+#
+# Generate the caBundle via:
+#   cat deploy/test/certs/ca.crt | base64 -w0
+# (See certctl/docs/acme-server.md "TLS trust bootstrap" section for the
+# end-to-end walkthrough — this is the single biggest first-time-deploy
+# footgun on cert-manager, captured as audit fix #9.)
+apiVersion: cert-manager.io/v1
+kind: ClusterIssuer
+metadata:
+  name: certctl-test-trust
+spec:
+  acme:
+    email: test@example.com
+    # Replace 'certctl-test' with your release name + adjust the
+    # profile path segment. Default profile path:
+    #   https://<service>.<namespace>.svc.cluster.local:8443/acme/profile/<profile-id>/directory
+    server: https://certctl-test.default.svc.cluster.local:8443/acme/profile/prof-test/directory
+    # caBundle: Audit fix #9. cert-manager validates the ACME server's
+    # TLS chain before submitting any account/order/finalize. With a
+    # self-signed bootstrap root, the ClusterIssuer MUST carry the root
+    # explicitly via this field.
+    caBundle: |
+      LS0tLS1CRUdJTiBDRVJUSUZJQ0FURS0tLS0tCi4uLgotLS0tLUVORCBDRVJUSUZJQ0FURS0tLS0tCg==
+    privateKeySecretRef:
+      name: certctl-test-trust-account-key
+    solvers:
+      # In trust_authenticated mode the solver is unused at the
+      # validation step but cert-manager still requires at least one
+      # solver in the spec. http01-via-ingress-nginx is the cheapest
+      # placeholder shape that round-trips correctly through cert-
+      # manager's validation webhooks.
+      - http01:
+          ingress:
+            class: nginx
diff --git a/deploy/test/acme-integration/conformance-lego.sh b/deploy/test/acme-integration/conformance-lego.sh
new file mode 100755
index 0000000..f43df89
--- /dev/null
+++ b/deploy/test/acme-integration/conformance-lego.sh
@@ -0,0 +1,56 @@
+#!/usr/bin/env bash
+#
+# Phase 5 — lego-driven RFC 8555 conformance test. Drives a real ACME
+# client (lego v4) against the certctl ACME server in trust_authenticated
+# mode and exercises the full happy-path: register → new-order →
+# finalize → cert download.
+#
+# Caller (`make acme-rfc-conformance-test`) brings up the certctl
+# docker-compose stack first; this script just runs lego against it.
+#
+# Skips cleanly when CERTCTL_ACME_DIR is unset (the operator probably
+# meant to run the make target instead of this script directly).
+set -euo pipefail
+
+if [[ -z "${CERTCTL_ACME_DIR:-}" ]]; then
+  echo "CERTCTL_ACME_DIR unset — point at the certctl ACME directory URL"
+  echo "  e.g. CERTCTL_ACME_DIR=https://localhost:8443/acme/profile/prof-test/directory"
+  exit 1
+fi
+
+WORKDIR="$(mktemp -d -t certctl-lego-conf-XXXXXX)"
+trap 'rm -rf "${WORKDIR}"' EXIT
+
+# Skip TLS verification — the test stack uses certctl's self-signed
+# bootstrap cert. Operators in production use --insecure-skip-verify=false
+# and pass --tls-bundle for the real CA.
+LEGO_INSECURE="--insecure-skip-verify"
+
+# Step 1: register a fresh account.
+echo "==> lego: register account"
+lego --server "${CERTCTL_ACME_DIR}" \
+     --email conformance@example.com \
+     --domains conformance.example.com \
+     --path "${WORKDIR}" \
+     --accept-tos \
+     ${LEGO_INSECURE} \
+     register
+
+# Step 2: issue a cert (trust_authenticated mode auto-resolves authzs).
+echo "==> lego: run (issue conformance.example.com)"
+lego --server "${CERTCTL_ACME_DIR}" \
+     --email conformance@example.com \
+     --domains conformance.example.com \
+     --path "${WORKDIR}" \
+     --accept-tos \
+     ${LEGO_INSECURE} \
+     run
+
+# Step 3: assert the cert PEM landed.
+CERT_FILE="${WORKDIR}/certificates/conformance.example.com.crt"
+if [[ ! -s "${CERT_FILE}" ]]; then
+  echo "FAIL: ${CERT_FILE} is missing or empty"
+  exit 1
+fi
+openssl x509 -in "${CERT_FILE}" -noout -subject -issuer -dates
+echo "PASS: lego conformance happy-path completed"
diff --git a/deploy/test/acme-integration/kind-config.yaml b/deploy/test/acme-integration/kind-config.yaml
new file mode 100644
index 0000000..0ee6f96
--- /dev/null
+++ b/deploy/test/acme-integration/kind-config.yaml
@@ -0,0 +1,34 @@
+# Phase 5 — kind-cluster shape for the cert-manager integration test.
+#
+# Single control-plane + single worker. Port 8443 (certctl ACME server)
+# and 80/443 (ingress-nginx for HTTP-01 solver) are extra-mapped onto
+# the host so the in-test workflow can curl the in-cluster services.
+#
+# Used by: deploy/test/acme-integration/certmanager_test.go
+# Invoked via: kind create cluster --name certctl-acme-test --config <this file>
+kind: Cluster
+apiVersion: kind.x-k8s.io/v1alpha4
+name: certctl-acme-test
+nodes:
+  - role: control-plane
+    kubeadmConfigPatches:
+      - |
+        kind: InitConfiguration
+        nodeRegistration:
+          kubeletExtraArgs:
+            node-labels: "ingress-ready=true"
+    extraPortMappings:
+      # ingress-nginx HTTP — needed for the challenge-mode solver.
+      - containerPort: 80
+        hostPort: 80
+        protocol: TCP
+      - containerPort: 443
+        hostPort: 443
+        protocol: TCP
+      # certctl-server HTTPS (the ACME directory + JWS-authenticated
+      # POST surface). Only required for out-of-cluster smoke tests; the
+      # in-cluster ClusterIssuer talks via Service DNS.
+      - containerPort: 30843
+        hostPort: 8443
+        protocol: TCP
+  - role: worker
diff --git a/deploy/test/loadtest/README.md b/deploy/test/loadtest/README.md
index 3d470f2..8bae1fa 100644
--- a/deploy/test/loadtest/README.md
+++ b/deploy/test/loadtest/README.md
@@ -313,7 +313,47 @@ deploy/test/loadtest/
 └── results/          (gitignored — k6 writes summary.{json,txt} here)
 ```
 
+## ACME flows (Phase 5)
+
+The `deploy/test/loadtest/k6/acme_flow.js` scenario hammers the
+unauthenticated ACME surface (directory + new-nonce + ARI synthetic
+lookups) at constant 100 VUs for 5 minutes. JWS-signed paths
+(new-account / new-order / finalize) are intentionally out of scope:
+k6 doesn't ship JWS, and bundling lego inside k6 would obscure the
+underlying-server p95 we're trying to measure. Instead, the
+`make acme-rfc-conformance-test` target drives lego against the same
+stack for the full happy-path conformance gate.
+
+Run it:
+
+```
+cd deploy/test/loadtest
+docker compose up -d certctl postgres
+k6 run --env CERTCTL_ACME_DIRECTORY=https://localhost:8443/acme/profile/prof-test/directory \
+       k6/acme_flow.js
+```
+
+### Baseline (ACME flows, 100 VUs × 5m)
+
+The baseline is operator-captured on a workstation-class machine with
+a single certctl-server container + a single postgres container.
+Re-capture after schema migrations or transport changes; commit the
+new numbers so regressions are visible in code review.
+
+| Metric                                     | Threshold | Last captured | Notes |
+|--------------------------------------------|-----------|---------------|-------|
+| `directory_duration` p95                   | < 500 ms  | _operator_    | Unauth GET; cache-friendly. |
+| `new_nonce_duration` p95                   | < 300 ms  | _operator_    | Single Postgres INSERT under the hood. |
+| `renewal_info_duration` p95 (synthetic id) | < 800 ms  | _operator_    | Synthetic cert-id → 4xx fast path. |
+| `http_req_failed` rate                     | < 1%      | _operator_    | Should be ~0 — failures here mean transport issues. |
+
+Capture command: `make loadtest` after pointing the compose stack at
+the ACME flow scenario. Operators with kind / cert-manager available
+should pair this with `make acme-cert-manager-test` for end-to-end
+verification.
+
 ## Audit references
 
 - API tier:       `cowork/issuer-coverage-audit-2026-05-01/RESULTS.md` fix #8.
 - Connector tier: `cowork/deployment-target-audit-2026-05-02/RESULTS.md` Bundle 10.
+- ACME flows:     Phase 5 master prompt (`cowork/acme-server-prompts/06-phase-5-certmanager-hardening-prompt.md`).
diff --git a/deploy/test/loadtest/k6/acme_flow.js b/deploy/test/loadtest/k6/acme_flow.js
new file mode 100644
index 0000000..8aa1f53
--- /dev/null
+++ b/deploy/test/loadtest/k6/acme_flow.js
@@ -0,0 +1,80 @@
+// Phase 5 — k6 scenario for the ACME issuance loop. Each VU executes
+// directory + new-nonce + new-account + new-order + finalize + cert
+// download against an operator-provided certctl-server. Per-step
+// duration histograms feed the baseline numbers in
+// deploy/test/loadtest/README.md (ACME flows section).
+//
+// Default scenario: 100 concurrent VUs for 5 minutes. Override via
+// K6_VUS / K6_DURATION env vars.
+//
+// Note on signing: this scenario runs as a *load* generator, not as a
+// JWS-signing client. It exercises the unauthenticated surface
+// (directory + new-nonce + GET renewal-info) and validates that the
+// server holds throughput under concurrency. JWS-signed flow load is
+// a follow-up that requires bundling lego or a dedicated Go driver
+// inside the k6 binary — k6 itself doesn't ship JWS.
+
+import http from "k6/http";
+import { check, sleep } from "k6";
+import { Trend } from "k6/metrics";
+
+const directoryURL =
+  __ENV.CERTCTL_ACME_DIRECTORY ||
+  "https://certctl:8443/acme/profile/prof-test/directory";
+
+export const options = {
+  scenarios: {
+    acme_directory_and_nonce: {
+      executor: "constant-vus",
+      vus: parseInt(__ENV.K6_VUS || "100", 10),
+      duration: __ENV.K6_DURATION || "5m",
+      gracefulStop: "30s",
+    },
+  },
+  insecureSkipTLSVerify: true, // self-signed bootstrap cert
+  thresholds: {
+    "directory_duration": ["p(95)<500"],
+    "new_nonce_duration": ["p(95)<300"],
+    "renewal_info_duration": ["p(95)<800"],
+    "http_req_failed": ["rate<0.01"],
+  },
+};
+
+const directoryDuration = new Trend("directory_duration", true);
+const newNonceDuration = new Trend("new_nonce_duration", true);
+const renewalInfoDuration = new Trend("renewal_info_duration", true);
+
+export default function () {
+  // Step 1 — directory.
+  let res = http.get(directoryURL);
+  directoryDuration.add(res.timings.duration);
+  check(res, { "directory 200": (r) => r.status === 200 });
+
+  if (res.status !== 200) return;
+  const dir = res.json();
+
+  // Step 2 — new-nonce.
+  if (dir.newNonce) {
+    res = http.head(dir.newNonce);
+    newNonceDuration.add(res.timings.duration);
+    check(res, {
+      "new-nonce 200 + Replay-Nonce": (r) =>
+        r.status === 200 && !!r.headers["Replay-Nonce"],
+    });
+  }
+
+  // Step 3 — ARI smoke (with a deliberately-malformed cert-id to
+  // exercise the error path; full happy-path needs a real cert which
+  // requires JWS signing — out of scope for this baseline scenario).
+  if (dir.renewalInfo) {
+    res = http.get(dir.renewalInfo + "/" + "aaaa.bbbb");
+    renewalInfoDuration.add(res.timings.duration);
+    // 400 (malformed cert-id, expected) OR 404 (cert not found).
+    check(res, {
+      "renewal-info 4xx for synthetic cert-id": (r) =>
+        r.status === 400 || r.status === 404,
+    });
+  }
+
+  sleep(1);
+}
diff --git a/docs/acme-server.md b/docs/acme-server.md
index 628d4fa..c5a2794 100644
--- a/docs/acme-server.md
+++ b/docs/acme-server.md
@@ -7,13 +7,15 @@ as an ACME issuer with no certctl-side modification — closing the
 "deploy a certctl agent on every K8s node" friction that costs deals to
 external PKI vendors today.
 
-> **Phase status (2026-05-03):** Phase 4 — closes the RFC 8555 surface
-> beyond the issuance happy-path: doubly-signed key rollover (§7.3.5),
-> revoke-cert via either account-key or cert-key (§7.6), and RFC 9773
-> ACME Renewal Information. ACME clients can now rotate their account
-> keys, revoke certs through the ACME surface (rather than only the
-> certctl GUI/API), and fetch ARI for proactive renewal scheduling.
-> Track shipped phases via `git log --grep='acme-server:'`.
+> **Phase status (2026-05-03):** Phase 5 — production hardening +
+> cert-manager integration test. Per-account rate limits applied at
+> 3 entry points (orders/hour, key-change/hour, challenge-respond/hour)
+> + a per-account concurrent-orders cap; a 1-minute scheduler loop
+> sweeps expired nonces / authzs / orders. A kind-driven cert-manager
+> integration test (gated by `KIND_AVAILABLE`) verifies the full
+> happy-path against a real cert-manager 1.15+ deployment. RFC
+> conformance is verified via lego against the same stack. Track
+> shipped phases via `git log --grep='acme-server:'`.
 
 ## Configuration
 
@@ -40,6 +42,11 @@ issuer connector). The struct definition lives in
 | `CERTCTL_ACME_SERVER_TLSALPN01_CONCURRENCY`      | `10`                   | 3     | Reserved. |
 | `CERTCTL_ACME_SERVER_ARI_ENABLED`                | `true`                 | 4     | Toggles the RFC 9773 ARI surface — both the `renewalInfo` URL in the directory document and the GET `/renewal-info/<cert-id>` handler. Set to `false` to drop ARI from the directory; ACME clients fall back to static renewal scheduling. |
 | `CERTCTL_ACME_SERVER_ARI_POLL_INTERVAL`          | `6h`                   | 4     | Server-policy `Retry-After` value the ARI handler emits on a 200 response. RFC 9773 §4.2 leaves this server-policy. Tighten to `1h` for short-lived certs; loosen to `24h` for standard 90-day certs. |
+| `CERTCTL_ACME_SERVER_RATE_LIMIT_ORDERS_PER_HOUR` | `100`                  | 5     | Per-account orders/hour cap. `0` disables. Hits return RFC 7807 + RFC 8555 §6.7 `urn:ietf:params:acme:error:rateLimited` with `Retry-After`. In-memory token-bucket; restart wipes the counter (eventual-consistency caps are acceptable). |
+| `CERTCTL_ACME_SERVER_RATE_LIMIT_CONCURRENT_ORDERS` | `5`                  | 5     | Per-account cap on simultaneously-active orders (status in pending/ready/processing). `0` disables. Same RFC 7807 + RFC 8555 §6.7 problem shape as the per-hour cap. |
+| `CERTCTL_ACME_SERVER_RATE_LIMIT_KEY_CHANGE_PER_HOUR` | `5`                | 5     | Per-account key-rollover cap. `0` disables. Default 5/hour: rollovers should be rare; a flood is an attack signal. |
+| `CERTCTL_ACME_SERVER_RATE_LIMIT_CHALLENGE_RESPONDS_PER_HOUR` | `60`       | 5     | Per-challenge-id respond cap. `0` disables. Defends against retry storms from a misbehaving client. Keyed by challenge-id (not account-id) so a flood against one challenge doesn't drain the account's whole budget. |
+| `CERTCTL_ACME_SERVER_GC_INTERVAL`                | `1m`                   | 5     | Tick interval for the ACME GC scheduler loop. On each tick: (1) DELETE used / expired nonces; (2) UPDATE pending authzs whose `expires_at < NOW()` to `expired`; (3) UPDATE pending/ready/processing orders whose `expires_at < NOW()` to `invalid`. Each sweep is a single SQL statement; the loop is idempotent + bounded by a 1m per-sweep timeout. `0` disables the loop. |
 
 ## Per-profile auth mode
 
@@ -206,7 +213,7 @@ at `internal/service/certificate.go:131`).
 | 2     | live        | orders + authzs + finalize + cert download (trust_authenticated mode end-to-end) |
 | 3     | live        | HTTP-01 + DNS-01 + TLS-ALPN-01 challenge validation (challenge mode end-to-end) |
 | 4     | live        | key rollover (RFC 8555 §7.3.5) + revoke-cert (§7.6) + ARI (RFC 9773) |
-| 5     | not yet     | cert-manager integration test + production hardening |
+| 5     | live        | rate limits + GC sweeper + kind-driven cert-manager integration test + lego conformance harness + k6 ACME-flow scenario |
 | 6     | not yet     | full operator-facing reference + walkthroughs + threat model |
 
 Track shipped phases via `git log --grep='acme-server:' --oneline`.
@@ -306,3 +313,76 @@ Window math:
 Disable ARI globally with `CERTCTL_ACME_SERVER_ARI_ENABLED=false`. The
 URL drops out of the directory; the route is still registered but
 returns 404 — clients fall back to static renewal scheduling.
+
+## Phase 5 — operational guidance
+
+### Rate limiting
+
+Production deployments serving multiple ACME profiles or fleets should
+keep the default rate limits in place. The four caps:
+
+- `RATE_LIMIT_ORDERS_PER_HOUR` (100) — per-account new-order cap. A
+  cert-manager Certificate that auto-renews at the 1/3 mark of its
+  validity (90-day cert → ~30-day renewal) consumes ~12 orders/year
+  per managed Certificate. 100/hour is generous for any plausible
+  fleet.
+- `RATE_LIMIT_CONCURRENT_ORDERS` (5) — per-account cap on
+  pending/ready/processing orders. Stops a runaway client from
+  starving DB-row throughput. Tune up only if you observe legitimate
+  bursts.
+- `RATE_LIMIT_KEY_CHANGE_PER_HOUR` (5) — rollovers are rare; a flood
+  is an attack signal. Tune down to 1/hour if your operator
+  procedure mandates manual rollovers only.
+- `RATE_LIMIT_CHALLENGE_RESPONDS_PER_HOUR` (60) — per-challenge cap,
+  defends against retry storms.
+
+Hits return RFC 8555 §6.7 `rateLimited` Problem with a `Retry-After`
+header. cert-manager 1.15+ honors the header; lego too. Older clients
+may not — that's the client's problem, not certctl's.
+
+The buckets are **in-memory + per-replica**. A 3-replica certctl-
+server fleet behind a load balancer effectively has 3× the configured
+throughput (each replica's bucket fills independently). For
+deployments where this matters operationally, the right answer is a
+shared rate-limit store — that's a follow-up; not blocking for the
+current threat model where same-account requests typically pin to
+the same replica via session affinity.
+
+### GC sweeper
+
+The scheduler runs the GC sweep every `GC_INTERVAL` (default 1m). Each
+sweep is three independent SQL statements:
+
+1. `DELETE FROM acme_nonces WHERE used = TRUE OR expires_at < NOW()`.
+2. `UPDATE acme_authorizations SET status='expired' WHERE status='pending' AND expires_at < NOW()`.
+3. `UPDATE acme_orders SET status='invalid', error=... WHERE status IN ('pending','ready','processing') AND expires_at < NOW()`.
+
+Each statement is bounded by a 1-minute per-sweep timeout. A failing
+sweep is logged + retried on the next tick; a tick that overruns its
+budget is skipped (the existing-tick atomic-Bool guard prevents
+overlap). Counts are exposed via `certctl_acme_gc_*` Prometheus
+metrics.
+
+### cert-manager integration test
+
+`make acme-cert-manager-test` brings up a kind cluster, installs
+cert-manager 1.15.0, helm-deploys certctl-server with
+`acmeServer.enabled=true`, and verifies a Certificate resource issues
+end-to-end. Skipped in CI by default (kind is too heavy for per-PR);
+operators run locally on workstation. See
+`deploy/test/acme-integration/` for the YAML + Go test harness.
+
+### lego RFC conformance harness
+
+`make acme-rfc-conformance-test` drives lego v4 against a hermetic
+certctl-server stack, exercising register → new-order → finalize.
+Operators run this when shipping behavior changes to the ACME surface
+to confirm a real third-party client still works.
+
+### k6 ACME flows scenario
+
+`deploy/test/loadtest/k6/acme_flow.js` exercises the unauthenticated
+surface (directory + new-nonce + ARI) at 100 VUs × 5m. JWS-signed
+flows are out of scope for k6 (no JWS support); they're covered by
+the lego conformance harness above. Baseline numbers + thresholds in
+`deploy/test/loadtest/README.md`.
diff --git a/internal/api/acme/ratelimit.go b/internal/api/acme/ratelimit.go
new file mode 100644
index 0000000..48881cb
--- /dev/null
+++ b/internal/api/acme/ratelimit.go
@@ -0,0 +1,166 @@
+// Copyright (c) certctl
+// SPDX-License-Identifier: BSL-1.1
+
+package acme
+
+import (
+	"errors"
+	"sync"
+	"time"
+)
+
+// Phase 5 — per-account rolling-hour rate limiter for ACME operations.
+//
+// Architecture:
+//   - In-memory token-bucket per (key, action). Restart wipes the
+//     buckets; orders/hour caps are eventual-consistency so this is
+//     acceptable. Persistent rate limiting is a follow-up if production
+//     telemetry shows abuse patterns we can't catch in a single restart
+//     cycle (master prompt criterion #11 explicitly accepts this).
+//   - Tokens-per-hour math: bucket capacity = perHour, refill rate =
+//     perHour / 3600 tokens/sec. A fresh bucket starts full; an over-
+//     limit caller drains it then has to wait for replenishment.
+//   - Key shape is action-specific: orders use accountID; key-rollover
+//     uses accountID; challenge-respond uses challengeID (so a flood
+//     against one challenge doesn't burn the whole account's budget).
+//
+// Concurrency: the outer map is RWMutex-guarded for create-on-demand;
+// per-bucket allow() takes a tiny per-bucket Mutex. Mirrors the
+// existing internal/api/middleware/middleware.go::keyedRateLimiter
+// pattern (different scope, same shape).
+
+// RateLimiter is the per-action token-bucket pool. Construct with
+// NewRateLimiter(); pass a single instance into ACMEService via
+// SetRateLimiter so all entry points share the same buckets.
+type RateLimiter struct {
+	mu      sync.RWMutex
+	buckets map[string]*rlBucket // keyed by "<action>|<keyID>"
+	clock   func() time.Time     // injectable for tests
+}
+
+// NewRateLimiter returns an empty RateLimiter. Buckets are created on
+// first reference, so a fresh limiter does no work until traffic
+// arrives.
+func NewRateLimiter() *RateLimiter {
+	return &RateLimiter{
+		buckets: make(map[string]*rlBucket),
+		clock:   time.Now,
+	}
+}
+
+// SetClock replaces the clock for tests. Production callers leave it
+// pointing at time.Now (the constructor default).
+func (r *RateLimiter) SetClock(now func() time.Time) {
+	if now != nil {
+		r.clock = now
+	}
+}
+
+// Allow returns true when the (action, keyID) bucket has at least one
+// token available — and consumes that token. perHour=0 disables the
+// limit (always true). Negative perHour is treated as 0.
+//
+// On hit (first call → first token consumed → returns true). Once
+// drained, further calls within the same hour return false until
+// elapsed-time refills the bucket.
+func (r *RateLimiter) Allow(action, keyID string, perHour int) bool {
+	if perHour <= 0 {
+		return true
+	}
+	bucketKey := action + "|" + keyID
+	r.mu.RLock()
+	b, ok := r.buckets[bucketKey]
+	r.mu.RUnlock()
+	if !ok {
+		r.mu.Lock()
+		b, ok = r.buckets[bucketKey]
+		if !ok {
+			b = &rlBucket{
+				capacity:   float64(perHour),
+				refillRate: float64(perHour) / 3600.0, // tokens/sec
+				tokens:     float64(perHour),
+				lastRefill: r.clock(),
+			}
+			r.buckets[bucketKey] = b
+		}
+		r.mu.Unlock()
+	}
+	return b.allow(r.clock)
+}
+
+// RetryAfter returns the duration the caller should wait before the
+// (action, keyID) bucket has at least one token again. Returns 0 when
+// at least one token is currently available. Used by the handler to
+// emit a Retry-After header on rateLimited responses.
+func (r *RateLimiter) RetryAfter(action, keyID string, perHour int) time.Duration {
+	if perHour <= 0 {
+		return 0
+	}
+	bucketKey := action + "|" + keyID
+	r.mu.RLock()
+	b, ok := r.buckets[bucketKey]
+	r.mu.RUnlock()
+	if !ok {
+		return 0
+	}
+	b.mu.Lock()
+	defer b.mu.Unlock()
+	if b.tokens >= 1 {
+		return 0
+	}
+	missing := 1 - b.tokens
+	if b.refillRate <= 0 {
+		// Shouldn't happen (Allow rejects perHour<=0 before bucket
+		// creation), but a divide-by-zero here would panic.
+		return time.Hour
+	}
+	secs := missing / b.refillRate
+	return time.Duration(secs * float64(time.Second))
+}
+
+// rlBucket is the per-(action, keyID) token bucket. Mirrors the shape
+// of internal/api/middleware/middleware.go::tokenBucket but with a
+// per-hour-shaped refill instead of per-second.
+type rlBucket struct {
+	mu         sync.Mutex
+	capacity   float64
+	refillRate float64 // tokens per second
+	tokens     float64
+	lastRefill time.Time
+}
+
+func (b *rlBucket) allow(clock func() time.Time) bool {
+	b.mu.Lock()
+	defer b.mu.Unlock()
+
+	now := clock()
+	// Monotonic-clock-safe via t.Sub(t) per Go time-package contract.
+	elapsed := now.Sub(b.lastRefill).Seconds()
+	if elapsed > 0 {
+		b.tokens += elapsed * b.refillRate
+		if b.tokens > b.capacity {
+			b.tokens = b.capacity
+		}
+		b.lastRefill = now
+	}
+	if b.tokens < 1 {
+		return false
+	}
+	b.tokens--
+	return true
+}
+
+// Action constants — keep one source of truth for the bucket-key
+// `<action>|...` prefix. Using untyped consts (not iota) so they
+// survive cross-process coordination if a follow-up adds shared-state
+// rate-limiting.
+const (
+	ActionNewOrder         = "new_order"
+	ActionKeyChange        = "key_change"
+	ActionChallengeRespond = "challenge_respond"
+)
+
+// ErrRateLimited is the sentinel service-layer entry points return on
+// a hit. Handler maps to RFC 7807 + RFC 8555 §6.7
+// `urn:ietf:params:acme:error:rateLimited` with Retry-After.
+var ErrRateLimited = errors.New("acme: rate limit exceeded")
diff --git a/internal/api/acme/ratelimit_test.go b/internal/api/acme/ratelimit_test.go
new file mode 100644
index 0000000..a02f2dd
--- /dev/null
+++ b/internal/api/acme/ratelimit_test.go
@@ -0,0 +1,159 @@
+// Copyright (c) certctl
+// SPDX-License-Identifier: BSL-1.1
+
+package acme
+
+import (
+	"sync"
+	"testing"
+	"time"
+)
+
+// Phase 5 — RateLimiter unit tests.
+
+func TestRateLimiter_DisabledWhenPerHourZero(t *testing.T) {
+	r := NewRateLimiter()
+	for i := 0; i < 10000; i++ {
+		if !r.Allow(ActionNewOrder, "acc-1", 0) {
+			t.Fatalf("Allow returned false on call %d with perHour=0", i)
+		}
+	}
+}
+
+func TestRateLimiter_DisabledWhenPerHourNegative(t *testing.T) {
+	r := NewRateLimiter()
+	if !r.Allow(ActionNewOrder, "acc-1", -5) {
+		t.Errorf("Allow returned false with perHour=-5; expected always-allow")
+	}
+}
+
+func TestRateLimiter_BucketCapacity(t *testing.T) {
+	// Frozen clock: a fresh bucket has perHour tokens. Drain exactly
+	// that many; the next call must return false.
+	now := time.Date(2026, 5, 3, 12, 0, 0, 0, time.UTC)
+	r := NewRateLimiter()
+	r.SetClock(func() time.Time { return now })
+
+	for i := 0; i < 100; i++ {
+		if !r.Allow(ActionNewOrder, "acc-1", 100) {
+			t.Fatalf("Allow returned false on call %d (within capacity)", i)
+		}
+	}
+	if r.Allow(ActionNewOrder, "acc-1", 100) {
+		t.Errorf("Allow returned true on the 101st call; expected limit hit")
+	}
+}
+
+func TestRateLimiter_PerKeyIsolation(t *testing.T) {
+	// Frozen clock — drain acc-1 to zero, then acc-2 should still have
+	// a full bucket (separate key).
+	now := time.Date(2026, 5, 3, 12, 0, 0, 0, time.UTC)
+	r := NewRateLimiter()
+	r.SetClock(func() time.Time { return now })
+
+	for i := 0; i < 100; i++ {
+		_ = r.Allow(ActionNewOrder, "acc-1", 100)
+	}
+	if r.Allow(ActionNewOrder, "acc-1", 100) {
+		t.Errorf("acc-1 should be rate-limited")
+	}
+	if !r.Allow(ActionNewOrder, "acc-2", 100) {
+		t.Errorf("acc-2 should be unaffected by acc-1's bucket; expected allow")
+	}
+}
+
+func TestRateLimiter_PerActionIsolation(t *testing.T) {
+	// Same key but different actions get different buckets.
+	now := time.Date(2026, 5, 3, 12, 0, 0, 0, time.UTC)
+	r := NewRateLimiter()
+	r.SetClock(func() time.Time { return now })
+
+	for i := 0; i < 5; i++ {
+		_ = r.Allow(ActionKeyChange, "acc-1", 5)
+	}
+	if r.Allow(ActionKeyChange, "acc-1", 5) {
+		t.Errorf("ActionKeyChange should be rate-limited")
+	}
+	// ActionNewOrder for the same key has its own (empty) bucket.
+	if !r.Allow(ActionNewOrder, "acc-1", 100) {
+		t.Errorf("ActionNewOrder for same key should be allowed (different bucket)")
+	}
+}
+
+func TestRateLimiter_RefillOverTime(t *testing.T) {
+	// Drain bucket; advance the clock; expect tokens replenished.
+	current := time.Date(2026, 5, 3, 12, 0, 0, 0, time.UTC)
+	r := NewRateLimiter()
+	r.SetClock(func() time.Time { return current })
+
+	for i := 0; i < 100; i++ {
+		_ = r.Allow(ActionNewOrder, "acc-1", 100)
+	}
+	if r.Allow(ActionNewOrder, "acc-1", 100) {
+		t.Fatalf("expected limit hit after draining bucket")
+	}
+	// Advance by 36 seconds: at 100/hour = 100/3600 tokens/sec ≈
+	// 0.0278/sec. 36 * 0.0278 = 1.00 tokens — exactly enough for 1
+	// more call.
+	current = current.Add(36 * time.Second)
+	if !r.Allow(ActionNewOrder, "acc-1", 100) {
+		t.Errorf("Allow returned false after 36s elapsed; expected ≥1 token replenished")
+	}
+}
+
+func TestRateLimiter_RetryAfter(t *testing.T) {
+	now := time.Date(2026, 5, 3, 12, 0, 0, 0, time.UTC)
+	r := NewRateLimiter()
+	r.SetClock(func() time.Time { return now })
+
+	// Drain to zero.
+	for i := 0; i < 100; i++ {
+		_ = r.Allow(ActionNewOrder, "acc-1", 100)
+	}
+	d := r.RetryAfter(ActionNewOrder, "acc-1", 100)
+	// 1 token at 100/hour = 36 seconds.
+	if d < 35*time.Second || d > 37*time.Second {
+		t.Errorf("RetryAfter = %v, expected ~36s", d)
+	}
+	// Allow above capacity — RetryAfter returns 0 on a fresh bucket.
+	if zero := r.RetryAfter(ActionNewOrder, "acc-fresh", 100); zero != 0 {
+		t.Errorf("RetryAfter for fresh bucket = %v, expected 0", zero)
+	}
+}
+
+func TestRateLimiter_ConcurrentAccess(t *testing.T) {
+	// Hammer 200 goroutines × 200 calls each = 40000 calls against a
+	// 1000-token bucket; assert no panic, no data race (run with -race),
+	// and that no more than 1000 calls succeeded.
+	now := time.Date(2026, 5, 3, 12, 0, 0, 0, time.UTC)
+	r := NewRateLimiter()
+	r.SetClock(func() time.Time { return now })
+
+	var (
+		wg      sync.WaitGroup
+		success int64
+		mu      sync.Mutex
+	)
+	for g := 0; g < 200; g++ {
+		wg.Add(1)
+		go func() {
+			defer wg.Done()
+			local := int64(0)
+			for i := 0; i < 200; i++ {
+				if r.Allow(ActionNewOrder, "shared-acc", 1000) {
+					local++
+				}
+			}
+			mu.Lock()
+			success += local
+			mu.Unlock()
+		}()
+	}
+	wg.Wait()
+	if success > 1000 {
+		t.Errorf("got %d successes, want ≤ 1000 (bucket capacity)", success)
+	}
+	if success < 1000 {
+		t.Errorf("got %d successes, want exactly 1000 (frozen clock, no refill)", success)
+	}
+}
diff --git a/internal/api/handler/acme.go b/internal/api/handler/acme.go
index d8b7568..d578c8d 100644
--- a/internal/api/handler/acme.go
+++ b/internal/api/handler/acme.go
@@ -274,6 +274,23 @@ func writeServiceError(w http.ResponseWriter, err error) {
 		})
 	case errors.Is(err, service.ErrACMEARIBadCertID):
 		acme.WriteProblem(w, acme.Malformed("ARI cert-id is malformed"))
+	case errors.Is(err, service.ErrACMERateLimited):
+		// RFC 8555 §6.7 + RFC 7807. The handler doesn't have the
+		// (action, key) tuple here so we can't emit a precise
+		// Retry-After; the entry-point handlers (NewOrder etc.) emit
+		// their own Retry-After header before delegating to the
+		// service, leaving this catchall for completeness.
+		acme.WriteProblem(w, acme.Problem{
+			Type:   "urn:ietf:params:acme:error:rateLimited",
+			Detail: "request rate limit exceeded; retry later",
+			Status: http.StatusTooManyRequests,
+		})
+	case errors.Is(err, service.ErrACMEConcurrentOrdersExceeded):
+		acme.WriteProblem(w, acme.Problem{
+			Type:   "urn:ietf:params:acme:error:rateLimited",
+			Detail: "too many concurrent orders for this account; finish or cancel pending orders before submitting more",
+			Status: http.StatusTooManyRequests,
+		})
 	default:
 		// Avoid leaking internal error text per master-prompt
 		// criterion #10 (operator-actionable errors with no info
diff --git a/internal/config/config.go b/internal/config/config.go
index 9d12d01..869a1dc 100644
--- a/internal/config/config.go
+++ b/internal/config/config.go
@@ -744,6 +744,42 @@ type ACMEServerConfig struct {
 	// certs. Setting: CERTCTL_ACME_SERVER_ARI_POLL_INTERVAL.
 	ARIPollInterval time.Duration
 
+	// RateLimitOrdersPerHour caps new-order requests per ACME account per
+	// rolling hour. 0 disables (no limit). Default: 100. Hits return RFC
+	// 7807 + RFC 8555 §6.7 `urn:ietf:params:acme:error:rateLimited` with
+	// a Retry-After header. In-memory token-bucket — restart wipes the
+	// counter, which is acceptable for orders/hour caps (eventual-
+	// consistency anyway). Setting:
+	// CERTCTL_ACME_SERVER_RATE_LIMIT_ORDERS_PER_HOUR.
+	RateLimitOrdersPerHour int
+
+	// RateLimitConcurrentOrders caps the number of orders an ACME account
+	// can have in pending/ready/processing state simultaneously. 0
+	// disables. Default: 5. Same Problem shape as the per-hour limit.
+	// Setting: CERTCTL_ACME_SERVER_RATE_LIMIT_CONCURRENT_ORDERS.
+	RateLimitConcurrentOrders int
+
+	// RateLimitKeyChangePerHour caps account-key rollovers per account
+	// per rolling hour. 0 disables. Default: 5 (rollovers should be rare;
+	// a flood is an attack signal). Setting:
+	// CERTCTL_ACME_SERVER_RATE_LIMIT_KEY_CHANGE_PER_HOUR.
+	RateLimitKeyChangePerHour int
+
+	// RateLimitChallengeRespondsPerHour caps challenge-respond requests
+	// per challenge per rolling hour. 0 disables. Default: 60 (defends
+	// against retry storms from a misbehaving client). Setting:
+	// CERTCTL_ACME_SERVER_RATE_LIMIT_CHALLENGE_RESPONDS_PER_HOUR.
+	RateLimitChallengeRespondsPerHour int
+
+	// GCInterval is the tick interval for the ACME GC scheduler loop.
+	// On each tick the loop sweeps expired nonces, transitions expired
+	// pending authzs to `expired`, transitions expired
+	// pending/ready/processing orders to `invalid`, and reaps Phase-2
+	// atomicity-window orphans (orders without a linked cert when one
+	// should exist). 0 disables the loop entirely. Default: 1m. Setting:
+	// CERTCTL_ACME_SERVER_GC_INTERVAL.
+	GCInterval time.Duration
+
 	// DirectoryMeta is the optional metadata advertised in the directory
 	// document per RFC 8555 §7.1.1.
 	DirectoryMeta ACMEServerDirectoryMeta
@@ -1779,18 +1815,23 @@ func Load() (*Config, error) {
 		// NonceTTL + DirectoryMeta. Order/Authz TTLs + concurrency
 		// caps + DNS01 resolver are reserved (Phases 2/3 read).
 		ACMEServer: ACMEServerConfig{
-			Enabled:                 getEnvBool("CERTCTL_ACME_SERVER_ENABLED", false),
-			DefaultAuthMode:         getEnv("CERTCTL_ACME_SERVER_DEFAULT_AUTH_MODE", "trust_authenticated"),
-			DefaultProfileID:        getEnv("CERTCTL_ACME_SERVER_DEFAULT_PROFILE_ID", ""),
-			NonceTTL:                getEnvDuration("CERTCTL_ACME_SERVER_NONCE_TTL", 5*time.Minute),
-			OrderTTL:                getEnvDuration("CERTCTL_ACME_SERVER_ORDER_TTL", 24*time.Hour),
-			AuthzTTL:                getEnvDuration("CERTCTL_ACME_SERVER_AUTHZ_TTL", 24*time.Hour),
-			HTTP01ConcurrencyMax:    getEnvInt("CERTCTL_ACME_SERVER_HTTP01_CONCURRENCY", 10),
-			DNS01Resolver:           getEnv("CERTCTL_ACME_SERVER_DNS01_RESOLVER", "8.8.8.8:53"),
-			DNS01ConcurrencyMax:     getEnvInt("CERTCTL_ACME_SERVER_DNS01_CONCURRENCY", 10),
-			TLSALPN01ConcurrencyMax: getEnvInt("CERTCTL_ACME_SERVER_TLSALPN01_CONCURRENCY", 10),
-			ARIEnabled:              getEnvBool("CERTCTL_ACME_SERVER_ARI_ENABLED", true),
-			ARIPollInterval:         getEnvDuration("CERTCTL_ACME_SERVER_ARI_POLL_INTERVAL", 6*time.Hour),
+			Enabled:                           getEnvBool("CERTCTL_ACME_SERVER_ENABLED", false),
+			DefaultAuthMode:                   getEnv("CERTCTL_ACME_SERVER_DEFAULT_AUTH_MODE", "trust_authenticated"),
+			DefaultProfileID:                  getEnv("CERTCTL_ACME_SERVER_DEFAULT_PROFILE_ID", ""),
+			NonceTTL:                          getEnvDuration("CERTCTL_ACME_SERVER_NONCE_TTL", 5*time.Minute),
+			OrderTTL:                          getEnvDuration("CERTCTL_ACME_SERVER_ORDER_TTL", 24*time.Hour),
+			AuthzTTL:                          getEnvDuration("CERTCTL_ACME_SERVER_AUTHZ_TTL", 24*time.Hour),
+			HTTP01ConcurrencyMax:              getEnvInt("CERTCTL_ACME_SERVER_HTTP01_CONCURRENCY", 10),
+			DNS01Resolver:                     getEnv("CERTCTL_ACME_SERVER_DNS01_RESOLVER", "8.8.8.8:53"),
+			DNS01ConcurrencyMax:               getEnvInt("CERTCTL_ACME_SERVER_DNS01_CONCURRENCY", 10),
+			TLSALPN01ConcurrencyMax:           getEnvInt("CERTCTL_ACME_SERVER_TLSALPN01_CONCURRENCY", 10),
+			ARIEnabled:                        getEnvBool("CERTCTL_ACME_SERVER_ARI_ENABLED", true),
+			ARIPollInterval:                   getEnvDuration("CERTCTL_ACME_SERVER_ARI_POLL_INTERVAL", 6*time.Hour),
+			RateLimitOrdersPerHour:            getEnvInt("CERTCTL_ACME_SERVER_RATE_LIMIT_ORDERS_PER_HOUR", 100),
+			RateLimitConcurrentOrders:         getEnvInt("CERTCTL_ACME_SERVER_RATE_LIMIT_CONCURRENT_ORDERS", 5),
+			RateLimitKeyChangePerHour:         getEnvInt("CERTCTL_ACME_SERVER_RATE_LIMIT_KEY_CHANGE_PER_HOUR", 5),
+			RateLimitChallengeRespondsPerHour: getEnvInt("CERTCTL_ACME_SERVER_RATE_LIMIT_CHALLENGE_RESPONDS_PER_HOUR", 60),
+			GCInterval:                        getEnvDuration("CERTCTL_ACME_SERVER_GC_INTERVAL", time.Minute),
 			DirectoryMeta: ACMEServerDirectoryMeta{
 				TermsOfService:          getEnv("CERTCTL_ACME_SERVER_TOS_URL", ""),
 				Website:                 getEnv("CERTCTL_ACME_SERVER_WEBSITE", ""),
diff --git a/internal/repository/postgres/acme.go b/internal/repository/postgres/acme.go
index 78d1714..0c1b41c 100644
--- a/internal/repository/postgres/acme.go
+++ b/internal/repository/postgres/acme.go
@@ -751,6 +751,85 @@ func (r *ACMERepository) AccountOwnsCertificate(ctx context.Context, accountID,
 	return count > 0, nil
 }
 
+// --- Phase 5 — concurrent-orders count + GC sweeps ---------------------
+
+// CountActiveOrdersByAccount returns the number of acme_orders rows
+// with the given account_id where status is in
+// {pending, ready, processing}. Used by the per-account
+// concurrent-orders rate limit.
+func (r *ACMERepository) CountActiveOrdersByAccount(ctx context.Context, accountID string) (int, error) {
+	var count int
+	err := r.db.QueryRowContext(ctx, `
+		SELECT COUNT(1)
+		FROM acme_orders
+		WHERE account_id = $1
+		  AND status IN ('pending', 'ready', 'processing')
+	`, accountID).Scan(&count)
+	if err != nil {
+		return 0, fmt.Errorf("acme: count active orders: %w", err)
+	}
+	return count, nil
+}
+
+// GCExpiredNonces deletes nonce rows that have been used or have
+// passed their expires_at. Returns rows-affected count for telemetry.
+// Phase 5 — called every GCInterval from the scheduler.
+func (r *ACMERepository) GCExpiredNonces(ctx context.Context) (int64, error) {
+	res, err := r.db.ExecContext(ctx, `
+		DELETE FROM acme_nonces
+		WHERE used = TRUE OR expires_at < NOW()
+	`)
+	if err != nil {
+		return 0, fmt.Errorf("acme: gc expired nonces: %w", err)
+	}
+	n, err := res.RowsAffected()
+	if err != nil {
+		return 0, fmt.Errorf("acme: gc expired nonces rows affected: %w", err)
+	}
+	return n, nil
+}
+
+// GCExpireAuthorizations transitions authzs in `pending` whose
+// expires_at < NOW() to `expired`. Authzs in valid/invalid are left
+// alone (they're already terminal). Returns rows-affected count.
+func (r *ACMERepository) GCExpireAuthorizations(ctx context.Context) (int64, error) {
+	res, err := r.db.ExecContext(ctx, `
+		UPDATE acme_authorizations
+		SET status = 'expired', updated_at = NOW()
+		WHERE status = 'pending' AND expires_at < NOW()
+	`)
+	if err != nil {
+		return 0, fmt.Errorf("acme: gc expire authorizations: %w", err)
+	}
+	n, err := res.RowsAffected()
+	if err != nil {
+		return 0, fmt.Errorf("acme: gc expire authorizations rows affected: %w", err)
+	}
+	return n, nil
+}
+
+// GCInvalidateExpiredOrders transitions orders in
+// pending/ready/processing whose expires_at < NOW() to `invalid` with
+// a server-internal error. Orders in valid/invalid are terminal and
+// untouched.
+func (r *ACMERepository) GCInvalidateExpiredOrders(ctx context.Context) (int64, error) {
+	const errBlob = `{"type":"urn:ietf:params:acme:error:serverInternal","detail":"order expired before issuance","status":500}`
+	res, err := r.db.ExecContext(ctx, `
+		UPDATE acme_orders
+		SET status = 'invalid', error = $1::jsonb, updated_at = NOW()
+		WHERE status IN ('pending', 'ready', 'processing')
+		  AND expires_at < NOW()
+	`, errBlob)
+	if err != nil {
+		return 0, fmt.Errorf("acme: gc invalidate expired orders: %w", err)
+	}
+	n, err := res.RowsAffected()
+	if err != nil {
+		return 0, fmt.Errorf("acme: gc invalidate expired orders rows affected: %w", err)
+	}
+	return n, nil
+}
+
 // scanACMEAccount is the shared shape for the SELECT-by-X account
 // queries above. Returns sql.ErrNoRows-wrapped repository.ErrNotFound
 // on miss; any other scan failure surfaces verbatim.
diff --git a/internal/scheduler/scheduler.go b/internal/scheduler/scheduler.go
index 3db4191..e9dd94d 100644
--- a/internal/scheduler/scheduler.go
+++ b/internal/scheduler/scheduler.go
@@ -77,6 +77,13 @@ type CRLCacheServicer interface {
 	RegenerateAll(ctx context.Context)
 }
 
+// ACMEGarbageCollector is the interface the scheduler's acmeGCLoop
+// invokes once per tick. The concrete implementation is *service.ACMEService.
+// Phase 5 — sweeps expired nonces / authzs / orders.
+type ACMEGarbageCollector interface {
+	GarbageCollect(ctx context.Context) error
+}
+
 // JobReaperService defines the interface for job timeout reaping used by the scheduler.
 type JobReaperService interface {
 	ReapTimedOutJobs(ctx context.Context, csrTTL, approvalTTL time.Duration) error
@@ -101,6 +108,7 @@ type Scheduler struct {
 	healthCheckService    HealthCheckServicer
 	cloudDiscoveryService CloudDiscoveryServicer
 	crlCacheService       CRLCacheServicer
+	acmeGC                ACMEGarbageCollector
 	jobReaper             JobReaperService
 	logger                *slog.Logger
 
@@ -118,6 +126,7 @@ type Scheduler struct {
 	cloudDiscoveryInterval        time.Duration
 	crlGenerationInterval         time.Duration
 	jobTimeoutInterval            time.Duration
+	acmeGCInterval                time.Duration
 	// agentOfflineJobTTL: per-tick threshold for reaping Running jobs whose
 	// owning agent has been silent. Bundle C / Audit M-016. Defaults below.
 	agentOfflineJobTTL      time.Duration
@@ -138,6 +147,7 @@ type Scheduler struct {
 	cloudDiscoveryRunning        atomic.Bool
 	crlGenerationRunning         atomic.Bool
 	jobTimeoutRunning            atomic.Bool
+	acmeGCRunning                atomic.Bool
 
 	// Graceful shutdown: wait for in-flight work to complete
 	wg sync.WaitGroup
@@ -174,6 +184,7 @@ func NewScheduler(
 		cloudDiscoveryInterval:        6 * time.Hour,
 		crlGenerationInterval:         1 * time.Hour,
 		jobTimeoutInterval:            10 * time.Minute,
+		acmeGCInterval:                1 * time.Minute,
 		// 5 minutes is 5×agentHealthCheckInterval default of 1m; an agent
 		// must miss multiple heartbeats before its in-flight jobs are reaped.
 		agentOfflineJobTTL: 5 * time.Minute,
@@ -287,6 +298,25 @@ func (s *Scheduler) SetJobReaperService(jr JobReaperService) {
 	s.jobReaper = jr
 }
 
+// SetACMEGarbageCollector wires the ACME GC service. Phase 5 — when
+// non-nil, an acmeGCLoop runs every acmeGCInterval and sweeps expired
+// nonces / authzs / orders. Optional: leaving nil disables the loop
+// (legacy behavior pre-Phase-5).
+func (s *Scheduler) SetACMEGarbageCollector(gc ACMEGarbageCollector) {
+	s.acmeGC = gc
+}
+
+// SetACMEGCInterval configures the interval at which the ACME GC sweep
+// runs. Default 1m. Operators with quiet fleets can lengthen to 5m;
+// operators expecting nonce-storms can shorten to 30s. Zero or
+// negative values are ignored.
+func (s *Scheduler) SetACMEGCInterval(d time.Duration) {
+	if d <= 0 {
+		return
+	}
+	s.acmeGCInterval = d
+}
+
 // SetAgentOfflineJobTTL sets the threshold past which a Running job whose
 // owning agent has gone silent is reaped to Failed. Bundle C / Audit M-016.
 // Zero or negative values are ignored (the default of 5 minutes is kept).
@@ -342,6 +372,9 @@ func (s *Scheduler) Start(ctx context.Context) <-chan struct{} {
 		if s.crlCacheService != nil {
 			loopCount++
 		}
+		if s.acmeGC != nil {
+			loopCount++
+		}
 		s.wg.Add(loopCount)
 
 		go func() { defer s.wg.Done(); s.renewalCheckLoop(ctx) }()
@@ -367,6 +400,9 @@ func (s *Scheduler) Start(ctx context.Context) <-chan struct{} {
 		if s.crlCacheService != nil {
 			go func() { defer s.wg.Done(); s.crlGenerationLoop(ctx) }()
 		}
+		if s.acmeGC != nil {
+			go func() { defer s.wg.Done(); s.acmeGCLoop(ctx) }()
+		}
 
 		// Signal that all loops are launched
 		close(startedChan)
@@ -1074,3 +1110,39 @@ func (s *Scheduler) runCRLGeneration(ctx context.Context) {
 
 // ErrSchedulerShutdownTimeout is returned when scheduler graceful shutdown times out.
 var ErrSchedulerShutdownTimeout = errors.New("scheduler graceful shutdown timeout")
+
+// acmeGCLoop runs every acmeGCInterval and invokes ACMEGarbageCollector.
+// Per CLAUDE.md "Scheduler idempotency" architecture decision: an
+// atomic.Bool guard prevents concurrent tick execution; the
+// sync.WaitGroup tracks the in-flight goroutine for graceful shutdown.
+// Phase 5.
+func (s *Scheduler) acmeGCLoop(ctx context.Context) {
+	ticker := time.NewTicker(s.acmeGCInterval)
+	defer ticker.Stop()
+
+	for {
+		select {
+		case <-ctx.Done():
+			return
+		case <-ticker.C:
+			if !s.acmeGCRunning.CompareAndSwap(false, true) {
+				s.logger.Warn("ACME GC sweep still running, skipping tick")
+				continue
+			}
+			s.wg.Add(1)
+			go func() {
+				defer s.wg.Done()
+				defer s.acmeGCRunning.Store(false)
+				// 1-minute timeout per sweep — the per-statement work is
+				// cheap (single DELETE / UPDATE per sweep, all on indexed
+				// columns), but bound the cycle so a stuck Postgres can't
+				// block the next tick.
+				opCtx, cancel := context.WithTimeout(ctx, time.Minute)
+				defer cancel()
+				if err := s.acmeGC.GarbageCollect(opCtx); err != nil {
+					s.logger.Warn("acme gc sweep failed (next tick will retry)", "error", err)
+				}
+			}()
+		}
+	}
+}
diff --git a/internal/service/acme.go b/internal/service/acme.go
index e4bd6d4..83061d9 100644
--- a/internal/service/acme.go
+++ b/internal/service/acme.go
@@ -56,6 +56,20 @@ type ACMERepo interface {
 	// Phase 4 — key rollover + revocation auth.
 	UpdateAccountJWKWithTx(ctx context.Context, q repository.Querier, accountID, expectedOldThumbprint, newThumbprint, newJWKPEM string) error
 	AccountOwnsCertificate(ctx context.Context, accountID, certificateID string) (bool, error)
+	// Phase 5 — per-account concurrent-order count + GC sweeps.
+	// CountActiveOrdersByAccount returns the number of orders in
+	// pending/ready/processing for the given account.
+	CountActiveOrdersByAccount(ctx context.Context, accountID string) (int, error)
+	// GCExpiredNonces deletes nonces whose expires_at < now() OR
+	// used = true. Returns rows-affected count for telemetry.
+	GCExpiredNonces(ctx context.Context) (int64, error)
+	// GCExpireAuthorizations transitions authzs in `pending` whose
+	// expires_at < now() to `expired`. Returns rows-affected count.
+	GCExpireAuthorizations(ctx context.Context) (int64, error)
+	// GCInvalidateExpiredOrders transitions orders in
+	// pending/ready/processing whose expires_at < now() to `invalid`
+	// with a server-internal error. Returns rows-affected count.
+	GCInvalidateExpiredOrders(ctx context.Context) (int64, error)
 }
 
 // CertificateRevoker is the minimum surface ACMEService needs to route
@@ -142,6 +156,12 @@ type ACMEService struct {
 	// and RenewalInfo returns the no-policy default window.
 	revoker         CertificateRevoker
 	renewalPolicies RenewalPolicyLookup
+
+	// Phase 5 — per-account rate limiter. cmd/server/main.go constructs
+	// an *acme.RateLimiter and wires it via SetRateLimiter. When unset
+	// (tests, legacy bootstrap) the limiter calls short-circuit to
+	// "always allow" — same shape as the validatorPool unset case.
+	rateLimiter *acme.RateLimiter
 }
 
 // NewACMEService constructs an ACMEService with the directory + nonce
@@ -196,6 +216,16 @@ func (s *ACMEService) SetRevocationDelegate(r CertificateRevoker) { s.revoker =
 // still returns 200.
 func (s *ACMEService) SetRenewalPolicyLookup(r RenewalPolicyLookup) { s.renewalPolicies = r }
 
+// SetRateLimiter wires Phase 5's per-account rate limiter. Optional —
+// when nil, the per-action rate-limit checks short-circuit to
+// "always allow" so the legacy code path stays unchanged for bootstrap
+// + tests that don't care about throttling.
+func (s *ACMEService) SetRateLimiter(r *acme.RateLimiter) { s.rateLimiter = r }
+
+// RateLimiter returns the wired limiter so the handler can compute
+// Retry-After durations on rate-limited responses without re-checking.
+func (s *ACMEService) RateLimiter() *acme.RateLimiter { return s.rateLimiter }
+
 // SetValidatorPool wires Phase 3's challenge validator pool.
 // cmd/server/main.go constructs an *acme.Pool at startup with the
 // per-type concurrency caps from cfg.ACMEServer. Optional —
@@ -334,6 +364,19 @@ var ErrACMEARIDisabled = errors.New("acme: ARI is disabled on this server")
 // not RFC 9773 §4.1 shape. Handler maps to 400 + malformed.
 var ErrACMEARIBadCertID = errors.New("acme: ARI cert-id is malformed")
 
+// Phase 5 sentinels.
+
+// ErrACMERateLimited is returned when the per-action rate limit fires.
+// Handler maps to RFC 7807 + RFC 8555 §6.7
+// `urn:ietf:params:acme:error:rateLimited` with a Retry-After header.
+var ErrACMERateLimited = errors.New("acme: rate limit exceeded")
+
+// ErrACMEConcurrentOrdersExceeded is returned by CreateOrder when the
+// account already has cfg.RateLimitConcurrentOrders orders in
+// pending/ready/processing. Handler maps to rateLimited (RFC 8555 §6.7
+// shape; the certctl-side cause is concurrency rather than per-hour).
+var ErrACMEConcurrentOrdersExceeded = errors.New("acme: concurrent orders limit exceeded")
+
 // BuildDirectory constructs the per-profile directory document.
 //
 // profileID resolution:
@@ -463,6 +506,13 @@ type ACMEMetrics struct {
 	RevokeCertFailTotal  atomic.Uint64 // rejected revocation (4xx)
 	RenewalInfoTotal     atomic.Uint64 // ARI 200
 	RenewalInfoFailTotal atomic.Uint64 // ARI 4xx
+
+	// Phase 5 — GC sweep counts (per-tick rows-affected, summed).
+	GCNoncesReapedTotal      atomic.Uint64
+	GCAuthzsExpiredTotal     atomic.Uint64
+	GCOrdersInvalidatedTotal atomic.Uint64
+	GCRunsTotal              atomic.Uint64
+	GCRunFailuresTotal       atomic.Uint64
 }
 
 // NewACMEMetrics returns a zeroed counter table. Concurrent callers
@@ -508,6 +558,11 @@ func (m *ACMEMetrics) Snapshot() map[string]uint64 {
 		"certctl_acme_revoke_cert_failures_total":       m.RevokeCertFailTotal.Load(),
 		"certctl_acme_renewal_info_total":               m.RenewalInfoTotal.Load(),
 		"certctl_acme_renewal_info_failures_total":      m.RenewalInfoFailTotal.Load(),
+		"certctl_acme_gc_nonces_reaped_total":           m.GCNoncesReapedTotal.Load(),
+		"certctl_acme_gc_authzs_expired_total":          m.GCAuthzsExpiredTotal.Load(),
+		"certctl_acme_gc_orders_invalidated_total":      m.GCOrdersInvalidatedTotal.Load(),
+		"certctl_acme_gc_runs_total":                    m.GCRunsTotal.Load(),
+		"certctl_acme_gc_run_failures_total":            m.GCRunFailuresTotal.Load(),
 	}
 }
 
@@ -783,6 +838,27 @@ func (s *ACMEService) CreateOrder(
 		s.metrics.bump(&s.metrics.NewOrderFailureTotal)
 		return nil, fmt.Errorf("acme: new-order requires SetTransactor + SetAuditService")
 	}
+	// Phase 5 — per-account orders/hour cap. Hits return rateLimited
+	// (RFC 8555 §6.7) before any DB work. Counter is in-memory; restart
+	// wipes (eventual-consistency caps are acceptable).
+	if s.rateLimiter != nil && s.cfg.RateLimitOrdersPerHour > 0 {
+		if !s.rateLimiter.Allow(acme.ActionNewOrder, accountID, s.cfg.RateLimitOrdersPerHour) {
+			s.metrics.bump(&s.metrics.NewOrderFailureTotal)
+			return nil, ErrACMERateLimited
+		}
+	}
+	// Phase 5 — concurrent-orders cap. We count
+	// pending/ready/processing orders for this account; if at-or-over
+	// the cap, reject. This is a DB read (no FOR UPDATE), so two
+	// requests racing under the threshold can both succeed and push
+	// the account one over — accepted as eventual-consistency.
+	if s.cfg.RateLimitConcurrentOrders > 0 {
+		count, cerr := s.repo.CountActiveOrdersByAccount(ctx, accountID)
+		if cerr == nil && count >= s.cfg.RateLimitConcurrentOrders {
+			s.metrics.bump(&s.metrics.NewOrderFailureTotal)
+			return nil, ErrACMEConcurrentOrdersExceeded
+		}
+	}
 	resolvedProfileID, err := s.resolveProfile(ctx, profileID)
 	if err != nil {
 		s.metrics.bump(&s.metrics.NewOrderFailureTotal)
@@ -1285,6 +1361,16 @@ func (s *ACMEService) RespondToChallenge(
 		s.metrics.bump(&s.metrics.ChallengeRespondFailTotal)
 		return nil, ErrACMEChallengePoolUnconfigured
 	}
+	// Phase 5 — per-challenge respond rate limit. Defends against retry
+	// storms from a misbehaving client. Keyed by challengeID (not
+	// accountID) so a flood against one challenge doesn't drain the
+	// account's whole budget.
+	if s.rateLimiter != nil && s.cfg.RateLimitChallengeRespondsPerHour > 0 {
+		if !s.rateLimiter.Allow(acme.ActionChallengeRespond, challengeID, s.cfg.RateLimitChallengeRespondsPerHour) {
+			s.metrics.bump(&s.metrics.ChallengeRespondFailTotal)
+			return nil, ErrACMERateLimited
+		}
+	}
 
 	ch, err := s.repo.GetChallengeByID(ctx, challengeID)
 	if err != nil {
@@ -1508,6 +1594,14 @@ func (s *ACMEService) RotateAccountKey(
 		s.metrics.bump(&s.metrics.KeyChangeFailTotal)
 		return nil, ErrACMEKeyRolloverInvalid
 	}
+	// Phase 5 — rollovers/hour cap. Defaults to 5/hour: a flood is an
+	// attack signal (key rotation should be rare). Keyed by accountID.
+	if s.rateLimiter != nil && s.cfg.RateLimitKeyChangePerHour > 0 {
+		if !s.rateLimiter.Allow(acme.ActionKeyChange, oldAccount.AccountID, s.cfg.RateLimitKeyChangePerHour) {
+			s.metrics.bump(&s.metrics.KeyChangeFailTotal)
+			return nil, ErrACMERateLimited
+		}
+	}
 
 	newThumbprint, err := acme.JWKThumbprint(newJWK)
 	if err != nil {
@@ -1816,3 +1910,56 @@ func mapACMERevocationReason(code int) string {
 		return string(domain.RevocationReasonUnspecified)
 	}
 }
+
+// GarbageCollect runs a single ACME GC sweep. Phase 5 — the scheduler
+// invokes this every cfg.GCInterval. Three independent sweeps:
+//
+//  1. Delete used / expired nonces.
+//  2. Transition expired pending authzs to `expired`.
+//  3. Transition expired pending/ready/processing orders to `invalid`.
+//
+// Each sweep is a single SQL statement (no per-row transactions) so a
+// large reap is one atomic write per sweep. Per-sweep errors are
+// logged-and-continued: a failing nonces sweep doesn't block the
+// authzs sweep. Returns the first error encountered (for caller
+// telemetry); per-sweep counts are recorded on metrics regardless.
+//
+// Idempotent — repeated runs are safe; the second run finds 0 rows.
+func (s *ACMEService) GarbageCollect(ctx context.Context) error {
+	s.metrics.bump(&s.metrics.GCRunsTotal)
+	var firstErr error
+
+	if n, err := s.repo.GCExpiredNonces(ctx); err != nil {
+		s.metrics.bump(&s.metrics.GCRunFailuresTotal)
+		if firstErr == nil {
+			firstErr = fmt.Errorf("acme gc: nonces: %w", err)
+		}
+	} else if n > 0 {
+		atomicAddUint64(&s.metrics.GCNoncesReapedTotal, uint64(n))
+	}
+
+	if n, err := s.repo.GCExpireAuthorizations(ctx); err != nil {
+		s.metrics.bump(&s.metrics.GCRunFailuresTotal)
+		if firstErr == nil {
+			firstErr = fmt.Errorf("acme gc: authzs: %w", err)
+		}
+	} else if n > 0 {
+		atomicAddUint64(&s.metrics.GCAuthzsExpiredTotal, uint64(n))
+	}
+
+	if n, err := s.repo.GCInvalidateExpiredOrders(ctx); err != nil {
+		s.metrics.bump(&s.metrics.GCRunFailuresTotal)
+		if firstErr == nil {
+			firstErr = fmt.Errorf("acme gc: orders: %w", err)
+		}
+	} else if n > 0 {
+		atomicAddUint64(&s.metrics.GCOrdersInvalidatedTotal, uint64(n))
+	}
+
+	return firstErr
+}
+
+// atomicAddUint64 adds delta to the counter. The metrics struct exposes
+// only `bump` (add 1) by default; this helper covers the
+// rows-affected-N case the GC needs.
+func atomicAddUint64(c *atomic.Uint64, delta uint64) { c.Add(delta) }
diff --git a/internal/service/acme_test.go b/internal/service/acme_test.go
index 8aa0098..cc9787b 100644
--- a/internal/service/acme_test.go
+++ b/internal/service/acme_test.go
@@ -164,6 +164,23 @@ func (f *fakeACMERepo) UpdateAccountJWKWithTx(ctx context.Context, q repository.
 func (f *fakeACMERepo) AccountOwnsCertificate(ctx context.Context, accountID, certificateID string) (bool, error) {
 	return false, nil
 }
+func (f *fakeACMERepo) CountActiveOrdersByAccount(ctx context.Context, accountID string) (int, error) {
+	return 0, nil
+}
+func (f *fakeACMERepo) GCExpiredNonces(ctx context.Context) (int64, error) {
+	n := int64(0)
+	for nonce, exp := range f.issued {
+		if time.Now().After(exp) {
+			delete(f.issued, nonce)
+			n++
+		}
+	}
+	return n, nil
+}
+func (f *fakeACMERepo) GCExpireAuthorizations(ctx context.Context) (int64, error) { return 0, nil }
+func (f *fakeACMERepo) GCInvalidateExpiredOrders(ctx context.Context) (int64, error) {
+	return 0, nil
+}
 
 // fakeTransactor is the repository.Transactor stand-in: runs fn
 // against the supplied querier (we just pass nil — fakes ignore it).