diff --git a/cowork/ci-pipeline-cleanup/baseline.md b/cowork/ci-pipeline-cleanup/baseline.md
new file mode 100644
index 0000000..163d571
--- /dev/null
+++ b/cowork/ci-pipeline-cleanup/baseline.md
@@ -0,0 +1,159 @@
+# CI Pipeline Cleanup — Phase 0 Baseline
+
+> Captured against repo HEAD `1de61e91cf07449356d9046a76499c86efe413b1` (operator tag `v2.0.66`) on 2026-04-30.
+> Each subsequent Phase that changes a number references this baseline.
+
+## Repo state
+
+**HEAD SHA:** `1de61e91cf07449356d9046a76499c86efe413b1`
+
+**Operator-stamped tag:** `v2.0.66`
+
+## ci.yml shape
+
+- Total lines: `1488`
+- Total named steps: `53`
+- Named regression-guard steps: 22 (enumerated below)
+
+### The 22 regression-guard steps
+
+```
+81:      - name: Forbidden auth-type literal regression guard (G-1)
+144:      - name: Forbidden bare InsecureSkipVerify regression guard (L-001)
+180:      - name: Forbidden bare FROM regression guard (H-001)
+201:      - name: Forbidden missing USER regression guard (M-012)
+228:      - name: Forbidden README JWT advertising regression guard (H-009)
+254:      - name: Forbidden api_key_hash JSON-shape regression guard (G-2)
+311:      - name: Forbidden plaintext HEALTHCHECK regression guard (U-2)
+360:      - name: Forbidden migration mount in compose initdb (U-3)
+417:      - name: Forbidden StatusBadge dead-key + TS phantom-field regression guard (D-1 + D-2)
+569:      - name: Forbidden client-side bulk-action loop regression guard (L-1)
+613:      - name: Forbidden orphan-CRUD client function regression guard (B-1)
+665:      - name: Forbidden strings.Contains(err.Error()) regression guard (S-2)
+868:      - name: QA-doc Part-count drift guard
+886:      - name: QA-doc seed-count drift guard
+938:      - name: Test-naming convention guard (hard-fail)
+982:      - name: Forbidden hardcoded source-count prose regression guard (S-1)
+1027:      - name: Documented orphan client fns sync guard (P-1)
+1063:      - name: Frontend page-coverage regression guard (T-1)
+1118:      - name: Bundle-8 / L-015 target=_blank rel=noopener regression guard
+1147:      - name: Bundle-8 / L-019 dangerouslySetInnerHTML regression guard
+1176:      - name: Bundle-8 / M-009 + M-029 Pass 1 mutation contract guard (hard zero)
+1220:      - name: Forbidden env-var docs drift regression guard (G-3)
+```
+
+## SA1019 site count
+
+- **Operator-on-workstation deliverable** — sandbox cannot run `staticcheck`.
+- ci.yml inline comment claims "6 sites" (`middleware.NewAuth × 3`, `csr.Attributes`, `elliptic.Marshal`).
+- Source-grep at HEAD shows:
+  - `internal/api/handler/scep.go`: `csr.Attributes` references present
+  - `internal/connector/issuer/local/local.go`: `elliptic.Marshal` historic refs (already migrated per bundle9_coverage_test.go byte-equivalence test)
+  - `cmd/server/main_test.go`: `middleware.NewAuth` references TBD
+- Operator must run `staticcheck ./... 2>&1 | grep SA1019` on workstation and update Phase 3 plan with the actual site list.
+
+## Dockerfile inventory (verified 4)
+
+```
+./Dockerfile.agent
+./Dockerfile
+./deploy/test/f5-mock-icontrol/Dockerfile
+./deploy/test/libest/Dockerfile
+```
+
+## Migration up/down balance
+
+- ups: `24`
+- downs: `24`
+- missing downs: `0`
+
+## OpenAPI ↔ handler parity gap (verified)
+
+- operationIds in api/openapi.yaml: `136`
+- r.Register calls in router.go: `149`
+- Gap to root-cause in Phase 9: 13 routes
+
+## docker-compose.test.yml sidecars
+
+```
+52:  certctl-tls-init:
+107:  postgres:
+135:  pebble-challtestsrv:
+150:  pebble:
+178:  step-ca:
+213:  certctl-server:
+363:  nginx:
+391:  certctl-agent:
+449:  libest-client:
+488:  apache-test:
+502:  haproxy-test:
+515:  traefik-test:
+533:  caddy-test:
+548:  envoy-test:
+562:  postfix-test:
+577:  dovecot-test:
+591:  openssh-test:
+613:  f5-mock-icontrol:
+631:  k8s-kind-test:
+648:  windows-iis-test:
+666:  certctl-test:
+```
+
+## Makefile::verify body (existing)
+
+```
+verify:
+	@echo "==> fmt"
+	@go fmt ./... | { ! grep -q '.'; } || (echo "gofmt produced changes — commit them" && exit 1)
+	@echo "==> go vet ./..."
+	@go vet ./...
+	@echo "==> golangci-lint run ./... (incl. staticcheck ST*)"
+	@which golangci-lint > /dev/null || (echo "Installing golangci-lint..." && go install github.com/golangci/golangci-lint/cmd/golangci-lint@latest)
+	@golangci-lint run ./... --timeout 5m
+	@echo "==> go test -short ./..."
+	@go test -short -count=1 ./...
+	@echo ""
+	@echo "verify: PASS — safe to commit"
+
+```
+
+## RAM headroom for collapsed vendor-e2e job
+
+- **Operator-on-workstation deliverable** — requires a prototype branch with the collapsed job + `docker stats` polling.
+- Per Phase 0 frozen decision 0.14: if peak RSS ≤ 12 GB on ubuntu-latest (16 GB ceiling), single-job collapse is approved.
+- If > 12 GB, fall back to bucketed-matrix design documented in `cowork/ci-pipeline-cleanup/decisions-revised.md`.
+
+## Coverage thresholds at HEAD
+
+```
+778:          if [ "$(echo "$SERVICE_COV < 70" | bc -l)" -eq 1 ]; then
+779:            echo "::error::Service layer coverage ${SERVICE_COV}% is below 70% (Bundle R-CI-extended floor — add tests, do not lower the gate)"
+782:          if [ "$(echo "$HANDLER_COV < 75" | bc -l)" -eq 1 ]; then
+783:            echo "::error::Handler layer coverage ${HANDLER_COV}% is below 75% (Bundle R-CI-extended floor — add tests, do not lower the gate)"
+786:          if [ "$(echo "$DOMAIN_COV < 40" | bc -l)" -eq 1 ]; then
+787:            echo "::error::Domain layer coverage ${DOMAIN_COV}% is below 40% threshold"
+790:          if [ "$(echo "$MIDDLEWARE_COV < 30" | bc -l)" -eq 1 ]; then
+791:            echo "::error::Middleware layer coverage ${MIDDLEWARE_COV}% is below 30% threshold"
+802:          if [ "$(echo "$CRYPTO_COV < 88" | bc -l)" -eq 1 ]; then
+803:            echo "::error::Crypto package coverage ${CRYPTO_COV}% is below 88% (Bundle R closure floor — add tests, do not lower the gate)"
+832:          if [ "$(echo "$LOCAL_ISSUER_COV < 86" | bc -l)" -eq 1 ]; then
+833:            echo "::error::Local-issuer coverage ${LOCAL_ISSUER_COV}% is below 86% (Bundle R closure floor — add tests, do not lower the gate)"
+842:          if [ "$(echo "$ACME_COV < 80" | bc -l)" -eq 1 ]; then
+843:            echo "::error::ACME issuer coverage ${ACME_COV}% is below 80% (Bundle R-CI-extended floor — add tests, do not lower the gate)"
+846:          if [ "$(echo "$STEPCA_COV < 80" | bc -l)" -eq 1 ]; then
+847:            echo "::error::StepCA issuer coverage ${STEPCA_COV}% is below 80% (Bundle L.B closure floor — add tests, do not lower the gate)"
+850:          if [ "$(echo "$MCP_COV < 85" | bc -l)" -eq 1 ]; then
+851:            echo "::error::MCP coverage ${MCP_COV}% is below 85% (Bundle K closure floor — add tests, do not lower the gate)"
+```
+
+## CodeQL workflow (no changes)
+
+- File: `.github/workflows/codeql.yml` (`81` lines)
+- Matrix: `[go, javascript-typescript]` — 2 status checks per push
+- Trigger: push to master, PR to master, weekly Sunday cron
+
+## Status check accounting (verified)
+
+Today: 1 `go-build-and-test` + 1 `frontend-build` + 1 `helm-lint` + 12 `deploy-vendor-e2e (<vendor>)` + 2 `deploy-vendor-e2e-windows (<vendor>)` + 2 `CodeQL Analyze (<lang>)` = **19 status checks per push**.
+
+After cleanup: 1 `go-build-and-test` + 1 `frontend-build` + 1 `helm-lint` + 1 `deploy-vendor-e2e` + 1 `image-and-supply-chain` + 2 `CodeQL Analyze (<lang>)` = **7 status checks per push**.
diff --git a/cowork/ci-pipeline-cleanup/decisions-revised.md b/cowork/ci-pipeline-cleanup/decisions-revised.md
new file mode 100644
index 0000000..ac1ee95
--- /dev/null
+++ b/cowork/ci-pipeline-cleanup/decisions-revised.md
@@ -0,0 +1,53 @@
+# CI Pipeline Cleanup — Deliberate Revisions of Bundle II Decisions
+
+This bundle deliberately revises two Bundle II frozen decisions. Both revisions are recorded here for audit trail and acknowledged in the per-Phase commits that implement them.
+
+## Bundle II decision 0.4 → revised by ci-pipeline-cleanup decision 0.5
+
+**Bundle II 0.4 (original):** "IIS e2e strategy — `mcr.microsoft.com/windows/servercore:ltsc2022` Windows containers via Docker Desktop on Windows hosts. Linux CI runners CAN'T run Windows containers, so the IIS e2e suite runs on a separate Windows-runner CI matrix job (or operator's local Windows host for development). Documented limitation."
+
+**ci-pipeline-cleanup 0.5 (revision):** Delete the Windows-runner CI matrix entirely.
+
+**Rationale for revision:**
+
+1. The matrix can't physically work on `windows-latest` GitHub-hosted runners today. Verified via the failure logs from CI run `25183374742` (commit `1de61e9`):
+   - `wincertstore` job: `error during connect: ... open //./pipe/docker_engine: The system cannot find the file specified` — Docker daemon not started in Windows-containers mode.
+   - `iis` job: image pulled successfully (so the new digest is correct), then died at `failed to create network deploy_certctl-test: could not find plugin bridge in v1 plugin registry: plugin not found` — `bridge` network driver doesn't exist on Windows Docker (uses `nat`).
+
+2. Even if both Docker-daemon and network-driver issues were fixed, the matrix would validate nothing of substance. Verified by source-grep: all 16 functions matching `TestVendorEdge_(IIS|WinCertStore)_*` in `deploy/test/vendor_e2e_phase3_to_13_test.go` are `t.Log` placeholders that exercise no IIS-specific behavior. The real IIS connector validation lives in `internal/connector/target/iis/` unit tests (run on Linux in `go-build-and-test` — already green per push).
+
+3. Bundle II decision 0.14 explicitly required operator manual smoke against a real instance for "verified" status in the vendor matrix. Moving IIS + WinCertStore validation to a documented operator playbook in `docs/connector-iis.md` satisfies that criterion better than a fake CI matrix that passes by skipping.
+
+**Preservation:** the `windows-iis-test` sidecar stays in `deploy/docker-compose.test.yml` under `profiles: [deploy-e2e-windows]` — operators on a Windows host can opt in via `docker compose --profile deploy-e2e-windows up -d windows-iis-test`. Linux CI never activates this profile.
+
+## Bundle II decision 0.9 → revised by ci-pipeline-cleanup decision 0.4
+
+**Bundle II 0.9 (original):** "CI parallelism — Each vendor e2e gets its own GitHub Actions matrix job. Vendor failures surface independently in the CI status check (operator sees 'K8s 1.31 vendor-edge fail' as a discrete check, not a generic 'integration tests failed')."
+
+**ci-pipeline-cleanup 0.4 (revision):** Single `deploy-vendor-e2e` job replaces the 12-job matrix; per-vendor visibility partially restored via skip-detection guard messages.
+
+**Rationale for revision:**
+
+1. The per-vendor granularity Bundle II decision 0.9 was designed to provide is fake signal. Verified by source-analysis at HEAD:
+   ```
+   $ grep -cE 't\.Log\(' deploy/test/{vendor_e2e_phase3_to_13,nginx_vendor_e2e}_test.go
+   deploy/test/nginx_vendor_e2e_test.go:9
+   deploy/test/vendor_e2e_phase3_to_13_test.go:106
+
+   $ awk '/^func TestVendorEdge_/{in_test=1; name=$2; has_assert=0; next}
+          in_test && /^}$/ {if (has_assert) print name; in_test=0}
+          in_test && /t\.(Fatal|Error|Errorf|Fatalf|Fail|Failf)/ {has_assert=1}' \
+          deploy/test/vendor_e2e_phase3_to_13_test.go deploy/test/nginx_vendor_e2e_test.go
+   TestVendorEdge_NGINX_HighConcurrencyDeployUnderLoad_E2E
+   ```
+   115 of 116 vendor-edge test functions are `t.Log`-only — they spin up a sidecar, log a one-line description of the vendor quirk, and return. Only 1 has a real assertion.
+
+2. Per-vendor status-check granularity costs ~9 sec setup overhead × 12 jobs = ~108 sec of pure runner waste per push (verified from CI run `25183374742` job timings).
+
+3. The single-job version partially restores per-vendor visibility via the skip-detection guard (decision 0.6): if a sidecar fails to start, the affected tests' SKIP names print in the CI output and the build fails. Operators see "TestVendorEdge_K8s_KubeletSyncWaitContract_DefaultTimeout60s_E2E SKIPPED: vendor sidecar 'k8s-kind' not reachable" — same per-vendor signal, just no longer rendered as a separate status-check row.
+
+**Preservation:** the per-test discoverability via `go test -run 'VendorEdge_<vendor>'` (Bundle II frozen decision 0.6) is unchanged. Only the matrix-jobs-per-vendor part of decision 0.9 is revised; the per-test naming convention stays.
+
+## Forward-looking note
+
+Both revisions are limited in scope to CI execution shape — they do NOT delete the test files, the sidecar definitions, or the documentation that Bundle II shipped. Future work could re-introduce per-vendor matrix jobs if test bodies are filled in with real assertions (transforming the t.Log placeholders into actual contract pins). At that point, decision 0.4 + 0.9 should be re-evaluated.
diff --git a/cowork/ci-pipeline-cleanup/frozen-decisions.md b/cowork/ci-pipeline-cleanup/frozen-decisions.md
new file mode 100644
index 0000000..7bcc0a9
--- /dev/null
+++ b/cowork/ci-pipeline-cleanup/frozen-decisions.md
@@ -0,0 +1,64 @@
+# CI Pipeline Cleanup — Frozen Decisions
+
+> 14 frozen decisions confirmed at Phase 0. Each subsequent Phase references the decision number it implements.
+
+## 0.1 — Trigger model
+
+Three-tier split, no mixing:
+- **On push/PR to master:** blocking, fast, every check earns its keep, target <10 min wall-clock.
+- **Daily cron + workflow_dispatch:** `security-deep-scan.yml` as-is; slow scans, best-effort, never blocks.
+- **On tag push (`v*`):** `release.yml` as-is; cross-platform binaries, ghcr.io push, SLSA provenance.
+
+## 0.2 — Extracted-script location
+
+`scripts/ci-guards/` at repo root. Operator runs `bash scripts/ci-guards/<id>.sh` locally. Contract documented in `scripts/ci-guards/README.md`.
+
+## 0.3 — Coverage threshold YAML format
+
+`.github/coverage-thresholds.yml`. Top-level keys are package paths; each entry has `floor:` (integer pct) + `why:` (multi-line string for load-bearing context). Bash step uses Python (already on the runner) to read the YAML — no `yq` dependency.
+
+## 0.4 — Vendor matrix collapse policy (REVISES Bundle II decision 0.9)
+
+Single `deploy-vendor-e2e` job replaces 12-job matrix. Bundle II decision 0.9 said "Each vendor e2e gets its own GitHub Actions matrix job" — this revision recognizes that 115/116 vendor-edge tests are `t.Log` placeholders, so per-vendor status-check granularity is fake signal. Skip-detection guard partially restores per-vendor visibility (SKIP messages name the vendor). Documented as deliberate revision in `cowork/ci-pipeline-cleanup/decisions-revised.md`.
+
+## 0.5 — Windows IIS validation deletion (REVISES Bundle II decision 0.4)
+
+Delete `deploy-vendor-e2e-windows` matrix entirely. Bundle II decision 0.4 said "the IIS e2e suite runs on a separate Windows-runner CI matrix job" — this revision recognizes that (a) the matrix can't physically work on `windows-latest` (Docker not started in Windows-containers mode; `bridge` driver missing on Windows Docker), and (b) all 16 IIS + WinCertStore tests are `t.Log` placeholders. Move validation to `docs/connector-iis.md::Operator validation playbook` per Bundle II decision 0.14's third criterion. The `windows-iis-test` sidecar stays in `deploy/docker-compose.test.yml` for operator local use.
+
+## 0.6 — Skip-detection guard semantics + EXPECTED_SKIPS allowlist
+
+After `go test -tags integration -run 'VendorEdge_'`, count `^--- SKIP:` lines. Allowlist: 6 JavaKeystore tests in `vendor_e2e_phase3_to_13_test.go` that legitimately t.Log without sidecar. Allowlist file at `scripts/ci-guards/vendor-e2e-skip-allowlist.txt`, one test name per line.
+
+## 0.7 — SA1019 closure approach
+
+Close each site individually with byte-equivalence tests where the deprecated API was load-bearing. Then flip `continue-on-error: true` → `false` in the SAME commit. Do NOT split — shipping the gate without closing sites would fail CI on master. Live verification: `staticcheck ./... 2>&1 | grep -c SA1019` returns 0 BEFORE flipping the gate.
+
+## 0.8 — Image-and-supply-chain placement
+
+Separate top-level job (not steps in `go-build-and-test`). Two reasons: (a) digest-validity needs network egress to multiple registries (Docker Hub, ghcr.io, mcr.microsoft.com), bundling into go-build blocks Go tests on registry latency. (b) `docker build` is parallel to Go tests; isolating lets it run concurrently.
+
+## 0.9 — Coverage PR-comment provider
+
+Default: lightweight self-hosted action that posts a per-PR comment via `gh pr comment`. Avoids paid SaaS. Operator can swap to Codecov/Coveralls later.
+
+## 0.10 — Docker build smoke scope
+
+Build all 4 Dockerfiles in the repo: `Dockerfile`, `Dockerfile.agent`, `deploy/test/f5-mock-icontrol/Dockerfile`, `deploy/test/libest/Dockerfile`. The test-sidecar Dockerfiles are load-bearing for vendor-e2e — a syntax error there silently breaks the e2e suite. Tagged `:smoke` and discarded.
+
+## 0.11 — OpenAPI ↔ handler parity exception YAML
+
+NEW `api/openapi-handler-exceptions.yaml`. Schema: `documented_exceptions:` list of `{route, why}` entries. The 13-route gap at HEAD is root-caused in Phase 9; most are likely health probes / metrics / SCEP-EST-OCSP wire endpoints that legitimately have no operationId.
+
+## 0.12 — Branch-protection-rule update timing
+
+Operator updates GitHub branch-protection rules in Phase 13 AFTER the new pipeline ships and runs green on a feature branch + on the first push to master. Required-checks list changes from 19 → 7 entries. Operator action only — agent cannot do this.
+
+## 0.13 — Make-target naming for new operator-side scripts
+
+- `make verify` (existing) — required pre-commit; gofmt + vet + lint + tests
+- `make verify-deploy` (new) — optional pre-push; digest-validity + OpenAPI parity + docker build smoke (server + agent only — fast subset for local)
+- `make verify-docs` (new) — required pre-tag; QA-doc Part-count + seed-count drift
+
+## 0.14 — RAM headroom verification methodology
+
+Phase 0 deliverable. Operator creates `prototype/ci-pipeline-cleanup-vendor-collapse` branch, runs the collapsed `deploy-vendor-e2e` job once, captures peak RSS via `docker stats --no-stream` snapshots every 30 sec, records max in this baseline doc. If max > 12 GB (75% of 16 GB ceiling), fall back to bucketed matrix (3 jobs × ~4 sidecars). If max ≤ 12 GB, single-job collapse is approved.