v0.6.0

Added

• Claim tiering + model-diversity judge slots (@prd-gen/verification).
  A claim whose own text names a deterministically executable verification
  method (grep/diff/time/kcov/exit-status/named gate) skips the judge panel
  and gets a synthesized rule-tier verdict instead
  ({kind:"rule",name:"rule-tier"}); architecture-typed claims get one
  judge per model in VerifyBudgetConfig.diversity_models (default
  ["haiku","sonnet"]) instead of N persona-prompted judges on one
  underlying model. JudgeVerdict.model records which model judged each
  claim; 10-verification-report.md renders a cross-model-agreement
  summary. Measured on the calibration fixture: 30 → 23 judge invocations
  (-23%).
• Explicit verification acceptance policy at implementation_gate.
  VerificationPolicyConfig (block_on, min_subjective_sampled_ratio,
  on_unsampled_below_ratio, on_cross_model_disagreement —
  composition-root-injectable, default null uses
  DEFAULT_VERIFICATION_POLICY) and evaluatePolicy(), a pure function
  turning verification results into pass / needs_attention / blocked.
  implementation_gate now shapes its "Implement / PRD only" question from
  that verdict — a bare "Implement" is never offered while blocked — and
  records any human derogation on post_specs.policy_derogation.
• Host-side external-judge executor (scripts/external-judge/). A
  zero-dependency Node CLI (judge.mjs) that posts a judge prompt to an
  OpenAI-compatible endpoint (Gemini via AI Studio, Mistral via La
  Plateforme) when a spawn_subagents invocation names a non-Anthropic
  model, plus a calibration harness (calibrate.mjs) gated on agreement
  ≥ 0.7. No API key configured produces an explicit skipped result, never
  a fabricated verdict.
• Bilingual (FR/EN) hard-output-rule detection: opt-out markers and
  per-rule topic/keyword signal lists now recognize French phrasing via a
  shared lexicon (rules/lexicon.ts); 12 previously English-only rule
  checks (crypto, input validation, output encoding, structured logging,
  alerting thresholds, API contract docs, deprecation strategy, etc.) gain
  an explicit opt-out path, audited by
  packages/validation/src/__tests__/opt-out-coverage.test.ts.
• Budget-gated haiku judge panel for self_check verification
  (PipelineState.verify_budget, composition-root-injected): default
  panel reduced to 1 judge/claim (2 for architecture claims); a budget
  gate asks the user (reduced sample / full fleet / skip verification)
  when the invocation count exceeds a configurable cap (default 20).
• submit_action_result / start_pipeline response-size bound
  (boundEnvelopeResponse): an oversized spawn_subagents action has
  every invocation's prompt replaced by an observable OmittedStub; the
  full unbounded action is recoverable via
  get_pipeline_state(run_id, format:"action").
• 10-verification-report.md now written by implementation_gate before
  the implementation decision is asked, carrying per-claim judge verdicts
  when available.
• Root pnpm lint typechecks every workspace package via tsc --noEmit
  (lint script added to all 9 TS packages) and runs in CI.

Fixed

• Root lint script was structurally broken: tsc --noEmit at the repo
  root found no tsconfig.json (only tsconfig.base.json exists), so it
  printed CLI help and exited non-zero — invisible because CI never ran
  it. Fixed by delegating to each package's own lint script via
  pnpm -r run lint; CI now runs the step so it cannot rot silently again.
• Hard-output-rule false positives on French-language PRD sections
  (technical_specification, cryptographic_standards, rate_limiting,
  secure_communication, GDPR consent, distributed_tracing) that had
  explicit, justified "non applicable" prose the English-only detector
  could not recognize.
• test_traceability_integrity's test-function pattern only matched
  Swift-style func test_xxx(, so bash-defined tests (test_xxx() { },
  function test_xxx() { }) were reported missing even when present.
• claim-extractor.ts's evidence-snippet window used a fixed ±N-line
  radius that ignored claim boundaries, letting an adjacent claim's
  wording bleed into the current claim's evidence and mis-tier it;
  snippet() now stops at the neighboring claim's own start line.
• file-export.ts wrote placeholder text for companion files whose
  source section(s) produced no content; such files are now omitted
  entirely, with the omission and its reason recorded in
  00-run-notes.md (numbering stays stable).
• renderJudgeVerdicts stringified the structured AgentIdentity judge
  field as "[object Object]" instead of "kind:name".

Fixed (release workflow)

• release.yml never wrote the real .mcpb SHA-256 back into
  server.json#/packages/0/file_sha256, leaving a permanent
  000...000 placeholder in every published release's manifest. The
  workflow now patches the checksum after packing the .mcpb and
  pushes the single-file update back to main as part of the same
  release job, so this fix is already in effect for the v0.6.0 tag
  itself. Fixes #23.

.mcpb bundle

SHA-256: c4897edab70e8a4db79dfc22167087891e3b62131dce1fe40280e9ed0f95d5e3