mjolnir-qa 1.0.5 → 1.0.8

package/CHANGELOG.md

@@ -9,6 +9,468 @@ Rule behavior changes (new rules, FP-rate changes against the corpus,
 severity changes) are first-class entries here — rule IDs are immutable
 once shipped, so this file is the record of what changed between versions.
 
+## [1.0.8] — 2026-09-11
+
+### Changes since 1.0.7
+
+- chore: sync smithery.yaml in the release cut step (Merged PR #76)
+
+## [1.0.7] — 2026-09-11
+
+### Changes since 1.0.6
+
+- docs: 1.0.6 CHANGELOG section lead-ins (Merged PR #75)
+
+## [1.0.6] — 2026-09-11
+
+### R10 2.0 preparation: breaking-set inventory + boundary-law guards (remediation/remote-first WI-25)
+
+Preparation-only increment: the 2.0 breaking-set proposal sheet, its migration draft, and the boundary-law guard tests — nothing breaking ships in this release.
+
+### Added
+
+- **2.0 breaking-set inventory** (`docs/2.0-BREAKING-SET.md`, WI-25): the
+  proposal sheet per the strategic blueprint's §28/§18 — two justified
+  breaking candidates (BS-1 default suppression expiry with the explicit
+  never-expire opt-out; BS-2 retirement completion into `RETIRED_RULE_IDS`),
+  each carrying its benefit>cost justification, its migration pointer, and a
+  PROPOSED decision line awaiting owner ratification, plus the locked
+  NOT-breaking list (`schemaVersion 1` additive extension, exit codes,
+  additive verbs, Node matrix, frozen surfaces). **Nothing is implemented in
+  this release** — nothing enters 2.0 "because large", and no frozen surface
+  breaks without evidence that `schemaVersion 1` cannot represent the
+  behavior.
+- **Migration guide draft** (`docs/MIGRATION-2.0-DRAFT.md`): the working
+  draft of the 2.0.0 guide (publication law: CHANGELOG + site with the
+  release itself) covering BS-1 (init config check → explicit `expires` /
+  `expires: false`; no silent retroactive expiry) and BS-2 (retired-rule
+  list, suppression cleanup, §15-lifecycle-honest disappearance causes).
+- **Boundary-law + non-goal guards** (`tests/contract/boundary-law.spec.ts`,
+  blueprint §9.1/§24/§36): the canonical layers (engine, forensics,
+  adapters, rules) never import upward into commands/ or the transports;
+  the MCP transport imports no detection machinery and rides the canonical
+  machine contract (pipeline → contract → runtime evidence → agent
+  transport is never reversed); the zero-network contract holds; the
+  dependency list carries no telemetry/cloud/hosted-backend package; src/
+  reads no telemetry configuration; and the breaking-set discipline is
+  drift-locked (every entry carries a Decision, nothing is implemented).
+
+### Changed
+
+- **Final capability matrix update** (`src/capabilities.ts`,
+  `docs/PLAYWRIGHT-CAPABILITIES.md`): the `to agents` column now carries the
+  foundational Agent Skill's evidence pointer (`src/commands/install-agents.ts`,
+  shipped with R8/WI-22) on every row — a `no → yes` flip in the same change
+  set that shipped its evidence, per the claim law.
+
+### Fixed
+
+- **MCP stdio bundle no longer prints the terminal Trust Report onto the
+  JSON-RPC stream** (`src/mcp/server.ts`): the standalone entry
+  (`node dist/mcp/stdio.mjs`, `npm run mcp`) dragged the CLI module in via
+  `import { runScan, CLI_VERSION } from "../cli.js"`, and cli.ts's entry
+  tail fired inside the bundle (`import.meta.url === argv[1]`), emitting the
+  full terminal Trust Report before/between JSON-RPC frames — a fatal
+  protocol violation for any MCP client. The transport now imports the
+  canonical homes directly (`engine/scan-pipeline.js`, `engine/version.js`),
+  the bundle contains no CLI entry tail, and the boundary-law guard bans the
+  `../cli.js` import from the MCP layer permanently. Found by the R10
+  bug-hunt smoke against the real stdio transport.
+- **Stale dist can no longer mask new code in spawned-binary tests**
+  (`tests/e2e/global-setup.ts`): the EXISTS-ONLY guard skipped the build
+  whenever a bundle was present, so the spawned stdio binary kept answering
+  from a pre-R8 catalog ("unknown tool: triage") while the suite stayed
+  green. The setup now rebuilds whenever any `src/**/*.ts` is newer than the
+  bundle (the same freshness discipline as the generated-docs drift gates).
+- **Freshness diagnosis names each drift class** (`checkArtifactFreshness`):
+  a revision bump (`rule@old -> new`), a rule retired since the render
+  (`rule@rev (retired)`), and a rule added since the render
+  (`rule@rev (new)`) are distinct remediations — a flat list hid which one
+  happened.
+- **`trust-report --from` reports the complete artifact set** it writes
+  (md + html + json), not just the MD path.
+- **The `provenance = bound` system-invariant item is now WIRED** (plan
+  §5.2 activation; `src/commands/release-trust.ts`): the release-trust
+  invariant previously hardcoded `provenance: UNSUPPORTED` even after the
+  machinery it waited for shipped. It is now PROVEN exactly when the
+  machine-anchored identity chain is proven — scope-integrity (runIdentity +
+  evidence graph, R4c) AND artifact-integrity (artifact scanId binding, R9)
+  both PASS — and stays UNSUPPORTED (recorded, non-blocking) otherwise. The
+  contract doc's activation sentence and the drift-lock are updated
+  accordingly; the shipped verdict is unchanged (PASS 12/12, provenance
+  bound).
+
+### R9 Trust Artifact integrity + HTML completion (remediation/remote-first WI-23+24)
+
+Trust Artifacts gain machine-anchored identity and a deterministic HTML surface; stale, wrong-run, revision-drifted, and unbound artifacts are now detectable.
+
+### Added
+
+- **Artifact integrity binding** (`src/commands/trust-report.ts`, R9): every
+  Trust Artifact (md · json · html) now embeds its IDENTITY — the machine
+  anchor (`scanId` from runIdentity), the bound commit when resolvable
+  (offline git read; null, never fabricated), the fired rule(rev) inventory
+  (deduped, sorted, undeclared revisions omitted — never a fabricated rev),
+  and the evidence inventory (totals, runtime-corroborated count, per-level
+  counts). Consumers detect **stale artifacts** (a scanId from another run),
+  **mismatched revisions** (rule-set drift, named per rule), and **unbound
+  artifacts** (pre-R9 producers) via `checkArtifactFreshness` — an unbound or
+  stale artifact is RECORDED, never assumed current.
+- **HTML Trust Artifact** (WI-23 completion, §18): deterministic,
+  self-contained `mjolnir-trust-report.html` — inline CSS only, zero external
+  resources, hostile interpolations escaped, byte-identical regen (same
+  ScanResult + label + commit → same bytes), the same five-question structure
+  as the MD. The command writes all three formats; `--from` gains an optional
+  `--commit <sha>` so the Action binds the artifact to the executing run's
+  HEAD.
+- **Artifact Integrity dimension wired** (`check:artifact-integrity` in
+  `src/commands/release-trust.ts`, R9 surface): the structural evaluation
+  asserts the identity binding, the three-format output, the freshness
+  detection, and the byte-regen/hostile-safety contract locks. The
+  release-trust contract's documented-unwired list is now EMPTY — all 12
+  canonical dimensions are wired and machine-evaluated.
+
+### R8 MCP runtime-evidence tools + Agent Safety (remediation/remote-first WI-21+22)
+
+The MCP transport learns the runtime-evidence tools, and every installed agent surface inherits the safety contract.
+
+### Added
+
+- **MCP runtime-evidence tools** (`src/mcp/server.ts`, WI-21): `forensics`,
+  `triage`, `pw-report` join the tool catalog as 1:1 mappings onto the SAME
+  engine functions the CLI verbs call — no MCP-only semantics. Parity is
+  drift-locked table-driven (`tests/mcp/parity.spec.ts`): for every new tool ×
+  every fixture class (Playwright JSON · JUnit XML · hostile corrupt report ·
+  no-reports directory) the MCP result deep-equals the canonical CLI
+  derivation, hostile inputs degrade to zero records on BOTH surfaces, and the
+  hostile parameter matrix (missing / empty / non-string / nonexistent path)
+  yields INVALID_PARAMS naming the target, never a crash. A crashing tool
+  never kills the server (`tests/mcp/crash-containment.spec.ts`): the failure
+  lands in the transport's existing catch as a structured INTERNAL error and
+  the server keeps answering. One scan in flight; zero network; the plugin
+  gate applies unchanged.
+- **Agent Safety dimension wired** (`check:agent-safety` in
+  `src/commands/release-trust.ts`, R8 surface): the structural evaluation
+  asserts the §17 safety wording on every installed skill surface, that the
+  MCP tool surface never opens the plugin trust gate, and that the agent edge
+  case (`fg-agent-unsafe-action`) stays registered in the False-Green Attack
+  Corpus. The release-trust contract's documented-unwired list shrinks to
+  artifact-integrity only (ships R9).
+
+### Changed
+
+- **Agent brief inherits the Constitution** (`src/commands/install-agents.ts`,
+  WI-22): every installed instruction surface (.claude/, .cursor/, .kilo/,
+  AGENTS.md) now carries the non-negotiable agent-safety contract — NEVER
+  declare trustworthiness without evidence · AGENT CLAIM ≠ VERIFICATION ·
+  NEVER manufacture, edit, or synthesize evidence · NEVER convert INCONCLUSIVE
+  to pass · NEVER suppress findings or weaken rules to get green — plus the
+  loop preconditions (FIX requires a proven actionable defect; RESCAN requires
+  changed-scope identification; PROOF requires fresh post-fix execution
+  evidence). Drift-locked by `tests/contract/agent-skill-surface.spec.ts`
+  (frozen surfaces only; safety wording asserted).
+
+### R7 Playwright capability matrix (remediation/remote-first WI-20)
+
+The Playwright capability matrix becomes a product surface with its own drift lock.
+
+### Added
+
+- **Playwright Capability Matrix** (`docs/PLAYWRIGHT-CAPABILITIES.md`,
+  `src/capabilities.ts`, WI-20): the product-depth surface — 12 Playwright
+  capabilities × 8 depth columns (detect · explain · produce evidence ·
+  correlate runtime · trust verdict · CLI · MCP · agents), every cell
+  explicitly classed (zero UNCLASSIFIED), every `yes` backed by a resolvable
+  evidence pointer (registered rule ID or in-repo artifact) with FAIL-CLOSED
+  validation: the generator refuses to render a claim on a dangling pointer.
+  Generated (`npm run docs:capabilities-playwright`) and drift-locked
+  (tests/contract/playwright-capabilities.spec.ts). The `to agents` column is
+  uniformly **no** until R8 ships the Agent Skill — stated, not implied.
+  Claims never exceed proven capability; rule counts stay out of the claim
+  surface entirely.
+
+### R6 forensic taxonomy + Selector Health v2 (remediation/remote-first WI-18+19)
+
+Forensic verdicts gain the semantic taxonomy, and Selector Health v2 replaces the locator heuristic.
+
+### Added
+
+- **Forensic verdict taxonomy** (`src/forensics/classify.ts`, WI-18): the
+  canonical §6 verdict set (likely-real-defect · environmental-failure ·
+  infrastructure-failure · flaky · retry-dependent · unstable-construction ·
+  **inconclusive default**) applied by a deterministic minimum-signal table —
+  a single weak signal can never classify confidently; conflicting signal
+  families force INCONCLUSIVE with an explicit `contradictory` evidence-state.
+  Every `TestVerdict` now carries a machine-visible `forensic` classification
+  (attempts + captured error text; sources without error text mark
+  `unsupported`, never a guess). Contradiction reconciliation
+  (`corroborates | contradicts | insufficient`) implements Contract H: the
+  runtime can corroborate but never silently weakens a static claim — a
+  contradiction renders the PAIR inconclusive while the claim stands.
+- **Selector Health v2** (`correlateSelectorHealth`, WI-19): runtime
+  correlation + concrete safe next actions; **no correlation ⇒ no claim** —
+  absent or merely-green runtime evidence yields no health claim in either
+  direction; the v1 static score is secondary and never altered here.
+
+### Changed
+
+- `TestRecord` gains an optional `errors` text surface (the trace ingester
+  populates it); `TestVerdict` gains the additive `forensic` field.
+
+### R5 trace ingester (remediation/remote-first WI-17)
+
+Trace forensics: bounded ingestion of Playwright trace.zip artifacts into the evidence core.
+
+### Added
+
+- **Playwright trace ingester** (`src/forensics/trace.ts`, WI-17): deterministic,
+  offline, bounded, version-aware ingestion of per-test traces — `trace.zip`
+  (a bounded, dependency-free ZIP reader: EOCD scan, central-directory
+  enumeration, stored/deflate members via `node:zlib` with a decompressed-output
+  cap) or raw `.trace`/`.ndjson` NDJSON streams. Action pairs become
+  Evidence-Core `TestRecord`s (start/end pairing, durations, per-action
+  errors; timeout errors render `timedOut`). `runForensics` recognizes trace
+  artifacts in both file and directory modes; the report source union gains
+  `playwright-trace` additively (`contractVersion 1` unchanged).
+- False-Green corpus cases for the trace surface: corrupt stream, truncated
+  stream, event-count overflow, unsupported version marker, zip without
+  `trace.trace` — all degrade to the zero-record exit-2 state, never a green
+  empty suite; plus positive controls (real stored zip + valid stream ingest
+  with paired durations) proving the rejections are precision, not blindness.
+
+### Changed
+
+- `ForensicsReport.source` + `RuntimeCorroboration.source` widened additively
+  with `"playwright-trace"`.
+
+### R4c Evidence Graph + Scope Integrity + Exit-Code proofs (remediation/remote-first)
+
+Every verdict now carries a machine-anchored evidence graph, scope-integrity accounting, and exit-code proofs.
+
+### Added
+
+- **Run Identity** (`src/engine/run-identity.ts`): the deterministic anchor —
+  `scanId = sha256(input snapshot fingerprint + rulesDigest + config
+fingerprint + engine version)`; set-identity semantics (input order does not
+  matter); every scan report carries `runIdentity` + `evidenceGraph` — the
+  chain-law links VERDICT ← EVIDENCE ← EXECUTION ← SCOPE ← SOURCE ← RULE(rev)
+  ← FIXTURE ← REPRODUCTION, each `ref` present only when its identity input
+  exists (no fabrication). The engine-version literal moved to the leaf module
+  `src/engine/version.ts` (cli.ts re-exports it as CLI_VERSION;
+  sync-sarif-version.cjs + version-consistency spec follow).
+- **Scope Integrity** (`ScanResult.scopeIntegrity`, additive): discovered /
+  analyzed / ignored / unrecognized / parseFailed / truncated counts +
+  `scopeVerdict` — PROVEN only when analyzed ≡ claimed scope; else PARTIAL
+  with named reasons. The terminal reporter renders the scope block and the
+  "repository verified" phrasing is forbidden output unless PROVEN. Walk-level
+  accounting: matcher exclusions (`onIgnored`) and unclaimed files
+  (`onUnrecognized`) are counted at the shared walk; parse failures are
+  counted at the rule stage.
+- **Exit-code decision proofs** (tests/blast-radius/scope-and-exit.spec.ts):
+  the frozen decision points exercised in both directions — trigger present →
+  frozen code, trigger absent → a different code — plus the closed frozen set
+  {0,1,2,10,20}.
+- Machine-contract doc regenerated with the three additive blocks
+  (`contractVersion 1` unchanged — additive within the schema).
+
+### Changed
+
+- Discovery accounting: the shared walk counts matcher-excluded files and
+  unclaimed files (ScanContext gains optional `onIgnored`/`onUnrecognized`;
+  all shared-walk adapters pass them through).
+
+### R4b False-Green Attack Corpus (remediation/remote-first)
+
+The False-Green Attack Corpus: hostile failure classes with mutation-based detection proofs.
+
+### Added
+
+- **tests/false-green/** — the adversarial corpus (plan §6, P0): 20 cases
+  across the plan's seven hostile classes (execution · parser · adapter ·
+  evidence · rule · mcp · agent failures), each declaring the seven
+  owner-required fields (INPUT / EXPECTED EXECUTION / EVIDENCE / VERDICT /
+  EXIT CODE / REPORT FIELDS / RELEASE IMPACT) and executed against real
+  surfaces with specific field bindings:
+  - execution: empty suite (score null + no-tests-found recorded), deadline
+    truncation, and the partial+findings never-blocks invariant (audit C5);
+  - parsers (through the real `runForensics` entry): corrupt JSON, truncated
+    Playwright report, malformed JUnit, unsupported schema → zero records →
+    exit-2 state — PARSER FAILURE ≠ CLEAN; duplicate retry-storm records stay
+    visible;
+  - adapters: scalar-jobs workflow fabricates nothing; broken YAML is SKIPPED
+    with accounting;
+  - rules: a throwing local plugin rule (QA-ACME-666) → `rulesCrashed ≥ 1`
+    with the scan completing — RULE CRASH ≠ CLEAN;
+  - evidence: missing/corrupt baseline → hasBaseline=false (exit 2); stale
+    baseline resolutions stay scoped to their capture; the foreign
+    baselineCommit is recorded (binding gate ships R4c);
+  - MCP: unknown tool / invalid params answer JSON-RPC errors, never success;
+  - agent: codegen and generated-header provenance classification — AGENT
+    CLAIM ≠ VERIFICATION.
+- **Mutation / assertion-strength protocol** (tests/false-green/mutation-
+  protocol.spec.ts): for every wired case and every report-field binding, the
+  false-green twin of the honest report (failure→success, partial→complete,
+  unknown→clean, crashed-rule→clean…) is injected and the case's assertion
+  must FAIL on it — a decorative assertion fails CI. Parser input twins flip
+  the hostile input to its benign form and require the observed verdict to
+  flip with it.
+- **Generated, drift-locked index** (npm run false-green:index + index.spec.ts):
+  one row per case with all seven declarations, the mutation inventory, and
+  the UNSURFACED rows (MCP transport internals / agent-action policy → R8;
+  artifact binding → R9) — recorded per Constitution §5, never silently
+  dropped. All seven plan classes present.
+
+### R4a Trust Constitution + Release Trust Verdict (remediation/remote-first)
+
+The Trust Constitution and the two-layer release-trust verdict algebra.
+
+### Added
+
+- **docs/TRUST-CONSTITUTION.md** — canonical law: CERTIFICATION-POLICY A1–A4
+  adopted as §1; the 18 PASS-forbidden conditions (verbatim); the closed status
+  algebra (PROVEN evidence-state → PASS/FAILED derivation, terminality rule,
+  record shape); the core law (`PASS = conclusion backed by sufficient
+evidence`); per-dimension applicability (UNSUPPORTED surfaces are recorded,
+  non-blocking, and drift-locked); publication honesty.
+- **docs/RELEASE-TRUST-CONTRACT.md** — the canonical 12 dimensions (fixed set,
+  fixed order, governance-locked): Engine/Evidence/Rule Integrity, Failure
+  Containment, Corpus Integrity, Contract Compatibility, Determinism, Scope
+  Integrity (ships R4c), Reproducibility, Zero-Network Compliance, Agent Safety
+  (R8), Artifact Integrity (R9).
+- New verb **`mjolnir release-trust`** emitting `mjolnir.release-trust@1` —
+  byte-deterministic (frozen key order, no timestamps, zero absolute paths),
+  per-dimension `evidence` + `determination` via the status algebra, verdict =
+  contract satisfaction (never a PROVEN count) with the binding system
+  invariant. Exit contract: 0 PASS · 1 non-PASS · 2 blocked context · 10 usage ·
+  20 internal. Drift-locked by tests/contract/release-trust-contract.spec.ts
+  (canonical set/order, binding resolution, derivation table + terminality,
+  byte-stability, path-freedom).
+
+### Changed
+
+- **release.yml**: the Release Trust Verdict gate is wired RELEASE-BLOCKING
+  pre-publish (Tests → Certification → CHANGELOG Gate → … → release-trust gate
+  → publish), running the BUILT binary; the verdict block + machine contract
+  ship with the GitHub Release (publication honesty — a missing proof renders
+  UNPROVEN, never omitted). No waiver path.
+
+### R4 blast radius audit (remediation/remote-first R4)
+
+The blast-radius audit: a machine-verified surface manifest with its own drift lock.
+
+### Added
+
+- **docs/BLAST-RADIUS-AUDIT.md** — the machine-verified surface manifest
+  (`npm run docs:blast-radius`): src inventory with per-area LOC, the internal
+  import fan-in ranking (change-blast candidates), the external dependency
+  allowlist, and the shipped surface (adapters, rules census, CLI flags, report
+  formats, frozen exit codes).
+- **tests/contract/blast-radius.spec.ts** — the machine-TESTABLE boundary
+  contract: the committed manifest must equal a fresh render; every external
+  import in src/ must belong to the allowlist (`yaml`, `ts-morph`,
+  `web-tree-sitter`, `tree-sitter-wasms`; node builtins are platform
+  contracts); every CLI flag parsed must appear in the manifest; every
+  `process.exit(N)` in src/ must be inside the frozen set (0/1/2/10/20).
+
+### P6 quarantine remediation (remediation/remote-first R3)
+
+Quarantine remediation: measured verdicts recorded, the quarantine ledger reconciled, and three rules restored to the live set.
+
+### Added
+
+- **docs/QUARANTINE-REMEDIATION.md** — the ledger-first quarantine view, generated
+  from the live registry (`npm run docs:quarantine-ledger`) and drift-locked
+  (tests/contract/quarantine-ledger.spec.ts): one row per live quarantine rule
+  with failure-mode class, disposition, and re-measure gate; historical section
+  records the governed retirements.
+- Python tree-sitter parse stage: `parsePythonAst` wired into the python
+  adapter's async `parseAst` hook (the §10 parse-or-fallback contract), with
+  `src/engine/python-ast.ts` structural queries — the first real python AST
+  substrate (the Sprint-8 "unwired" caveat is closed and re-pinned honestly).
+
+### Changed
+
+- **QA-PY-007** (detectorRevision 4, AST rework): fires only on ≥2-statement
+  with-blocks or broad root exception types — the adjudicated FP core
+  (single-statement/specific-type) suppressed. Corpus: pytest-dev 167 → 11,
+  pallets-click 16 → 1 live findings.
+- **QA-TQUAL-009** (detectorRevision 2, AST rework): skips Cypress command
+  chains (`cy.`-rooted — the driver awaits them) and deliberate `void`
+  discards. Corpus: cypress-realworld-app 10 → 0.
+- **QA-PW-147** (detectorRevision 2, final attempt): AST arm fires only on real
+  test/it declarations — code-as-data (`test('test')` inside lint-rule test
+  strings) can never fire. Corpus: eslint-plugin repo 32 → 0.
+- **QA-ENV-001** (detectorRevision 4, final attempt): OS-path sub-pattern
+  dropped (20/20 adjudicated FP — deliberate path fixtures, same undecidability
+  as the wave-2 host drop); locale/local-time families kept. Corpus: grafana
+  7 → 4.
+- Measurement: orphaned verdicts (findings the reworks suppressed) archived to
+  `tests/corpus/verdicts/archive/` per the established prune flow; the three
+  fully-reworked rules fall below the n ≥ 10 threshold and ship UNMEASURED
+  until owner re-adjudication (measured census 77 → 74 of 79; the
+  certification floor test documents the P6 invalidations).
+
+### P3c Jenkins (remediation/remote-first R2)
+
+Jenkins support: a bounded Jenkinsfile scanner and the QA-CI Jenkins arms (retry masking, catchError rescue, silent swallow).
+
+### Added
+
+- Jenkinsfile detection: the root `Jenkinsfile` (declarative and scripted
+  pipelines) is now discovered and scanned as a TEXT-target kind — a bounded,
+  string-aware Groovy block scanner (`sh` segments, `catchError` blocks,
+  `try`/`catch` pairs); no new language grammar (master-plan P3c wording).
+- New rule **QA-CI-014** "try/catch swallows a verification-stage failure" —
+  a `try` running a gate whose `catch` neither rethrows, calls `error(...)`,
+  marks `currentBuild.result`, nor downgrades via `unstable()`. BORN
+  QUARANTINE (§15.5): opt-in via `--strict` until corpus-measured.
+
+### Changed
+
+- **QA-CI-002** (detectorRevision 4): Jenkinsfile routing — the lexical
+  `|| true` scan now reaches `sh` strings.
+- **QA-CI-008** (detectorRevision 4): Jenkinsfile arms —
+  `catchError(buildResult: 'SUCCESS')` wrapping a gate, and `unstable()` used
+  as a rescue for a failed verification stage (master-plan P3c shapes;
+  `buildResult: 'UNSTABLE'` is a visible downgrade and never fires).
+- **QA-CI-009** (detectorRevision 3): Jenkinsfile arm — `sh` running a
+  verification gate with `returnStatus: true` discards the exit code.
+- Measurement: sidecar + `MEASURED_FP` re-recorded for QA-CI-002/008/009
+  (corpus re-run: no corpus repo carries a root Jenkinsfile, so the
+  classified verdict evidence carries over unchanged).
+
+### P3b Azure DevOps (remediation/remote-first R1)
+
+Azure DevOps support: guarded azure-pipelines.yml parsing, the QA-CI Azure arms, and the adapter's honest accounting.
+
+### Added
+
+- Azure DevOps pipeline detection: `azure-pipelines.yml` at the repo root is now
+  discovered and scanned (`azure-pipelines` adapter, safe-YAML machinery shared
+  with the GitHub Actions parser — alias-bomb guard, depth cap, prototype-safe
+  keys; docs/AZURE-DEVOPS.md).
+- New rule **QA-CI-013** "Verification gate conditioned so it can never fail the
+  pipeline" — `condition: failed()` rescue, `condition: false`, `enabled: false`
+  on Azure verification gates. BORN QUARANTINE (§15.5): opt-in via `--strict`
+  until corpus-measured; never silent-core.
+- docs/AZURE-DEVOPS.md — platform recipe with the frozen exit-code contract.
+
+### Changed
+
+- **QA-CI-001** (detectorRevision 3): Azure DevOps arm — `continueOnError: true`
+  on a verification step or a gate-bearing job (same mechanism, framework-tagged
+  `azure-pipelines`).
+- **QA-CI-002** (detectorRevision 3): Azure routing — the lexical `|| true` scan
+  now reaches `bash:`/`pwsh:` script blocks in azure-pipelines.yml.
+- **QA-CI-007** (detectorRevision 3): Azure DevOps arm — `retryCountOnTaskFailure`
+  on verification tasks.
+- **QA-CI-008** (detectorRevision 3): Azure DevOps arm — verification gate jobs
+  conditioned `always()` / `succeededOrFailed()` (master-plan P3b shape).
+- Measurement: sidecar + `MEASURED_FP` re-recorded at detectorRevision 3 for
+  QA-CI-001/002/007/008; corpus re-run showed zero QA-CI count drift (no corpus
+  repo carries a discoverable azure-pipelines.yml), so the existing classified
+  verdicts remain the measurement evidence.
+
 ## [1.0.5] — 2026-09-10
 
 ### Changes since 1

package/README.md

@@ -146,7 +146,7 @@ QA impact:   False-green risk (FALSE-GREEN)
 Measured FP: 11% (19 hand-classified corpus verdicts)
 FP risk:     low (author estimate)
 Languages:   yaml
-Frameworks:  github-actions
+Frameworks:  github-actions, azure-pipelines
 
 WHAT WAS FOUND (real detector output, not a mockup)
   Job `security-scan` runs a verification gate under `continue-on-error: true`.
@@ -262,7 +262,7 @@ have no such requirement.)
 
 ## What Mjölnir finds
 
-**<!-- census:total-rules -->77 rules<!-- /census:total-rules -->** in four families — **test hygiene**, **test quality**,
+**<!-- census:total-rules -->79 rules<!-- /census:total-rules -->** in four families — **test hygiene**, **test quality**,
 **Playwright**, **CI integrity** — over TypeScript/JavaScript, Python,
 Java, C# and GitHub Actions YAML, covering Playwright in all four bindings
 plus pytest, JUnit, TestNG, NUnit, xUnit, MSTest, Jest, Vitest and Mocha,
@@ -442,9 +442,9 @@ Rung by rung: [docs/TERMINOLOGY.md](docs/TERMINOLOGY.md).
 
 ### How much of this is measured
 
-**<!-- census:measured-of-total -->77 of 77<!-- /census:measured-of-total --> rules carry a false-positive rate measured against real OSS code**
+**<!-- census:measured-of-total -->74 of 79<!-- /census:measured-of-total --> rules carry a false-positive rate measured against real OSS code**
 (≥ 10 hand-classified findings each — [docs/FP-AUDIT.md](docs/FP-AUDIT.md)).
-The other <!-- census:unmeasured -->0<!-- /census:unmeasured --> ship on the author's estimate and say so, per rule, in
+The other <!-- census:unmeasured -->5<!-- /census:unmeasured --> ship on the author's estimate and say so, per rule, in
 `mjolnir explain`; `mjolnir rules --unmeasured` lists them, and every scan
 footer reports how many of the rules that actually _fired_ are measured.
 
@@ -685,7 +685,7 @@ artifacts.
   product does what the requirement asked for.
 - **A 100 is not proof of a good suite.** Whether your suite covers your
   actual risk is a different question, and this tool does not answer it.
-- **<!-- census:unmeasured-of-total -->0 of 77<!-- /census:unmeasured-of-total --> rules ship on an estimate**, not a measured rate — disclosed
+- **<!-- census:unmeasured-of-total -->5 of 79<!-- /census:unmeasured-of-total --> rules ship on an estimate**, not a measured rate — disclosed
   per rule, not buried here.
 - **E1 is not E2.** Heuristic findings are worth reading, not worth
   applying blindly.

package/dist/cli.d.mts

@@ -80,7 +80,7 @@ interface RuntimeCorroboration {
    * constrains TRUE-FLAKE derivation, which lives in the analysis, not
    * in the provenance label.
    */
-  source: "playwright-json" | "junit-xml" | "jest-json" | "vitest-json";
+  source: "playwright-json" | "junit-xml" | "jest-json" | "vitest-json" | "playwright-trace";
   /** Number of tests executed in the finding's file (any level). */
   testsExecuted: number;
   /**
@@ -304,6 +304,62 @@ interface ScanResult {
      */
     rulesCrashed?: number;
   };
+  /**
+   * Scope Integrity block (product-gap master plan §7, R4c): the
+   * claimed-vs-analyzed accounting. `scopeVerdict` is PROVEN only when
+   * every discovered file was analyzed — no matcher exclusions, no
+   * unrecognized files, no parse failures, no truncation. Additive
+   * within schemaVersion 1.
+   */
+  scopeIntegrity?: {
+    /** Files discovery claimed for adapters. */
+    discovered: number;
+    /** Files that reached (and survived) the rule stage. */
+    analyzed: number;
+    /** Files excluded by the ignore matcher (counted at the walk). */
+    ignored: number;
+    /** Files the walk saw but no adapter claims. */
+    unrecognized: number;
+    /** Discovered files whose parse/analysis threw (counted, never fatal). */
+    parseFailed: number;
+    /** Named truncation events (deadline, file caps). */
+    truncated: number;
+    /** PROVEN only when analyzed ≡ claimed scope; else PARTIAL + reasons. */
+    scopeVerdict: "PROVEN" | "PARTIAL";
+    /** The named scope reasons, present only when PARTIAL. */
+    reasons?: string[];
+  };
+  /**
+   * Run Identity (R4c): the deterministic anchor binding verdict ←
+   * evidence ← execution ← scope ← source ← rule(rev). Present when the
+   * execution was machine-anchored; never fabricated.
+   */
+  runIdentity?: {
+    scanId: string;
+    inputFingerprint: string;
+    rulesDigest: string;
+    configFingerprint: string;
+    engineVersion: string;
+  };
+  /**
+   * Evidence Graph (R4c): the chain-law links (VERDICT ← EVIDENCE ←
+   * EXECUTION ← SCOPE ← SOURCE ← RULE(rev) ← FIXTURE ← REPRODUCTION).
+   * A link's `ref` is present only when its identity input exists — the
+   * CHAIN is always emitted so unbound links stay visible.
+   */
+  evidenceGraph?: {
+    chain: Array<{
+      link: string;
+      ref?: string;
+    }>;
+    runId?: {
+      scanId: string;
+      inputFingerprint: string;
+      rulesDigest: string;
+      configFingerprint: string;
+      engineVersion: string;
+    };
+  };
   /**
    * Local incremental cache report (Beta-to-Stable plan, M5.2). Present
    * only when the scan ran with `--cache`; additive within
@@ -848,6 +904,17 @@ declare function isValidFindingRecord$1(f: unknown): f is Omit<Finding, "ruleId"
  */
 declare function runScan$1(args: CliArgs, hooks?: ScanHooks): Promise<ScanResult>;
 //#endregion
+//#region src/engine/version.d.ts
+/**
+ * Engine version, as a leaf module: `runIdentity` (R4c) needs the engine
+ * version inside scan-pipeline WITHOUT importing cli.ts (a cycle — cli
+ * imports the pipeline). The literal follows the same discipline as
+ * CLI_VERSION and SARIF's driver.version: kept in sync on release by
+ * scripts/sync-sarif-version.cjs and guarded by the version-consistency
+ * spec. cli.ts re-exports this as CLI_VERSION.
+ */
+declare const ENGINE_VERSION = "1.0.8";
+//#endregion
 //#region src/cli-io.d.ts
 /**
  * Shared process IO sinks (certification-audit Phase 5, G6): the variadic
@@ -870,17 +937,6 @@ declare function runDoctorCommand(argv: string[], io?: {
 //#endregion
 //#region src/cli.d.ts
 declare const runScan: typeof runScan$1, buildUniversalRules: typeof buildUniversalRules$1, fallbackWorkspace: typeof fallbackWorkspace$1, pathMatchesGlob: typeof pathMatchesGlob$1, isValidFindingRecord: typeof isValidFindingRecord$1, discoverRuntimeReport: typeof discoverRuntimeReport$1, KNOWN_RULE_IDS: ReadonlySet<string>, OVERLAP_META_BY_RULE_ID: ReadonlyMap<string, OverlapMeta>, EVIDENCE_OVERRIDES: ReadonlyMap<string, string>, SUITE_INVALIDATING_RULE_IDS: ReadonlySet<string>;
-/**
- * Tool version for `mjolnir --version`.
- *
- * A literal, not a package.json read: the shipped artifact is a single
- * bundled `dist/cli.mjs`, so resolving package.json at runtime depends on
- * where the file happens to sit after install. This follows the same
- * discipline as SARIF's `driver.version` — kept in sync by
- * `scripts/sync-sarif-version.cjs` on release and guarded by
- * `tests/version-consistency.spec.ts` locally.
- */
-declare const CLI_VERSION = "1.0.5";
 /** A usage-error detail: the offending token, when one exists. */
 interface UsageErrorDetail {
   /** The unknown flag or rejected value (e.g. `--nope`, `loud`). */
@@ -1050,4 +1106,4 @@ declare function runHelpCommand(argv: string[], io?: {
 }): number;
 declare function isEntryPoint(): boolean;
 //#endregion
-export { CLI_VERSION, type CliArgs, EVIDENCE_OVERRIDES, KNOWN_RULE_IDS, OVERLAP_META_BY_RULE_ID, type Output, SUITE_INVALIDATING_RULE_IDS, type ScanHooks, UsageErrorDetail, buildUniversalRules, discoverRuntimeReport, err, exitForFindings, fallbackWorkspace, internalErrorMessage, isEntryPoint, isValidFindingRecord, levenshtein, main, nearestFlags, out, parseArgs, pathMatchesGlob, runBadgeCommand, runBaselineCommand, runCiInstall, runCreateRuleCommand, runDebtCommand, runDiffCommand, runDoctorCommand, runDoctorPlaywright, runExplainCommand, runFixCommand, runForensicsCommand, runHandoverCommand, runHelpCommand, runImpactCommand, runInitCommand, runMutationCommand, runPrCommentCommand, runPwReportCommand, runRulesCommand, runScan, runScanCommand, runStatsCommand, runSuppressions, runTriageCommand, runVerifyCommand, usageErrorMessage };
+export { ENGINE_VERSION as CLI_VERSION, type CliArgs, EVIDENCE_OVERRIDES, KNOWN_RULE_IDS, OVERLAP_META_BY_RULE_ID, type Output, SUITE_INVALIDATING_RULE_IDS, type ScanHooks, UsageErrorDetail, buildUniversalRules, discoverRuntimeReport, err, exitForFindings, fallbackWorkspace, internalErrorMessage, isEntryPoint, isValidFindingRecord, levenshtein, main, nearestFlags, out, parseArgs, pathMatchesGlob, runBadgeCommand, runBaselineCommand, runCiInstall, runCreateRuleCommand, runDebtCommand, runDiffCommand, runDoctorCommand, runDoctorPlaywright, runExplainCommand, runFixCommand, runForensicsCommand, runHandoverCommand, runHelpCommand, runImpactCommand, runInitCommand, runMutationCommand, runPrCommentCommand, runPwReportCommand, runRulesCommand, runScan, runScanCommand, runStatsCommand, runSuppressions, runTriageCommand, runVerifyCommand, usageErrorMessage };