mjolnir-qa 0.4.0 → 0.5.2

package/CHANGELOG.md

@@ -9,29 +9,1034 @@ 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.
 
-## [0.4.0] — 2026-08-27
+## [0.5.2] — 2026-09-05
 
-### Changed
+### npm 12 pack-shape repair of the release pipeline
 
-- **BREAKING: Rebranded from QA Doctor to Mjölnir.** Package name is now
-  `mjolnir-qa` (bin: `mjolnir`). Config file: `mjolnir.config.json`.
-  Data directory: `.mjolnir/`. Badge: `mjolnir-badge.json`.
-- Score label: "SCORE" → "WORTHINESS".
-- Verdicts: "HEALTHY" → "WORTHY", "CRITICAL" → "UNWORTHY".
-- Environment variable: `QA_DOCTOR_ASCII` → `MJOLNIR_ASCII`.
-- SARIF tool.driver.name: "Mjölnir".
-- Repository: `github.com/Sergey-Bar/Mjolnir`.
-- CLI: all help text, error messages, usage strings reference `mjolnir`.
-- Generated workflows: `mjolnir.yml`, `npx mjolnir-qa@latest`.
+### Fixed
 
-### Added
+- The fresh-install gate (`tests/integrations/registry-install.spec.ts`),
+  `tests/integrations/package-smoke.spec.ts`, and
+  `tests/e2e/journey-1-first-run.spec.ts` parse `npm pack --json` through
+  a shared shape-tolerant helper
+  (`tests/helpers/npm-pack-json.ts`): npm 12 changed the output from an
+  array to an object keyed by package name, which stopped the v0.5.1
+  publish at the gate — tag cut, nothing shipped, by design.
+- `release.yml` upgrades to `npm@11` (the proven line) instead of
+  `npm@latest`: toolchain majors must be deliberate, verified changes,
+  never implicit drift on the publish path.
+- `registry-install.spec.ts` cleanup no longer cascades a second error
+  when `beforeAll` fails early (the cascade buried the real diagnosis).
 
-- `--format mermaid` — test-architecture diagram (Sprint 9).
-- `--tone blunt` — opt-in blunter messages (Sprint 9).
-- Milestones — first flawless scan / first debt reduction announced once.
-- New MJÖLNIR ASCII art logo (minimal Nordic hammer).
+## [0.5.1] — 2026-09-05
+
+### Verification Trust Evolution, Phase 8 — Local Extensibility (plan §18)
+
+### Added — folder-based external rules, zero network
+
+- **`mjolnir-rules/` contract** (`src/plugins/local-rules.ts`): a
+  workspace directory loaded from the scan target root alongside npm
+  plugins. Two file kinds: **JSON rule manifests** (declarative regex
+  patterns — NO code executed; id/title/severity/category/appliesTo/
+  patterns/message/why/fix/languages/frameworks) and **JS modules**
+  (`rules: QADoctorRule[]`, full-Node trust, same posture as npm
+  plugins). Missing directory → no-op.
+- **Same trust contract as core/npm plugins:** reserved core prefixes
+  rejected (case-insensitive spoofing guard); bad metadata/regexes
+  degrade to warning entries (QA-PLUGIN-000), never a crash; external
+  rules carry the full trust metadata shape and are born
+  quarantine/unmeasured. **Core-tier clamp**: an external rule
+  declaring `tier: "core"` is clamped to `extended` with a load
+  warning — core requires a measured FP rate from the committed corpus
+  sidecar, which external rules cannot have.
+- **Tier caps obeyed, filter unified:** the quarantine exclusion filter
+  in `buildUniversalRules` now consults the tier map that includes
+  plugin/external tiers — plugin-declared quarantine rules were
+  previously excluded only when the core registry knew the ID; the
+  unified filter excludes them from non-strict scans exactly like core
+  (post-scan cap still observable under `--strict`; the cli-scan-arms
+  test updated to cover both sides).
+- **Drift-checked:** `mjolnir rules --md --external` renders the
+  catalog from the LOADED external rules with a provenance column
+  (`core`/`external`) — an on-disk edit changes the next render; the
+  catalog can never drift from what actually ships.
+  `scripts/generate-capability-matrix.ts --external <root>` writes a
+  workspace-local `MJOLNIR-RULES-MATRIX.md` with provenance "external"
+  (unmeasured by definition — outside the corpus sidecar). The
+  committed matrix stays core-registry-only and byte-stable.
+- **S-8 disclosure:** external rule surfaces appear in the scan's
+  plugin disclosure block.
+
+Registry/marketplace explicitly deferred (plan §18).
+
+### Verification Trust Evolution, Phase 7 — Agentic QA Trust (plan §17)
+
+### Added — Agentic Trust Profile (plan §17.2, §17.4)
+
+- **Provenance detection** (`src/engine/provenance.ts`): per-file static
+  markers, EXTRACTED not invented — the industry-wide generated-file
+  header convention ("auto-generated"/"generated by"/"do not edit") and
+  the Playwright codegen recorder fingerprint (its default test title
+  is exactly `'test'`). Honest boundary, documented in the module:
+  healed and MCP-edited files WITHOUT markers are statically
+  indistinguishable from hand-written ones — the classifier returns
+  `unmarked`, never a guess.
+- **`ScanResult.agenticProfile`** (additive): share of test files with
+  detected generative markers + the findings split across those
+  surfaces, with the boundary in the profile's own `note`. Surfaced in
+  the terminal report only when something was actually detected
+  (silence over noise); JSON carries it on every scan.
+- **§17.4 — same evidence standard regardless of author:** the profile
+  is metadata only. It never changes scoring, evidence levels, tier
+  caps, or rule behavior. AI tests earn trust through exactly the same
+  evidence as human ones.
+
+### Added — locator.normalize() framework-standards rules (plan §17.3)
+
+Two new frozen `QA-PW-*` rules, BORN QUARANTINE (§17 exit gate:
+measured before leaving provisional), fixtures both directions:
+
+- **QA-PW-146 — CSS/XPath string selector instead of a normalized
+  locator** (warning, heuristic): `css=`/`xpath=` engine prefixes, bare
+  id/class/attr CSS, and `nth-child` chains inside
+  `.locator()`/`waitForSelector()`/`page.$` — aligned with Playwright's
+  locator standard. Fills the TS gap (the brittle-selectors family
+  covers Java/C#/Python; the retired QA-PW-112 was style police, this
+  is markup-coupling detection). Not auto-fixable: the normalized
+  getter depends on app semantics only a human knows — the fix field
+  carries the concrete `getByRole`/`getByTestId`/`getByText`
+  suggestion. Both fixtures + all three cached OSS corpora (0 fires —
+  real suites use normalized locators) recorded.
+- **QA-PW-147 — codegen default test title** (info, observation): a
+  committed spec still titled `'test'`/`'test N'` is an unreviewed
+  recording artifact — the §17.1 provenance marker as a finding.
+
+Registry: 99 rules (73 measured). Docs pages + capability matrix
+regenerated.
+
+### Verification Trust Evolution, Phase 6 — Runtime Evidence (plan §16)
+
+### Added — runtime corroboration + the honest L0–L5 trust ladder
+
+Built on the existing forensics ingestion (no greenfield):
+`packages/playwright-reporter` → `mjolnir.report.json` → the
+`ForensicsReport` pipeline that `forensics`/`triage`/`pw-report`
+already consume. Findings gain two additive, optional fields
+(schemaVersion 1 unchanged):
+
+- **`runtimeCorroboration`** — what a real run report vouches for:
+  `level: "file" | "test" | "defect"`, the report source, executed-test
+  count, and (test/defect level) the containing test's full verdict
+  (attempts, final status, passed-on-retry, ever-failed, skipped).
+- **`trustLevel`** — the L0–L5 ladder, derived deterministically:
+  L0 (E0 observation) / L1 (E1 heuristic) / L2 (E2 deterministic) are
+  the static-only ceiling; L3 (file executed) / L4 (the containing
+  test executed — matched via the report's spec declaration lines) /
+  L5 (the run verdict directly corroborates the defect class — a
+  FLAKY-RISK finding whose test actually flaked, retried, or timed
+  out) exist ONLY when runtime corroboration is present. The
+  no-static-only-L4/L5 invariant is structurally enforced in
+  `deriveTrustLevel` and locked by tests across the full
+  findingType×confidence×evidenceLevel matrix.
+
+- **Matching is honest by construction:** Playwright JSON reports now
+  carry the spec declaration line (additive `line` on
+  `TestRecord`/`TestVerdict`); a finding is tied to a specific test
+  only when the report's declaration spans place it there — otherwise
+  corroboration stays at file level, and files the report never ran
+  get NOTHING (no fabricated evidence). JUnit XML has no locations:
+  file-level only.
+- **Scan wiring:** `runScan` auto-discovers a run report next to the
+  scan target (`mjolnir.report.json` — the reporter package's default
+  output — or a `test-results/` directory), runs the existing
+  forensics ingestion, and stamps findings. No report → findings
+  unchanged (honest "runtime evidence: not available"). A hostile
+  report degrades the scan, never fails it.
+- **Reporters split verified vs assumed (plan §16):** the terminal
+  footer reports `Runtime evidence: N/M findings corroborated by a
+real run report (trust L3–L5)` or the explicit not-available line;
+  finding cards show `trust L4 · runtime: test executed`. SARIF
+  results carry `trustLevel` + `runtimeCorroboration` in properties.
+
+### Verification Trust Evolution, Phase 5 — Framework Expansion (plan §15, D7 closed)
+
+### Added — FrameworkDimension enforced (plan §15.1, defect D7 closed)
+
+- **Real dependency parsing per build system:** `package.json` (JSON),
+  `pom.xml` (Maven `<dependency>` blocks), `build.gradle(.kts)`
+  (dependency-statement coordinates), and EVERY `.csproj` at the root
+  (`<PackageReference Include="…">` attributes — the old "first
+  `.csproj` only" defect is closed), plus requirements*.txt for Python.
+- **Per-file framework tags** (`ParsedFile.frameworkTags`): derived from
+  the file's OWN imports/usings/import-lines — `@playwright/test` →
+  "playwright", `cypress` → "cypress", `org.junit.*` → "junit",
+  `NUnit` → "nunit", `import selenium` → "selenium", etc. AST-truth for
+  Java/C#/TS; import-line scan for Python (no AST seam by design).
+- **`rule.frameworks ∩ file.frameworkTags` filtering** with
+  open-when-unknown: a rule that declares `frameworks` runs on a file
+  only when the file's tags intersect it; files without tags and rules
+  without `frameworks` are always analyzed — the dimension narrows, it
+  never silently drops evidence. Shared in
+  `src/engine/adapter.ts` (`frameworkFilterApplies`), enforced in the
+  TS/Java/C#/Python adapters' `runRules`.
+- **Generalized config-gating (§15.2):** `configRule: true` +
+  `configFiles: string[]` (regex sources) replaces the hard-coded
+  `playwright.config.*` regex that lived in the TS adapter and
+  duplicated inside all five config rules (QA-PW-121/122/141/143/144
+  migrated). Discovery knows the config filename conventions
+  (`cypress.config.*` added alongside `playwright.config.*`).
+
+### Added — Cypress integration (first framework per the §15 order)
+
+Three rules in the new frozen `QA-CYP-*` namespace, BORN QUARANTINE
+(§15.5) with fixtures both directions (`tests/fixtures/QA-CYP-*/`):
+
+- **QA-CYP-001 — fixed `cy.wait(n)`** (warning): numeric-literal
+  `cy.wait(3000)` — the Cypress hard-sleep idiom. Alias waits
+  (`cy.wait('@route')`) are the legitimate form and never fire.
+  File gate: framework tag, `.cy.*` extension, or `cy.*` API usage in
+  the file (real Cypress suites rarely import cypress — surfaced by the
+  first measurement against cypress-example-kitchensink). MEASURED on
+  the kitchensink corpus: 12 TP (viewport-switch fixed waits) / 3 FP
+  (doc-example artifacts, the suite intentionally demonstrates the API
+  — QA-PY-003 precedent) → 20% FP at n=15, 95% Wilson
+  [7.1%, 45.2%]; tier extended (band-consistent), corpus
+  cypress-io-kitchensink added with baselines.
+  cypress-realworld-app scanned as precision evidence: 0 fires (the
+  suite uses alias waits exclusively), baseline recorded.
+- **QA-CYP-002 — focused test (`.only`)** (error): committed
+  `it.only`/`describe.only`/`context.only` de-schedules the rest of the
+  suite. Quarantine, unmeasured (fixtures only) pending a measured
+  Cypress corpus with the pattern.
+- **QA-CYP-003 — `chromeWebSecurity: false`** (error): deterministic
+  config defect via the generalized `configFiles` gate. Quarantine,
+  unmeasured pending corpus.
+
+### Added — Selenium cross-language reach (plan §15.3, JV/CS/Py reuse)
+
+Three rules in the new frozen `QA-SE-*` namespace sharing one sequence
+detector (hard sleep followed by an element lookup within 3 lines —
+the sleep standing in for an explicit `WebDriverWait`), all BORN
+QUARANTINE:
+
+- **QA-SE-001** (Java): `Thread.sleep` → `findElement`/`click`/
+  `sendKeys` within 3 lines.
+- **QA-SE-002** (C#): `Thread.Sleep`/`Task.Delay` →
+  `FindElement`/`Click`/`SendKeys` within 3 lines. MEASURED on the
+  SeleniumHQ/selenium .NET webdriver suite: 3 TP + 1 FP (a
+  sleep-inside-polling-loop cadence — loop-body containment is the
+  documented residue) → 25% FP at n=4; below the n≥10 measurement bar,
+  stays quarantine/unmeasured with verdicts recorded
+  (tests/corpus/verdicts/SeleniumHQ-selenium.jsonl, baseline added).
+- **QA-SE-003** (Python): `time.sleep` → `find_element`/`click`/
+  `send_keys` within 3 lines.
+
+### Registry
+
+- 97 rules (was 91); registry/doctor/scaffolder ID validators widened
+  for the new frozen namespaces (QA-CYP-_, QA-SE-_, QA-WDIO-_,
+  QA-PPTR-_, QA-APM-* reserved). Measured coverage 73/97 (75%); the
+  §20.1(a) unmeasured-count ratchet now tracks the PRE-EXISTING set
+  mechanically (`introduced` ≤ 0.5.0) — new waves onboard under §20.1(b)
+  - the per-framework exit gate, not the global count freeze.
+- Defect ledger: **D7 closed** (target phase 5). Capability matrix
+  regenerated (97 rules); docs pages regenerated; corpus baselines
+  added for the three Phase 5 measurement repos.
+
+### Verification Trust Evolution, Phase 4 — Common QA Semantic Model (plan §14, behavior-neutral)
+
+### Added — `src/engine/qa-model.ts`: the normalized QA concept IR (extract-only, no scan wiring)
+
+The plan-§14 vocabulary — Test, TestBoundary, Setup/Teardown, Fixture,
+Action, Locator, Wait, Assertion, Mock, NetworkInteraction, Retry,
+Navigation, Interaction, Lifecycle — now exists as a typed model
+(`QaNode`/`QaSemanticModel`, all 14 concepts in `EXTRACTOR_COVERAGE`)
+with per-language extractors over the ALREADY-existing parse stage:
+ts-morph for TS/JS (tests/hooks via the scorer's it/test vocabulary,
+calls via the measured rule vocabularies of qa-pw-002/004/005/101–145,
+qa-test-001/003/004/006, jest/vi mock+retry), tree-sitter for Java/C#
+(test boundaries REUSE `javaTestMethods`/`csharpTestMethods` verbatim —
+the model cannot drift from the rules' scoping; call/hook/retry
+vocabularies copied from the QA-JV-_/QA-CS-_ rules), regex boundaries
+for Python (def test_, @pytest.fixture, time.sleep, assert).
+
+- **Extracted, not invented:** every classification table cites the
+  rule whose measured vocabulary it copies (headers in qa-model.ts);
+  the shared `isHelperIdiom` helper moved to `jv-cs-ast.ts` so the
+  QA-CS-103 rule and the model share ONE implementation.
+- **Adoption is additive:** NOTHING in the scan pipeline imports the
+  module — BEHAVIOR-NEUTRAL by construction, golden/corpus locks
+  untouched (proven: locks byte-identical with the model present).
+- **Coverage is honest:** `EXTRACTOR_COVERAGE` documents per-language
+  gaps (e.g. Java has no fixture/interaction/lifecycle extractor;
+  Python extracts only test/fixture/assertion/wait) — visible gaps,
+  never silent claims (No False Proof).
+- **Equivalence proven:** `tests/qa-model.spec.ts` re-expresses the
+  QA-JV-103 / QA-CS-103 / QA-CS-102 oracles over the model
+  (`testVerifies`, ancestor-chain containment via the
+  `firstAncestorCallNamed` generalization now carried as
+  `node.ancestors`) and asserts FINDING-IDENTICAL results against the
+  rules on the committed fixture corpora + synthetic edge shapes —
+  the model can carry these rules without changing any output.
+  Awaitedness (qa-pw-002's consumption oracle) rides on TS nodes as
+  `node.awaited`.
+
+### Verification Trust Evolution, Phase 3 — Java/C# semantic upgrade (plan §13)
+
+### Changed — three JV/CS rules migrated to L2 tree-sitter analysis (EVIDENCE-BACKED, detectorRevision 2)
+
+The Phase 0.5 parse stage's tree-sitter trees (`parseJavaAst`/`parseCSharpAst`,
+delivered via `ParsedFile.ast`) are now consumed by rules: a new L2
+structural-analysis layer (`src/engine/jv-cs-ast.ts`) provides test-method
+scoping by annotation/attribute, invocation structure, and call/argument
+containment; the rule contract gains an optional `astQuery` hook whose
+regex path is a MANDATORY fallback (no AST ⇒ regex, never a second
+detector, plan §13.2). No type/symbol semantics are promised for JV/CS
+(no Roslyn, no classpath) — semantic depth is L2 per plan §13.4.
+
+- **QA-JV-103 (test without assertions), 50% FP (n=20) → 25.9% FP
+  (n=58), quarantine → extended.** Test boundaries now come from real
+  `method_declaration` nodes (any `@Test`/`@org.junit.Test`, argumented
+  TestNG forms included). The assertion oracle adds the two measured
+  rev-1 FP classes: Playwright's THROWING waits (`waitFor*` except
+  `waitForTimeout` — they throw on timeout, so the wait IS the
+  verification; 6 rev-1 FPs) and `verify*/check*/assert*` helper
+  calls (4 rev-1 FPs). Remaining FP class, documented as the L2
+  boundary: assertions behind arbitrarily-named helpers (keycloak's
+  `testValidationValid`, playwright-java's `testEnterKey`,
+  appsmith's `check`).
+- **QA-CS-103 (test without assertions), 95% FP (n=20) → 0% FP (n=9),
+  quarantine → core.** The rev-1 oracle missed the whole Shouldly
+  extension family — 17 of the 19 FPs were `ShouldBeOfType<...>(
+...).Message.ShouldBe(...)`-style assertion-rich tests (spectre-console,
+  310 count-lock fires → 0). The rev-2 oracle counts Shouldly chains,
+  `Assert`/`Should` receivers, the Verify snapshot framework,
+  PascalCase `verify*/check*/assert*` helpers (Humanizer's
+  `VerifyAnalyzerAsync` — the lowercase-only form was a Java camelCase
+  inheritance bug), C# throwing waits (`WaitFor*Async` except
+  `WaitForTimeoutAsync`), and conditional `throw new
+*Assertion*Exception` (playwright-dotnet's ConventionTests). Grammar
+  error-node guard: a truncated parse never produces a finding.
+- **QA-CS-102 (`Thread.Sleep`/`Task.Delay` hard sleep), 65% FP (n=20) →
+  8.3% FP (n=24), quarantine → core.** Tree-sitter invocation scoping
+  excludes the measured environment-simulation classes: delays inside
+  route/expose/server delegates (`Route*Async`, `SetRoute`,
+  `ExposeFunction*`, 10 rev-1 FPs + delta), deliberate infinite/negative
+  blocks (`Task.Delay(-1)`, `int.MaxValue`, `Timeout.Infinite*`),
+  `Task.WhenAny` timeout races, and runner payload fixtures
+  (`RunAndWaitFor{Request,RequestFinished,Response}Async(() => Task.Delay…)`,
+  `UnrouteAllAsync` timing-window sleeps stay flagged — their delay
+  creates the timing window the assertion measures). Documented
+  trade-off: one rev-1 TP (a sub-second artificial-timing delay inside a
+  route delegate) is no longer flagged; the structural boundary cannot
+  read that intent.
+
+All three: verdict corpus reconciled per the §07 loop (41 superseded rows
+removed, 49 fresh rows adjudicated from source with per-row notes, class-B
+positive fixtures for QA-CS-102/QA-JV-103 and class-C negative fixtures
+for all three), count-lock baselines refreshed (microsoft-playwright-java
+QA-JV-103 97→43, microsoft-playwright-dotnet QA-CS-102 52→21,
+spectre-console QA-CS-103 310→0, Humanizer 34→2; keycloak +602 is a
+recall gain — the rev-1 regex missed `void x() throws Exception`
+signatures, tree-sitter scoping does not), and
+`detector-revisions.json` bumped (sidecar entries stay; measurements
+stamped rev 2). QA-JV-102 stays LEXICAL rev 1 — no migration for
+symmetry (plan §12.4).
+
+### Verification Trust Evolution, Phase 2 — quarantine-cluster triage (plan §12.2)
+
+### Deprecated — 21 rules retired per docs/RULE-LIFECYCLE.md (measured 100% FP, premise wrong)
+
+Every rule below measured 100% FP (zero TPs at n ≥ 10, `docs/FP-AUDIT.md`)
+on real-world code. Per the lifecycle policy the severity is downgraded to
+`info` (non-blocking everywhere) and `falsePositiveRisk` is set to `high`;
+the code and fixtures stay in the repo, the frozen ID is never reused, and
+any salvageable detection idea ships under a NEW rule ID. If you gated CI
+on these findings, they no longer block at `info` severity; add an explicit
+`severityOverrides` entry in `mjolnir.config.json` to restore blocking.
+
+- **QA-PW-005** (business logic in `page.evaluate()`): in test files,
+  branching inside evaluate is the only way to reach browser state — every
+  measured use was browser-only test instrumentation or the API under test,
+  never leaked app logic. No successor.
+- **QA-PW-103** (missing timeout on navigation): bare `goto()` to the app
+  under test is the universal navigation idiom; per-call budgets are
+  config territory. No successor.
+- **QA-PW-105** (`expect.poll` without timeout): the default poll timeout
+  is a hard bound that raises — the claimed masking harm cannot occur.
+  No successor.
+- **QA-PW-107** (`toBeVisible` on toast/banner/modal): presence +
+  auto-retry semantics is what suites assert; viewport visibility is a
+  different question the suites are not asking. No successor.
+- **QA-PW-108** (`toHaveText` coupling): asserting self-owned markup's
+  exact text is a legitimate strong assertion; "whose markup is this" is
+  not statically decidable. No successor.
+- **QA-PW-112** (testid naming convention): hardcoding kebab-case as _the_
+  convention is the defect — conventions are repo-local. Successor idea
+  (mixed-conventions-within-one-repo check) requires a NEW ID.
+- **QA-PW-114** (legacy element handles): the auto-wait harm needs a
+  timing window; every measured use was immediate reads or deliberate
+  existence checks. No successor.
+- **QA-PW-118** (`networkidle` waits): the flake source is environmental
+  background traffic, absent by construction where the rule fired; harm is
+  not code-detectable. No successor.
+- **QA-PW-119** (module-level state order dependence, was `error`): FPs
+  scatter across ≥5 legitimate infrastructure idioms (per-test teardown
+  harnesses, counters, vi.hoisted fixtures, memoized shared infra) — an
+  error-severity rule at 0 TP / 24 is actively misleading. Successor idea
+  (cross-test write→read dataflow analysis) requires a NEW ID.
+- **QA-PW-120** (missing environment guard): file-level keyword
+  co-occurrence does not imply engine dependence; even the corpus's e2e
+  specs are engine-agnostic. No successor.
+- **QA-PW-145** (no a11y assertions): absence of optional coverage is not
+  a defect finding; the heuristic fires on every UI spec by construction.
+  Successor idea (a11y-coverage reporting) requires a NEW ID.
+- **QA-TQUAL-001** (mock-only verification, 0 TP / 26): spies observe the
+  real unit's output — the mock call IS the observable contract;
+  stand-in vs observer is not statically decidable. No successor.
+- **QA-PY-006** (empty test body `pass`): 18/20 FPs were pytester
+  test-data scripts; genuinely collected empty tests don't occur in real
+  code. No successor.
+- **QA-PY-008** (mock-only verification): boundary mocking and real-output
+  spies are contract testing, not mock theater. No successor.
+- **QA-PY-010** (random/time without freeze): wall-clock reads are the
+  measured subject in timing/throttle tests; freezing would defeat them.
+  No successor.
+- **QA-JV-108 / QA-CS-108** (hardcoded environment URL): FPs split across
+  HAR-replay fixtures, route-mocked origins, and proxy-failure tests — no
+  mechanically discriminable shape (the M-06 header concedes a fake-TLD
+  lookahead fixes zero measured FPs). No successor.
+- **QA-JV-110 / QA-CS-110** (no a11y assertions): same absence-heuristic
+  premise failure as QA-PW-145. Successor idea requires a NEW ID.
+- **QA-JV-111 / QA-CS-111** (blanket route mock): route-API self-tests and
+  fixture setup; no provable exclusion, and framework gating cannot help.
+  Revival on an application-repo corpus would need a NEW ID.
+
+### Changed — Phase 2 retunes (EVIDENCE-BACKED, detectorRevision 2 per plan §07)
+
+Each retune targets the single fixable root cause its measured FP cohort
+shares (full evidence table: docs/RULE-LIFECYCLE.md "Phase 2
+quarantine-cluster triage"). Detection-logic changes bump the rules'
+`detectorRevision` to 2, invalidating the revision-1 measurements (stale →
+provisional → re-measure, plan §07); must-not-fire fixtures now encode the
+measured FP shapes so the retunes are regression-locked in both directions.
+Tier stays quarantine until re-measurement says otherwise.
+
+- **QA-PW-102** (load-wait instead of assertion): the wait now fires only
+  when it is the TERMINAL wait (no `expect`/assert/`expect.poll` follows)
+  and is skipped when consumed by `expect(...).rejects` — the "instead of
+  an assertion" premise, now actually checked. Clears all 20 measured FPs
+  (vite HMR synchronization waits); keeps the must-fire no-assertion shape.
+- **QA-ENV-001** (environment coupling): the fixed-port sub-pattern no
+  longer matches loopback endpoints (`localhost`/`127.x`) — they are the
+  suite's own fixture containers (Azurite / DynamoDB Local / Mongo), the
+  entire measured FP cohort. Dotted hostnames and non-loopback IPv4
+  literals with ports still fire; OS-path, locale, and local-time-getter
+  sub-patterns are unchanged.
+- **QA-PY-003** (no-assertion test): the verification vocabulary gains
+  `pytest.warns` / `pytest.deprecated_call` / `pytest.fail`, and a
+  `test_*` function referenced by name elsewhere in its file (pytester-
+  style test data) is skipped — the collected assertion lives in the
+  parent test.
+- **QA-PY-004** (bare truthiness assert): boolean-predicate calls are
+  skipped — `assert isinstance(x, T)` type guards and
+  `assert s.startswith(...)`-style content predicates are real checks
+  (the measured FP clusters). Bare identifier/attribute asserts still
+  fire. This rule was measured at 45% FP (n=20), not 100%; with the
+  clusters removed the re-measurement is expected to approach the
+  extended band (≤30%), pending delta classification.
+- **QA-PY-007** (raises without match): `pytest.raises(X) as exc_info`
+  followed by an assert/expect on `exc_info.value` is skipped — the
+  message IS verified without `match=` (8 of 13 measured FPs). The
+  genuinely vague raises blocks still fire; remaining FP residue
+  (single-possible-exception blocks) is not statically decidable.
+- **QA-PY-105** (Playwright-Python test without assertions): a called
+  helper whose name asserts (`assert_*`/`expect_*`/`verify_*`/`check_*`)
+  or waits (`wait_for_*`) counts as verification — assertions delegated
+  to imported helpers, the entire measured FP cohort (streamlit e2e).
+- **QA-JV-106 / QA-CS-106 / QA-PY-104** (brittle selectors): no pattern
+  change — the measured 100% FP rows predate Bug Map M-06's removal of
+  the querySelector/QuerySelectorAsync/query_selector patterns, so the
+  measurement described a detector that no longer ships. The §07 fix is
+  the detectorRevision bump to 2 (stale → provisional → re-measure);
+  the surviving xpath=/nth-child/absolute-path patterns stay
+  quarantine-tier until re-measured.
+
+### Changed — §07 loop closed: rev-3 delta re-measurement (corpus rescans + verdict reconciliation)
+
+The revision-2/3 detectors were re-run over the corpus repos that
+produced the retuned rules' original verdicts, and the committed verdict
+corpus was reconciled to the current detectors (the §07 loop the
+capability matrix calls "re-measure"):
+
+- **Delta method:** every committed row whose finding no longer fires
+  was removed (superseded — it described a pre-retune detector); every
+  NEW finding was adjudicated from source context per
+  `tests/corpus/verdicts/README.md` criteria, with a quota sample of 20
+  per rule on the largest surfaces (QA-PY-004/007) and full coverage on
+  the small ones. Net: 157 rows retired, 71 adjudicated rows appended,
+  1423 → 1337 classified verdicts, measured coverage 78 → **72/91**
+  (the six rules whose remaining classified counts fell below n=10 —
+  QA-PW-102, QA-PY-104, QA-PY-105, QA-ENV-001, QA-JV-106, QA-CS-106 —
+  are now unmeasured at their current revisions; their sidecar entries
+  were removed accordingly and they stay quarantine/extended-tier
+  pending re-measurement).
+- **Re-measured envelopes (rev 3):** QA-PY-003 82% FP (n=17; remaining
+  FPs are pytest doc examples and pytester collection fixtures —
+  accepted residue), QA-PY-004 76% (n=21; remaining FPs are
+  pytest's own predicate-call idiom families: fnmatch/samefile/
+  isimportable membership checks and deliberate `assert False`
+  fail-marker DATA), QA-PY-007 79% (n=34; remaining FPs are
+  single-possible-exception API-error contracts where the raised type
+  IS the assertion — not statically decidable). All three stay
+  quarantine-tier with their detectorRevision 3 stamped in the sidecar
+  and FP-AUDIT; their accepted-residue FP causes are documented in
+  docs/RULE-LIFECYCLE.md's triage table.
+- **QA-PW-102 / QA-PY-105: fully cleared** — 0 findings on the
+  previously-20/20-FP corpus slices at revision 2.
+- **QA-ENV-001 (rev 3):** the wave-2 delta (20/20 FP on the rev-2
+  detector) showed no host shape is statically decidable — the
+  fixed-port sub-pattern is dropped entirely; OS-path, locale, and
+  local-time-getter sub-patterns stay (their remaining 6 findings on
+  vite are path-literal test DATA, adjudicated FP and recorded).
+- **§07 comparability guard:** the FP regression governor
+  (`checkFpRegression`) no longer compares measurements across
+  detectorRevision boundaries — a revision mismatch means the old rate
+  does not describe the current detector, so it is neither flagged nor
+  blessed.
+
+### Verification Trust Evolution, Phase 2 — detectionStrategy enum (D6 closed, scan-behavior-neutral)
+
+### Changed — D6 enum migration (plan §12.1; metadata-only)
+
+- **`detectionStrategy` is now the enforced §09.6 enum** (`LEXICAL | AST |
+SEMANTIC | FRAMEWORK | RUNTIME`) instead of free text: `src/rules/rule.ts`
+  types `RuleMeta.detectionStrategy` as the union, and a registry ratchet
+  (tests/rules.registry.spec.ts) fails CI when any rule omits it or carries
+  a non-enum value — the "free text" drift class cannot reintroduce itself.
+  All 91 registry rules were migrated in place. This is a metadata migration,
+  not a detection change: no `run()` body, pattern, or scoping was touched,
+  so scan findings, golden fixtures, and corpus baselines are byte-identical
+  (the Phase 0 classification of this work as BEHAVIOR-NEUTRAL for scan
+  findings, plan §06). Per §11.3 the enum conversion is metadata naming only
+  and does NOT bump any `detectorRevision` — every measurement stays valid.
+- **Legacy nuance preserved, not deleted:** the richer free-text
+  declarations ("regex pattern + inside-string oracle", "parsed YAML +
+  test-command gate", …) moved verbatim into a new optional
+  `detectionNotes` field rendered alongside the enum in the rule docs
+  pages. The capability matrix now renders the declared enum directly in a
+  single "Detection strategy (enum)" column (the provisional
+  "Enum (proposed)" guess column is gone — the declared value IS the
+  enum); `UNCLASSIFIED` renders only for an undeclared value.
+- Mapping applied: regex-over-text/absence-sweep detectors → `LEXICAL`
+  (incl. QA-TQUAL-002's AST-stripped text pattern, which is a text pass
+  over code-only text); ts-morph node-walk detectors (QA-PW-002,
+  QA-PW-005) → `AST`; GitHub-Actions workflow-structure detectors
+  (QA-CI-001/009/010) → `FRAMEWORK`. `SEMANTIC` and `RUNTIME` remain
+  reserved (no rule ships either yet).
+
+### Verification Trust Evolution, Phase 1 exit — dedicated corpora + wave-5 measurement (plan §11.5/§08)
+
+### Added — dedicated corpora (§11.5) and the wave-5 measurement
+
+- **CORPUS 19 → 34 repos** (`tests/corpus/audit.ts`): 14 new real-world
+  repos chosen by evaluating each candidate's unmeasured-rule fire count
+  at HEAD before committing (`vitest-dev-vitest`, `streamlit-streamlit`,
+  `apache-airflow`, `iluwatar-java-design-patterns`,
+  `spectreconsole-spectre-console`, `Humanizr-Humanizer`,
+  `cypress-realworld-app`, `keycloak-keycloak`, `appsmithorg-appsmith`,
+  `getsentry-sentry`, `github-docs`, `vercel-next-js`, `hashicorp-vault`,
+  `nocodb-nocodb`). This is the plan's dedicated-Corpora requirement for
+  §11.5: CI-workflow density for the starved QA-CI-* rules and JV/CS
+  application repos so the QA-JV/QA-CS rules measure on consumer code,
+  not just the Playwright bindings themselves (D5). Candidates whose scan
+  truncated against the budget (`n8n`, `posthog`, `vscode`) were rejected —
+  a partial scan can never be count-locked.
+- **Committed class-B/C fixture corpora (§08):**
+  `tests/corpus/positive-fixtures/` (realistic anti-pattern exhibits that
+  MUST fire — class-B recall evidence, every fire classifies TP) and
+  `tests/corpus/negative-fixtures/` (realistic legitimate code that must
+  NOT fire — class-C precision evidence, any fire is a recorded FP).
+  These give rules whose patterns are rare in the wild a measurement-grade
+  verdict surface, versioned with the verdicts that classify them via
+  `local:` corpus URLs. Excluded from self-scan, vitest, eslint, prettier
+  and the test tsconfig — they are DATA.
+- **`corpus-sample.ts --unmeasured-only`:** the verdict-harvesting loop
+  can sample only rules without a valid measurement, so classification
+  effort goes to the exit gate instead of re-sampling measured rules.
+- **`audit.ts --only=<name>,<name>`:** re-check or re-record a corpus
+  subset (used to re-verify a repo after a transient truncation without
+  rescanning everything); the completeness threshold now applies to the
+  filtered set. Clones use `git -c core.longpaths=true` (scoped, not a
+  global config change) and the per-repo scan budget is 60s → 120s; the
+  audit job gets `NODE_OPTIONS=--max-old-space-size=8192` (sentry's
+  repo-scale parse OOMs the default heap) and a 60-minute timeout.
+- **Measured coverage 43 → 78 of 91** (1423 classified verdicts, 0 blank,
+  0 UNSURE): 35 rules newly measured at n ≥ 10 — QA-TEST-001, QA-TEST-006,
+  QA-TEST-010, QA-TQUAL-002, QA-TQUAL-009, QA-TQUAL-011, QA-CI-001,
+  QA-CI-002, QA-CI-005, QA-CI-008, QA-PW-003, QA-PW-004, QA-PW-104,
+  QA-PW-113, QA-PW-115, QA-PW-117, QA-PW-121, QA-PW-123, QA-PW-140,
+  QA-PW-141, QA-PW-142, QA-PW-144, QA-PY-001, QA-PY-009, QA-PY-011,
+  QA-PY-012, QA-PY-103, QA-PY-105, QA-JV-101, QA-JV-102, QA-JV-109,
+  QA-CS-103, QA-CS-107 (plus the four previously counted). **Phase 1
+  exit gate MET: unmeasured 48 → 13 (≤ 20).**
+- **Explicit tier declarations for all 35 newly measured rules, set from
+  the measured FP band (§11.2: ≤10% core, ≤30% extended, >30%
+  quarantine):** core — QA-PW-003, QA-PW-104, QA-PW-113, QA-PW-117,
+  QA-PW-121, QA-PW-140, QA-PY-001, QA-PY-009, QA-PY-011, QA-PY-103,
+  QA-JV-101, QA-JV-109; extended — QA-TQUAL-011, QA-CI-002, QA-CI-007,
+  QA-PW-141, QA-PW-142, QA-PW-144; quarantine — QA-TEST-001, QA-TEST-006,
+  QA-TEST-010, QA-TQUAL-002, QA-TQUAL-009, QA-PW-004, QA-PW-115,
+  QA-PW-123, QA-PY-012, QA-PY-105, QA-CI-001, QA-CI-005, QA-CI-008,
+  QA-CI-010, QA-CS-103. Every measured entry carries `detectorRevision`
+  in the sidecar. **Evidence-backed, not silent:** the quarantine
+  demotions remove error-severity deductions from default scans, which is
+  the tier policy working as designed — the failing specs were updated
+  with their reasoning inline (gate tests re-anchored on QA-CI-009, the
+  demo repo's CI grew real CI-009 exhibits, the demo/hero assets and
+  `fix` command scan with `--strict`).
+- **Recall-floor + corpus baselines:** 34 count-locked baselines recorded
+  (non-partial scans only); the §20.6 recall floor extends to the
+  expanded registry.
+
+### Changed — scan-behavior fallout of the measured demotions (explained)
+
+- `mjolnir fix` now scans with `--strict`: an auto-fixable rule that is
+  measured into quarantine (QA-TEST-001's `.only` fix) must still be
+  fixable — hiding it would make `fix` a no-op on its own target debt.
+- The demo repo's CI workflow gained genuine CI-009 exhibits (piped and
+  `;`-sequenced test commands) so the demo keeps demonstrating the
+  NEEDS-WORK band now that its QA-TEST-001/QA-CI-001/PW-004 debt is
+  quarantine-capped to info (non-deducting); `docs:demo`/`docs:hero` scan
+  with `--strict` so the committed assets and their drift locks stay in
+  sync with the precision-contract spec.
+
+### Verification Trust Evolution, Phase 1 — measurement infrastructure
+
+### Added — UNSURE adjudication gate + QA-PW-101 measured (plan §11.5)
+
+- **UNSURE ceiling ratchet:** `npm run fp-audit:generate` now fails when the
+  UNSURE backlog grows beyond the committed
+  `tests/corpus/verdicts/unsure-ceiling.json` (the §11.5 mechanism: UNSURE
+  never counts into `n` but always triggers review). The ceiling only moves
+  DOWN via documented adjudication; upward movement requires an explicit
+  `--update` whose diff names every rule that grew. Criteria live in
+  `tests/corpus/verdicts/README.md`.
+- **QA-PW-101 is measured (the D5 "parked on 20 UNSURE" defect resolved):**
+  all 20 UNSURE verdicts were adjudicated by reading the cited sources at
+  repo HEAD — **20 TP, 0 FP, n=20** (next-auth's session-sync sleeps before
+  reading session state; sveltejs/kit's 100 ms request-observation windows
+  that false-pass when a stray refresh starts after the window). The rule
+  now declares `tier: "core"` (0% FP ≤ 10%, n ≥ 10, revision 1, recall
+  floor satisfied: fires in 4 corpus baselines) — the measurement-dependent
+  default would resolve it to core anyway, but measured rules declare
+  their tier explicitly (the D3 Step 2 invariant).
+- **QA-TQUAL-009 UNSURE row resolved → FP** (tanstack-query angular
+  inject-query:529 — a deliberate `void` fire-and-forget whose assertions
+  still fail the run via vitest's unhandled-rejection handling; the
+  FALSE-GREEN diagnosis does not hold). n=2 — below the measurement bar,
+  still unmeasured/PROVISIONAL.
+- **Coverage:** 42/91 → **43/91** measured rules; UNSURE backlog 21 → **0**.
+
+### Changed — D3 two-step tier fix (plan §11.2; scan-behavior-neutral)
+
+- **Step 1 (explicit tiers):** every rule now declares its `tier`
+  explicitly — the 42 formerly implicit-core rules received an explicit
+  `tier: "core"` matching their effective tier. Byte-identical scan
+  behavior (the declared value equaled the omitted-tier default);
+  generated docs updated with an explained diff (the matrix's
+  "explicit tier declarations" line moved 49/91 → 91/91).
+- **Step 2 (measurement-dependent default):** the omitted-tier default is
+  no longer unconditionally core. A rule that omits `tier` resolves via
+  `effectiveTier` (new `src/rules/measurement.ts`): **core** only with a
+  valid corpus measurement (n ≥ 10 at a matching `detectorRevision`),
+  otherwise **extended**, displayed with the new **PROVISIONAL** status.
+  PROVISIONAL is a display status (`tier extended/unmeasured`), not a
+  tier value — no schema churn. Scan findings are byte-identical: the
+  pipeline enforces only `quarantine` (severity/info + E0 caps, --strict
+  filter), and every quarantine rule declares its tier explicitly;
+  overlap-dedup consumes declared tiers only, and no demoted rule
+  participates in an `overlapWith` tier comparison that could change a
+  survivor set (all overlap groups are single-declarer/single-target).
+- **The D3 demotion (38 rules drop from effective core to
+  extended/PROVISIONAL):** QA-TEST-001, QA-TEST-006, QA-TEST-010,
+  QA-TQUAL-002, QA-TQUAL-009, QA-TQUAL-011, QA-PW-003, QA-PW-004,
+  QA-PW-101, QA-PW-104, QA-PW-113, QA-PW-115, QA-PW-116, QA-PW-117,
+  QA-PW-121, QA-PW-123, QA-PW-124, QA-PW-125, QA-PW-140, QA-PW-141,
+  QA-PW-142, QA-PW-144, QA-CI-001, QA-CI-002, QA-CI-005, QA-CI-007,
+  QA-CI-008, QA-CI-009, QA-CI-010, QA-PY-001, QA-PY-009, QA-PY-011,
+  QA-PY-012, QA-PY-101, QA-PY-103, QA-PY-105, QA-PY-106, QA-PY-107.
+  The 4 measured implicit-core rules (QA-PW-002, QA-PY-002, QA-JV-105,
+  QA-CS-101) keep core. Unmeasured-effective-core count: 38 → **0**.
+  **Suite-invalidating callout (plan §11.2):** QA-TEST-001 (focused test
+  committed, `suiteInvalidating: true`) is among the demoted rules — its
+  findings still void the suite's pass claim and still gate CI at
+  severity=error; only its tier/statistics status changed. QA-PY-001 is
+  the suite-invalidating Python sibling (same class).
+- **Registry ratchet (plan §20.3, enforced in code):** new
+  `tests/registry-ratchet.spec.ts` fails CI on ANY unmeasured rule in
+  effective core (`tier core ⇒ valid MEASURED_FP with matching
+detectorRevision, FP ≤ 10%, n ≥ 10`), on any detectorRevision mismatch
+  (§20.5: stale ⇒ provisional), and — recall floor (§20.6) — on a core
+  rule that fires nowhere in the corpus baselines. §20.1 evidence-state
+  monotonicity: the measured ratio may only improve without a
+  machine-detectable `MEASUREMENT-EXCEPTION` marker in this CHANGELOG.
+- **`mjolnir doctor`**: `MAX_UNMEASURED_CORE` lowered 40 → **0** (Phase 1
+  exit gate: 0 unmeasured in effective core, now enforced); tier checks
+  consume `effectiveTier` + stale-measurement logic. Display surfaces
+  (`mjolnir explain`, `mjolnir rules`, generated rule docs, capability
+  matrix) render the PROVISIONAL status honestly.
 
-## [Unreleased]
+### Verification Trust Evolution, Phase 0 + Phase 1 prep
+
+### Added — Rule Capability Matrix (Phase 0)
+
+- **Capability Matrix v0** (`docs/RULE-CAPABILITY-MATRIX.md` +
+  `docs/RULE-CAPABILITY-MATRIX.json`): generated per-rule inventory of all
+  91 rules from the registry + `MEASURED_FP` + the verdict corpus —
+  category, languages, frameworks, declared detection strategy with a
+  provisional enum mapping (Phase 0 contract proposal), semantic depth,
+  measured flag, FP rate with sample size, corpus size and diversity.
+  Unknown fields render as `UNCLASSIFIED` — visible gaps are the
+  deliverable. Regenerated with `npm run docs:capability`; drift-locked by
+  `tests/capability-matrix.spec.ts` and the generated-docs-drift CI job.
+- **Declared-vs-measured cross-check report** (ledger class D9): the
+  matrix names every measured rule whose declared tier violates the FP
+  ceilings (core > 10%, extended > 30%) and produces the D3 demotion list
+  (38 unmeasured rules currently in effective core — Phase 1 input).
+  Current run: **0 D9 mismatches** (all measured > 30% FP rules already
+  declare `quarantine`).
+- **Defect ledger recorded** into the matrix metadata (plan §02, D1–D8
+  with owning phase per defect).
+
+### Added — detectorRevision scaffold (Phase 1 prep)
+
+- **`tests/corpus/detector-revisions.json` sidecar**: hand-maintained,
+  one entry per measured rule (all at revision 1 today), diffable.
+  `MEASURED_FP` entries now carry `detectorRevision` stamped from the
+  sidecar by `fp-audit:generate`; `docs/FP-AUDIT.md` gains a
+  `detectorRev` column. Measurement inheritance law (§07): a measurement
+  belongs to a specific detector implementation, not merely to a rule
+  ID. Drift lock extended in `tests/measured-fp-generated.spec.ts`
+  (sidecar covers exactly the measured set; revisions are positive
+  integers matching the sidecar).
+
+### Fixed — packaging: offline grammar loading (Phase 0.5 spike, D2)
+
+- **`tree-sitter-wasms` and `web-tree-sitter` moved to `dependencies`**
+  (web-tree-sitter keeps its exact `0.25.6` pin — 0.26.x cannot load the
+  prebuilt grammar files). The published CLI's dependency tree now
+  carries the tree-sitter Java/C# grammars, so `npm install mjolnir-qa`
+  can load them offline once the Phase 0.5 parse-stage wiring consumes
+  them. Removed the misleading `!dist/**/*.wasm` files exclusion (the
+  grammars ship via the dependency, not the bundle).
+- **Pack-smoke regression test**: `tests/package-smoke.spec.ts` asserts
+  the packed package declares both as runtime dependencies and that the
+  java/c_sharp grammars resolve inside the installed dependency tree.
+
+### Fixed — adapter header claims (D4)
+
+- `src/adapters/java.ts` no longer claims to be a "Second tree-sitter
+  consumer" — it is a regex-layer adapter; the tree-sitter-java grammar
+  and the async `parseJavaAst` seam exist but are not wired into the
+  synchronous scan (D1). Same honest correction for
+  `src/adapters/csharp.ts` and `src/engine/adapter.ts` (whose header
+  still claimed tree-sitter "arrives in R2 with Python").
+
+### Verification Trust Evolution, Phase 0.5 — async parse stage (D1)
+
+### Changed — parse stage wired into the scan pipeline (D1 closed, BEHAVIOR-NEUTRAL)
+
+- **Async parse stage between discovery and rule execution** (plan §10.1):
+  `runScan`'s per-file loop now awaits an optional `LanguageAdapter.parseAst`
+  hook before running rules. `runRules` and every rule stay synchronous and
+  consume the tree via `ParsedFile.ast` — the engine is NOT async end-to-end,
+  only the one inherently-async seam (WASM grammar load) is. `main()` and the
+  script entry points await the returned promise.
+- **Java and C# adapters implement `parseAst`** backed by the previously-dead
+  `parseJavaAst`/`parseCSharpAst` (defect D1: tree-sitter AST was built and
+  tested in Sprint 8 but never consumed). Parse failure or a missing grammar
+  resolves `undefined` and rules fall back to the regex path — never fatal.
+  No rule consumes the AST yet (Phase 3 wires specific JV/CS rules), so scan
+  findings are byte-identical: golden lock, corpus baselines, and generated
+  assets all unchanged (BEHAVIOR-NEUTRAL mode, plan §06).
+- **Parser lifecycle management** (plan §10.3): one memoized `Parser` per
+  grammar, bounded by a fixed-size parse-slot semaphore
+  (`MAX_CONCURRENT_PARSES`); every per-file tree is released via
+  `ParsedAst.dispose()` (`tree.delete()`) in a `finally`-equivalent position
+  that runs on normal completion, rule crash, per-file budget expiry, and
+  adapter throw; `releaseTreeSitterResources()` tears down the memoized
+  parsers after each scan (library-consumer hygiene). No leak path depends
+  on rules completing successfully.
+- **Call-graph consequences**: `computeImpact` (and therefore
+  `runImpactCommand`) is async; `scripts/corpus-sample.ts`,
+  `scripts/generate-readme-demo.ts`, and `scripts/generate-readme-hero.ts`
+  await `runScan`; the full test suite (~40 call sites across 25 spec files)
+  awaits the now-async commands. No behavior change anywhere.
+- **Verification evidence** (exit gate §10): golden lock byte-identical
+  (3/3), corpus count-lock unaffected (no JV/CS rule reads the AST),
+  capability-matrix + FP-AUDIT generated docs byte-identical, pack smoke
+  re-proven manually: `npm pack` → clean `npm install` → offline grammar
+  load + parse of real Java and C# source → offline `mjolnir` scan of a
+  Java fixture detecting QA-JV-102. `web-tree-sitter` stays pinned to
+  exactly `0.25.6` (§10.5, documented 0.26.x breakage).
+
+### Added — score instrument redesign (hammer states)
+
+- **ScoreState model** (`src/reporter/score-state.ts`): one pure source of
+  truth for band / verdict / color / headline per score — critical 0–49,
+  warning 50–79, trusted 80–99, forged 100. `verdictFor`, the terminal
+  gauge and the badge all delegate to it.
+- **The hammer is now the score instrument** (terminal): a state-colored
+  hammer block renders above the WORTHINESS line — cracked (0–49),
+  strained with partial runes (50–79), charged with energy arcs (80–99),
+  halo + lightning at 100. A plain-text caption (`[CRACKED]` /
+  `[STRAINED]` / `[CHARGED]` / `[FORGED]`) carries the state without
+  color; ASCII fallback included.
+- **Trusted is aurora-cyan, forged is white-gold** on every surface
+  (terminal palette, web tokens, brand README). Green is no longer a
+  score color — it survives for non-score success contexts only.
+- **Findings render as cards** (terminal): Problem → Impact → Fix →
+  Verify with the evidence tag and measured FP rate beside the title;
+  rules with >3 findings collapse under one "same fix applies" header;
+  non-verbose shows 10 cards with an overflow line, `--verbose` shows
+  everything.
+- **FORGED block at 100** replaces the bare FLAWLESS VICTORY line in
+  unicode mode (trophy retained inside; the `*** FLAWLESS VICTORY ***`
+  ASCII contract string is preserved).
+- **PR comments show score drift** (`Score: 72/100 (+5 since baseline
+<sha>)`) using the new additive `score` field in the baseline JSON, and
+  carry per-finding evidence tags.
+- **Badge thresholds aligned** with the reporter: ≥80 / ≥50 / 100
+  (was ≥90 / ≥75 / ≥50), colors `red` / `yellow` / `important` /
+  `success`; the message at 100 reads `100/100 · forged`.
+
+### Fixed — security & detection-regression audit (`.planning/AUDIT-2026-08-30-QA.md`)
+
+- **QA-1 (P0, detection regression):** QA-TEST-003's M0-#4 header rewrite
+  required a space after `async` — `async()=>{` bodies silently stopped
+  matching. `\s*` restores them without reopening the bare-`return` exemption.
+- **QA-2 (P0, detection regression):** single-walk discovery applied the
+  UNION of every adapter's `dirSkips`, so Python's `env` (= virtualenv)
+  and Java's `build` (= Gradle output) hid directories from every other
+  language — real TS test dirs (e.g. withastro/astro's `test/units/env/`)
+  silently vanished from scans. `dirSkips` are now applied per owning
+  language at file level.
+- **QA-3 (P1):** `isInsideEmbeddedCode` truncated the masked run at the
+  first space inside a literal, so embedded test-data
+  (`'test(" foo", function () {})'`) was classified as live code and
+  fired 6 FP findings on eslint-plugin-playwright's ruleTester tables.
+- **QA-4 (P1):** config `exclude` was never validated —
+  `exclude: [1, {}, null]` crashed the scan (exit 20). Non-string entries
+  are now a fixable usage error (exit 10), with a defense-in-depth filter
+  in pattern compilation.
+- **QA-5 (P1):** an unparseable suppression `expires` value silently
+  degraded to "expired" via NaN comparisons; it is now a fixable
+  validation error at load time.
+- **QA-6 (P1, documented-behavior fix):** the README's 90-day suppression
+  policy was only applied at write time by the `ignore` command —
+  hand-written entries without `expires` stayed active forever. The
+  default is now enforced at enforcement time, anchored at the config
+  file's mtime; `mjolnir suppressions` labels the default explicitly.
+- **QA-7 (P1):** plugin reserved-prefix spoof rejection was
+  case-sensitive — `"qa-test-001"` walked straight past it. Matching is
+  now case-insensitive.
+- **QA-8 (P2):** suppression `files` globs written with Windows
+  backslashes could never match the normalized finding paths; both sides
+  are normalized before matching.
+- **QA-9 (P2, defense in depth):** option-shaped `--base` values are
+  refused at the git layer even when a programmatic caller bypasses
+  `parseArgs` (option-injection → `--upload-pack=` command execution).
+- **QA-10 (P2):** finding metadata (hostile filenames, plugin messages)
+  reached the terminal and markdown PR comments unsanitized — ANSI
+  escapes and control characters are now stripped, and the PR comment
+  escapes markdown-significant characters.
+- **QA-11 (P2):** the JUnit XML scan was quadratic on unclosed
+  `<testcase` floods (minutes of CPU for a 20 MB hostile report); the
+  scanner is now linear with a bounded-time regression test.
+- **QA-12 (P2):** `loadBaseline`/`loadStats` are now total over arbitrary
+  JSON — `null` finding entries used to crash `diff`, and hostile stats
+  shapes leaked string junk into totals.
+- **QA-13 (P3, defense in depth):** SARIF artifact URIs normalize
+  backslashes — `encodeURI` leaves `\` literal, which is not a valid
+  RFC 3986 uri-reference character.
+- **QA-15 (P1):** the code-text mask iterated code points instead of
+  UTF-16 code units — one emoji made the mask shorter than the text and
+  silently disabled masking for the whole file.
+- **QA-16 (P2):** files larger than the 1 MiB discovery cap were dropped
+  silently; the skip is now counted and flagged (`file-size`), keeping
+  the scan honest about what it did not read.
+
+### Changed — corpus baseline
+
+- Regenerated after review: M0-#2's QA-TEST-004 duplicate-count fix
+  (−5 on withastro-astro — every removed finding was a same-position
+  duplicate), M0-#4's function-body/return-narrowing detections (net
+  +1/+1/+8/+6 across vite/svelte-kit/tanstack-query/eslint-plugin-playwright,
+  spot-reviewed on real code), QA-3/QA-6 masking suppressions
+  (embedded-code test data no longer fires), and QA-15's mask-alignment
+  fix (QA-PW-105/108 match against `codeText`, whose offsets were
+  misaligned after any astral character — +27/+22/+1 restored matches on
+  vite/astro/tanstack-query).
+
+### Added — Open-Beta E2E test plan (Tier 1–5, ~1,700 new test assertions)
+
+- **Coverage ratchet:** statements/branches/functions/lines at literal
+  100% per file (`perFile: true`). Every branch arm, catch path, and
+  fallback in `src/**` is exercised or provably dead — ~30 dead guards
+  (unreachable `?? ""` fallbacks, unreachable-`if` arms behind
+  `noUncheckedIndexedAccess`, a dead proof-abort block in `fix`) were
+  removed or exported as testable pure helpers with the golden lock,
+  self-scan and determinism tripwires green throughout.
+- **E2E journeys** (`tests/e2e/`): nine spec files run the built
+  `dist/cli.mjs` as a real child process — tarball first-run (the path
+  that historically caught macOS-only CLI breakage), CI PR flow with
+  changed-scope attribution against real git fixtures, the baseline →
+  diff → stats loop, the fix flow, the forensics flow, explain/rules,
+  create-rule onboarding, the config journey (gate/severityOverrides/
+  ignore/expiry honored end-to-end), and a full exit-code contract sweep
+  across every documented command.
+- **Precision & accuracy:** every scorer number verified against the
+  documented benchmark (deductions 8/3/1, E2/E1/E0, honesty cap 99,
+  error ceiling 95, suite-invalidating ceiling 49, smoothing) with
+  fast-check property invariants (≥1,000 iterations: score bounds,
+  monotonicity, order-symmetry), Selector Health exact score vectors,
+  hand-computed forensics math, terminal-footer/JSON deduction
+  consistency, Mermaid well-formedness, and a three-verdict-band proof
+  (WORTHY / NEEDS WORK / UNWORTHY each reached for its stated reason).
+- **Regression & integration:** adapter→reporter matrix (one finding
+  asserted on terminal, JSON, SARIF, and Mermaid), plugin flow
+  integration (valid plugin + reserved-prefix rejection), cross-file
+  analysis, monorepo containment (a workspace scan never reports
+  siblings), a mutation guard (each sampled rule's finding drops to zero
+  when the offending line is removed), and baseline forward-compat.
+- **Nightly stress workflow** (`.github/workflows/stress.yml`, never
+  PR-blocking): 10k-file mixed-language synthetic repo under a 120 s
+  budget with no partial degradation, pathological trees (200-deep
+  nesting, 10k-char lines, unicode filenames, LF/CRLF/BOM mixes,
+  junctions, malformed specs), a 20-run byte-identical + flat-RSS soak,
+  the networked registry-install smoke (previously skipped locally), and
+  4 concurrent scans of one read-only target. Fixture generators live in
+  `tests/stress/`.
+- **Stability:** the 1 MB masking budget now takes the median of 3 runs
+  with a 15 s ceiling (coverage instrumentation slows string-heavy loops
+  ~3x and there is no in-worker marker to detect it; the ratio-based
+  linearity test remains the true non-quadratic guard), and the
+  scale-benchmark's 3k-file budget moved 20 s → 25 s with the same
+  rationale. The vitest global-setup builds `dist/` once before any
+  worker starts — parallel-file builds used to wipe dist mid-suite
+  (tsdown cleans `outDir`) and fail unrelated E2E spawns with
+  module-not-found.
+- **Fixed (flagged by the new E2E sweep):** `doctor --bogus` ignored the
+  unknown flag and scanned the CWD as a surprise full run; it now prints
+  usage and exits 10, matching the flag-error parity of every other
+  subcommand (unit + E2E regression tests added).
+- **Soak drift artifacts** moved from `coverage/` (a CI-generated dir the
+  tool's own QA-CI-005 rule correctly flags as "consumed but never
+  generated" in stress.yml) to a dedicated `soak-drift/` dir.
+
+## [0.5.0] — 2026-08-29
+
+### Added — measurement is now visible at the point of use
+
+- Only 15 of 91 rules carry a false-positive rate measured against real OSS
+  code; that fact previously lived only in `docs/FP-AUDIT.md` and
+  `mjolnir doctor`. Now surfaced everywhere a user looks:
+  - The scan footer reports how many of the rules that _fired_ are measured.
+  - `mjolnir rules --unmeasured` / `--measured` filter the catalog; a new
+    "FP (measured)" column in `rules --md`; a "Measured FP rate" row on every
+    `docs/rules/` page and in `mjolnir explain`.
+  - JSON findings carry `measuredFpRate` and `measuredFpN` (additive —
+    `schemaVersion` is still 1).
+- `src/rules/measured-fp.generated.ts` bakes the rates into the shipped
+  package (the raw verdicts are not packed); regenerated by
+  `npm run fp-audit:generate`, drift-locked by a test, and now the single
+  source `mjolnir doctor` reads.
+- Scoring is unchanged — this is visibility only.
+- **Corpus expanded 6 → 13 repos** so the previously-silent rule families
+  (QA-TEST, QA-TQUAL, most QA-PW, QA-CI-001) fire on real consumer code:
+  added `next-auth`, `vite`, `sveltekit`, `astro`, `TanStack/query`,
+  `eslint-plugin-playwright`, `playwright-pytest`. `corpus:sample` and
+  `corpus:regression` now scan with `--strict` (quarantine rules were
+  invisible to both before). `docs/FP-AUDIT.md` is 15/91 (down from 19 —
+  see the dispatch fix below, which retired the leaked cross-language
+  verdicts for QA-PW-101/112 and QA-TEST-004/QA-ENV-001 on Java/Python
+  repos); the ~250 new corpus findings are queued for classification in
+  `tests/corpus/verdicts/`, not counted until read.
+
+### Changed — help and README lead with the one command
+
+- `mjolnir --help` and the README quickstart now open with
+  `mjolnir --scope changed` as _the_ product, and group the other subcommands
+  into Everyday / When-something's-flaky / Occasional instead of a flat list
+  of 16 equals. A one-line first-run hint appears after a bare full-repo scan
+  with no config. No subcommand removed or renamed.
+
+### Fixed (rule-bug-hunt wave)
+
+- **Cross-language dispatch leak**: `appliesTo: "test-files"` mapped to all
+  four language adapters, so the 42 TypeScript/Playwright-only rules that
+  use it (QA-PW-\*, QA-TEST-\*, QA-TQUAL-\*) ran against `.py`, `.java` and
+  `.cs` files too. On the corpus this produced ~140 false positives on
+  `microsoft/playwright` Java bindings alone (QA-PW-101, QA-PW-112,
+  QA-TEST-004, QA-ENV-001) and inflated several baselines. `legacyAppliesTo`
+  now maps `"test-files"` to `["typescript"]` only; cross-language coverage
+  is the QA-PY/QA-JV/QA-CS families' job. Regression test added.
+- **QA-PW-103** (missing timeout): no longer fires on assertion strings that
+  contain Playwright code as _test data_ (`code: "await page.goto('/x')"` in
+  playwright-mcp) — guarded by `isInsideEmbeddedCode`.
+- **QA-TEST-004** (hard sleep): dropped the bare `sleep(N)` pattern and now
+  requires `await` and a non-zero argument. `sleep(10).then(...)` and
+  `queryFn: () => sleep(10)` are mock-latency, not test-body pauses — this
+  cut TanStack Query's count from 1648 to 157.
+- **QA-PW-002** (unawaited assertion): matches only Playwright's 31
+  web-first async matchers instead of any `to*` name, so
+  `expect(res.status()).toBe(200)` on a variable named `page` is no longer
+  flagged.
+- **QA-TQUAL-009** (assertion in unawaited promise chain): the `.then()`
+  callback body is now paren-matched instead of grabbing the next `{`, so a
+  sibling `.then(res => res.text())` inside an awaited `Promise.all` no
+  longer reaches into an unrelated block; the await/return check also sees
+  an `await` sitting one line above the `.then(`.
+
+### Fixed (rule-bug-hunt wave 2 — CI + Python + order-dependence)
+
+- **QA-CI-002** (`|| true` swallows exit code): only fires now when the
+  swallowed command is a verification gate. `docker compose down || true`,
+  `pkill … || true`, `rm -rf … || true` are ordinary teardown — flagging
+  them as FALSE-GREEN was wrong. Gate detection is now shared with QA-CI-001
+  (`src/rules/ci/verification-gate.ts`).
+- **QA-CI-009** (exit code not propagated): `playwright` alone was treated as
+  a test command, so `npx playwright install --with-deps; npx playwright
+test` (install first, test last) was flagged even though the test's exit
+  code IS the step's. Now requires `playwright test`, and skips any
+  `setup; <test>` sequence where the test command runs last.
+- **QA-CI-010** (tests skipped where they must block): the condition matcher
+  used `[!=]=`, which also matched `==` — so `if: github.event_name ==
+'pull_request'` (run **only** on PRs) was flagged as _skipping_ tests on
+  PRs. Now `!=` only, plus positive matches on `== 'push'` / `'schedule'` /
+  `'workflow_dispatch'`.
+- **QA-PY-012** (tautological assertion): the patterns lacked the `g` flag,
+  so `regex.exec()` never advanced — only the **first** `assert True` and the
+  first `assert x == x` in a file were reported; every later one was missed
+  (and the loop spun to its 1000-iteration guard each time). Fixed; corpus
+  count rose 4 → 5 as the previously-missed assertions surfaced.
+- **QA-PY-009** (commented-out test): `# main()` in a comment ("call main()
+  here") was flagged as a disabled test. The `main(` pattern now requires
+  the `pytest.main` namespace.
+- **QA-PW-119** (order dependence): three bugs. (1) `let [a, b] = …` /
+  `let { page } = …` destructuring was split on `,` into junk names like
+  `[a` that were interpolated into `new RegExp(…)` — a crash risk; it is now
+  skipped. (2) typed module-level `let x: Foo<T> = …` was invisible to the
+  declaration regex (the `<>` broke it) and is now detected. (3) a `before*`
+  hook with a destructured param — `beforeEach(async ({ page }) => {` — had
+  its body located at the _param_ brace, so assignments in the hook body
+  were treated as in-test and flagged; corpus count dropped 45 → 11 on the
+  worst-affected repo.
+- **QA-PW-116** (storageState without expiry): the canonical Playwright auth
+  pattern — a `setup` project / `*.setup.ts` / `globalSetup` regenerating
+  the state each run — is now recognised as a freshness mechanism, not
+  flagged.
 
 ### Fixed (adversarial-audit hardening wave)
 
@@ -54,7 +1059,7 @@ once shipped, so this file is the record of what changed between versions.
 - **JSON/SARIF truncation removed**: results were silently capped at 50
   findings, including machine consumers. The full finding set is now in
   JSON/SARIF; only terminal display is capped (with an honest count).
-- **`qa-doctor fix` path containment**: plugin-supplied finding paths can
+- **`mjolnir fix` path containment**: plugin-supplied finding paths can
   no longer write outside the scan root (`../` traversal refused).
 - **Symlinks are no longer followed** during test-file discovery in any
   adapter — prevents scanning outside the repo and link cycles.
@@ -62,6 +1067,80 @@ once shipped, so this file is the record of what changed between versions.
   (`QA-JV`, `QA-CS`, `QA-PLUGIN`).
 - `doctor:playwright` bad-usage exit code unified to 10 (was 2).
 
+### Fixed — false positives confirmed by reading source, each locked by a fixture
+
+Every entry below was verified by opening the cited file and reading the
+surrounding code, not inferred from the rule's description. Each is now locked
+by a `must-not-fire` fixture so the class cannot return silently.
+
+- **QA-TQUAL-011** matched a test identifier anywhere inside a comment block,
+  so any JSDoc header containing the sequence `test (` fired. Confirmed on
+  `tests/package-smoke.spec.ts:2` — `* Package publish integrity smoke test
+(Test Hardening Plan, P0 #2).` Now requires the identifier to be the first
+  token on the commented line. Locks:
+  `tests/fixtures/QA-TQUAL-011/must-not-fire/prose-mentioning-test.spec.ts`.
+- **QA-PW-004** fired on selectors passed as arguments to the function under
+  test. Confirmed on `tests/selector-health.spec.ts:33` —
+  `expect(classifyLocator("page.locator('xpath=//div')")).toBe("xpath")`. The
+  rule must read raw text to see selector content, so masking cannot fix it;
+  it now consults `codeText` as an oracle about the match position instead.
+  Locks: `tests/fixtures/QA-PW-004/must-not-fire/selector-as-argument.spec.ts`.
+- **QA-ENV-001**, **QA-PW-123**, **QA-PW-142** fired on code samples embedded
+  in strings as test data. Confirmed on
+  `tests/rule-sprint8-java-csharp.spec.ts` lines 147, 157, 603 — e.g.
+  `text: 'page.navigate("http://localhost:3000/checkout")'`. Now skipped when
+  the enclosing string literal holds both a nested quote and call syntax.
+  Locks: `tests/fixtures/QA-ENV-001/must-not-fire/code-as-test-data.spec.ts`.
+- **QA-CI-001** fired on `continue-on-error` regardless of what the step did.
+  Confirmed on this repo's own workflows: `ci.yml:48` (badge artifact
+  generation) and `mjolnir.yml:35` (advisory diff, which carries a comment
+  explaining that exit 1 is expected there). Now gated on an allowlist of
+  verification commands. Locks:
+  `tests/fixtures/QA-CI-001/must-not-fire/reporting-steps.yml`.
+
+### Fixed — true positive acted on
+
+- **QA-TQUAL-001** on `tests/adapters.spec.ts:110` was correct. The test was
+  named "counts skipped files on stat failure", its own comment admitted it
+  could not simulate a stat failure, it asserted the skip callback was _not_
+  called, and it never asserted on `ctx.testFiles` — the actual output of
+  `discoverTestFiles`. The test was rewritten to assert on real output and
+  renamed to match what it verifies. The rule was left unchanged.
+
+### Changed — BREAKING: scoring
+
+- **Normalization denominator is now test declarations, not test files.**
+  File count was gameable: adding empty spec files raised the score without
+  adding verification.
+- **`SMOOTHING_C` is 1 (Laplace), was 5.** At 5 it tripled the denominator of a
+  two-declaration repo, diluting real density away.
+- **Findings may declare `suiteInvalidating: true`**, capping the score at 49
+  (UNWORTHY) regardless of exposure. Density can express how much of a suite is
+  questionable; it cannot express whether the suite ran at all. Applied to
+  QA-TEST-001 and QA-PY-001. Deliberately not applied to QA-PW-003, which
+  detects both `test.only()` and `page.pause()` — the flag is per-rule.
+- **A score of 100 now requires zero deductions.** Normalization could
+  previously round a real finding up to a perfect score.
+- `NORMALIZATION_K` remains **unfitted**. See `docs/SCORING.md`.
+
+### Changed — QA-CI-001 severity
+
+- Step-level `continue-on-error` findings are now `error`, previously `warning`.
+  The self-scan gate filters on `severity === "error"`, so a warning could never
+  fail CI — which is how `continue-on-error` stayed live in this repo's own
+  workflows while the tool reported zero errors.
+- Title broadened to "continue-on-error masks a failing verification gate",
+  accurate to the allowlist now used.
+
+### Removed
+
+- **All 49 FABRICATED verdicts removed.** They had been produced by reasoning
+  about what each rule's description implied rather than by reading the source
+  at the cited file and line — fabricated evidence with a real-looking
+  provenance, inside the mechanism built to prevent exactly that.
+  Subsequently, 381 verdicts were classified from real source reading via
+  corpus:sample (see docs/FP-AUDIT.md).
+
 ### Changed (adversarial-audit hardening wave)
 
 - QA-TQUAL-002 and QA-PW-004 detection now runs on a comment-stripped /
@@ -79,7 +1158,7 @@ once shipped, so this file is the record of what changed between versions.
 - Upgrade-Plan-v3 Phase 0.1: expanded the false-positive corpus with two
   additional Python repos (`pytest-dev/pytest`, `psf/requests`) so all
   QA-PY-001..012 rules are exercised against real code via
-  `npm run corpus:audit`.
+  `npm run corpus:regression`.
 - Added `CHANGELOG.md` (this file) per Upgrade-Plan-v3 critical item #3:
   user-visible rule behavior changes get a first-class entry from now on.
 - Upgrade-Plan-v3 Phase 1: five new TypeScript/Playwright rules —
@@ -93,24 +1172,20 @@ once shipped, so this file is the record of what changed between versions.
   QA-PY-104 (brittle selectors), QA-PY-105 (UI-driving test without
   assertions), QA-PY-106 (shared page/context across tests), QA-PY-107
   (`networkidle` wait), QA-PY-108 (hardcoded environment URLs).
-
 - Upgrade-Plan-v3 Phase 6: Plugin API — declare third-party rule packages
-  in `qa-doctor.config.json` (`"plugins": [...]`). Security model: no
+  in `mjolnir.config.json` (`"plugins": [...]`). Security model: no
   sandbox (same trust as ESLint/Vitest plugins); reserved core rule-ID
   prefixes rejected; load failures degrade honestly as QA-PLUGIN-000
   warnings without affecting exit codes. Plus cross-file duplicate-test-name
   detection (`src/engine/cross-file.ts`).
-
-- Upgrade-Plan-v3 Phase 0.2: new `@qa-doctor/playwright-reporter` package
+- Upgrade-Plan-v3 Phase 0.2: new `mjolnir-qa-playwright-reporter` package
   (`packages/playwright-reporter/`) — official Playwright JSON reporter
-  wrapper for QA Doctor's forensics pipeline; default output
-  `qa-doctor.report.json` is the CLI's auto-discovery convention.
-
+  wrapper for Mjölnir's forensics pipeline; default output
+  `mjolnir.report.json` is the CLI's auto-discovery convention.
 - Upgrade-Plan-v3 Phase 3: ts-morph AST precision layer behind the `ast`
   seam (`src/engine/ts-ast.ts`). QA-PW-002 and QA-PW-005 migrated from
   regex to syntax-tree detection (legacy regex kept as fallback). No
   scoring changes — golden lock byte-identical.
-
 - Upgrade-Plan-v3 Phases 4+5: new language adapters — Java
   (`src/adapters/java.ts`, rules QA-JV-101..105) and C#/.NET
   (`src/adapters/csharp.ts`, rules QA-CS-101..104). Core Playwright rule
@@ -118,6 +1193,102 @@ once shipped, so this file is the record of what changed between versions.
   shared browser state, wait-timeout misuse. Rule-ID registry now accepts
   QA-JV/QA-CS families.
 
+### Changed — rule tiers assigned from measured FP rates
+
+- Every rule with a measured rate in `docs/FP-AUDIT.md` now carries the tier
+  its rate warrants (`core` ≤ 10% FP · `extended` ≤ 30% · `quarantine` above
+  that or unmeasured). Demoted to quarantine: QA-CS-102, QA-CS-106, QA-CS-108,
+  QA-CS-111, QA-ENV-001, QA-JV-103, QA-JV-106, QA-JV-108, QA-JV-111,
+  QA-PY-004, QA-PY-006, QA-PY-007. Set to extended: QA-CS-105, QA-TEST-004.
+  Promoted to core: QA-CS-101 (0% FP, n=20), QA-JV-105 (10% FP, n=20).
+  Quarantined rules still ship and are still documented — they are opt-in via
+  `--strict` rather than shaping the default report.
+- `mjolnir rules` (`--json` and `--md`) now exposes each rule's `tier`, and
+  every generated page under `docs/rules/` shows it in the metadata table.
+
+### Fixed — documentation claims a `grep` disproved
+
+- Every generated rule page told the reader to reproduce corpus counts with
+  a `corpus:audit` script. That script had been renamed to `corpus:regression` in
+  the Tempering plan and the generator string was never updated — the command
+  printed on 91 published pages did not exist. Same dead name in
+  `docs/PUBLISHING.md`
+  (`corpus:audit:update` → `corpus:regression:update`).
+- **`docs/FP-AUDIT.md` under-reported the rule base as 84 rules when the
+  registry holds 91.** The coverage denominator was built by grepping source
+  for `id: "QA-…"`, which silently missed the seven rules that the Phase 6
+  families declare as positional factory arguments — QA-CS-106/110/111,
+  QA-JV-106/110/111 and QA-PY-104, every one of them Java/C#/Python. The
+  honesty document was quietly shrinking the newest adapters' coverage. The
+  generator now imports the registry directly (and is TypeScript, so the
+  `.d.mts` shim is gone); a regression test locks the denominator to
+  `RULES.length`.
+- `docs/README.md` described a `docs/plans/` directory that no longer exists
+  and called a completed plan "current work".
+- Residual `qa-doctor` naming removed from user-facing CLI output
+  (`mjolnir explain`, `mjolnir stats`), from comments that contradicted the
+  code they described (`baseline.ts` cited `.qa-doctor/` while writing
+  `.mjolnir/`), and from this changelog's own unreleased section.
+- The Playwright reporter package is renamed throughout:
+  `mjolnirReporter` / `MJOLNIR_REPORT_FILE` / `mjolnir.report.json`. The
+  package is unpublished, so no consumer breaks.
+
+### Added — guards
+
+- CI now runs `npm run test:coverage`. Its absence is why coverage fell from
+  ~96% to 92.6% between releases without anyone noticing.
+- `tests/docs-consistency.spec.ts` now asserts that every `npm run <script>`
+  referenced in tracked docs and source actually exists in `package.json` —
+  the general fix for the dead-command class above, not a one-off patch.
+- ~130 unit tests recovering branch coverage on the code-text maskers, the
+  shared position helpers, ignore-pattern resolution, and the per-arm
+  behavior of ten Playwright rules plus QA-PY-010.
+
+### Infrastructure — automated npm publishing is live
+
+- `0.5.0` is the first version published to npm by CI. `release.yml` now
+  publishes via **OIDC trusted publishing** (no `NODE_AUTH_TOKEN`
+  anywhere) with `--provenance`; the published tarball carries a SLSA
+  provenance attestation (`npm audit signatures`). Every subsequent
+  release is `git push --follow-tags` and nothing else.
+- The blocker was a mismatch in the npmjs.com Trusted Publisher config
+  (`Sergey-bar` vs the real `Sergey-Bar` — npm matches the OIDC
+  `repository` claim case-sensitively), which surfaced as
+  `OIDC token exchange error - package not found` / `ENEEDAUTH`. Fixed on
+  npmjs.com; re-run against the existing `v0.5.0` tag via
+  `workflow_dispatch`.
+
+### Known gaps
+
+- **19 of 91 rules carry a measured FP rate** (n ≥ 10, from 381 hand-classified
+  corpus verdicts). The other 72 ship on an unverified assumption; `mjolnir
+doctor` reports this and will fail once a majority is classified.
+- `NORMALIZATION_K` is unfitted.
+- Statements/branches coverage sits at 94.8%/87.7% against a 95/88 aspiration;
+  the enforced floor is 94/87 with the gap documented in `.planning/STATE.md`.
+
+## [0.4.0] — 2026-08-27
+
+### Changed
+
+- **BREAKING: Rebranded from QA Doctor to Mjölnir.** Package name is now
+  `mjolnir-qa` (bin: `mjolnir`). Config file: `mjolnir.config.json`.
+  Data directory: `.mjolnir/`. Badge: `mjolnir-badge.json`.
+- Score label: "SCORE" → "WORTHINESS".
+- Verdicts: "HEALTHY" → "WORTHY", "CRITICAL" → "UNWORTHY".
+- Environment variable: `QA_DOCTOR_ASCII` → `MJOLNIR_ASCII`.
+- SARIF tool.driver.name: "Mjölnir".
+- Repository: `github.com/Sergey-Bar/Mjolnir`.
+- CLI: all help text, error messages, usage strings reference `mjolnir`.
+- Generated workflows: `mjolnir.yml`, `npx mjolnir-qa@latest`.
+
+### Added
+
+- `--format mermaid` — test-architecture diagram (Sprint 9).
+- `--tone blunt` — opt-in blunter messages (Sprint 9).
+- Milestones — first flawless scan / first debt reduction announced once.
+- New MJÖLNIR ASCII art logo (minimal Nordic hammer).
+
 ## [0.3.x] — prior releases
 
 See git history; per-rule Trust Metadata `introduced` fields record the