E2its Myrmion AI Factory Factory Pr Review

Factory PR Review — Push Gate + Assistive Reviewer

Shared Protocol — Referenced by: ALL agents that ship code (IMPLEMENT, QA, DEVOPS, CODESIGN/BLUEPRINT for spec-bearing PRs) + the check-push-preflight.sh hook. Position in the SDLC: runs immediately before git push as a local quality gate. Does NOT replace human review on the PR — it complements it by catching blockers before they hit the PR.

Why a push gate, not a merge gate

A merge gate (CI on PR) is too late: the work is already on the remote, the author has context-switched, and reviewers waste a round trip on findings the author could have caught locally. The push gate runs against origin/{base}..HEAD, exits non-zero on hard-blockers, and produces the same finding catalog the post-push review would. The assistive --review {PR} mode is kept for already-open PRs (ad-hoc review, second opinion, pre-merge final pass).

Hard-blocker exit semantics:

• exit 0 → no blockers, push proceeds
• exit 1 → blockers found, push blocked with humanised findings
• exit 2 → tooling failure (missing dep, repo state error) — NOT a blocker, push proceeds with warning

When this skill activates

The push hook is the default integration. Manual invocation is for the assistive cases.

Operation modes

Mode 1 — --preflight (push gate, default)

Runs locally against the current branch's diff vs its base. NEVER touches the remote. NEVER posts to a PR.

.claude/skills/factory-pr-review/scripts/preflight.sh [--base origin/main] [--json]

Steps (executed by preflight.sh):

1. Resolve base (origin/main by default, or read from .claude/rules/branching.md default_base_branch).
2. Compute git diff --name-only origin/{base}..HEAD.
3. Docs-only fast-lane — if every changed path matches **/*.md, docs/**, .context/templates/**, .gitignore (and none under .github/workflows/**), exit 0 with fast-lane: docs-only note. Skip remaining checks.
4. Run detect_change_type.py → flags JSON.
5. Run check_docs_sync.py --git-range origin/{base}..HEAD --json → docs findings.
6. If any docs/spec/*/dev_plan.md in diff, run check_dev_plan_task_format.py --git-range origin/{base}..HEAD --json → IMPLEMENT plan task-format findings (orphan ### X.N h3 vs canonical - [ ] [X.N] checkbox; status: READY plans with zero unchecked tasks).
7. If has_openapi: true, run check_openapi_diff.sh origin/{base} <spec-path>.
8. If has_asyncapi: true, run check_asyncapi_diff.sh origin/{base} <spec-path>.
9. Framework-aware checks (run unconditionally — see § Framework-aware Hard Blocks below).
10. Aggregate findings; print summary; exit 1 if any blocker, 0 otherwise.

Mode 2 — --review {PR-URL or branch}

Full audit against an open PR (or local branch). Produces the structured review (assets/review_template.md) and OPTIONALLY posts via post_review.py (only with explicit user confirmation — see Phase 6).

Framework-aware Hard Blocks

These extend the generic hard blocks (SKILL.md Phase 4 in the upstream skill) with framework-specific gates. ALL apply universally; some are scoped to materialised projects vs the framework meta repo.

Block 11 (Governance-bump miss) is the framework-meta equivalent of "missing CHANGELOG entry". It enforces the rule that lives in the root CLAUDE.md Generation Standards §2.

Blocks 13-18 are produced by Phase 0 Coherence Audit (semantic, agentic). 13-15 are deterministic-leaning but require the agent to interpret rename intent and lock-step pair semantics; 16 is split by context (one variant active per repo); 17-18 require semantic judgment over commit message + manifest bump kind. All six read their context-specific configuration from config/coherence-context.json (audit.* keys).

Block 19 (Cyclomatic complexity) is produced by axis 6 in Phase 3 (per-axis analysis). The skill factory-complexity-check invokes a project-configured MCP (default Semgrep, RDR-ratified at SETUP Q23.1) on the cumulative branch diff; severity routing depends on config/quality.json.complexity.pr_blocker (default false → advisory). The skill fails open — missing config, unavailable MCP, or unparseable response degrade to a banner-only advisory.

Framework artefact ↔ docs sync matrix

Extends references/docs-sync-checklist.md with the framework's own artefacts. Only applies to materialised projects (the meta repo has no docs/spec/ tree).

Workflow (--preflight)

Phase 0 — Coherence Audit (agentic, semantic)

Runs BEFORE all other phases when the diff is governance-sensitive (touches any path under audit.root_sets of config/coherence-context.json, minus audit.exclusions). For non-sensitive diffs (e.g. pure UI under src/components/, build-tooling-only), Phase 0 is skipped — go straight to Phase 1.

The audit is performed by the agent (Claude), NOT by a deterministic script. Reason: rename intent, lock-step semantics, commit↔diff coherence, and bump-severity judgment all require contextual reasoning that regex/grep cannot reliably perform without false negatives. The deterministic phases (1-6) stay deterministic and cheap; Phase 0 is the one phase that pays a model-call cost in exchange for completeness.

Scope discipline: Phase 0 reads ONLY (a) files in the diff and (b) files in the collateral set built in Step 0.3. It NEVER scans the whole repo. The collateral set is bounded by audit.root_sets minus audit.exclusions.

Execution split: Phase 0 has two sides. Steps 0.1–0.5 are agent-side (LLM judgment) — the agent reads config/coherence-context.json, runs git grep for collateral, applies the six detectors, and emits the marker. Step 0 inside preflight.sh is script-side (deterministic verification) — it only reads the marker and decides whether to block, never executes the audit itself. The agent is responsible for running Phase 0 before pushing; the push gate fails closed if the marker is absent. Trigger entry points for the agent: when the user asks "review this branch" / "audit before push" / explicitly invokes the factory-pr-review SKILL — the agent recognises the SKILL description match and runs Phase 0.1–0.5 inline.

Step 0.0 — Skip checks

• Detached HEAD or protected branch → skip with note (defence-in-depth, not Phase 0's concern).
• Diff matches the docs-only fast-lane (CLAUDE.md Generation Standards §3) → skip Phase 0; existing fast-lane handles.
• config/coherence-context.json missing → emit humanised warning, treat as downstream with empty allowlists, continue.

Step 0.1 — Load context

Read config/coherence-context.json. Parse:

• context → meta | downstream. Selects which Block 16 variant is active and which lock-step pair types apply.
• audit.root_sets → directories the audit is allowed to scan when expanding to collateral.
• audit.exclusions → globs globally excluded.
• audit.placeholder_allowlist_paths → globs where {{X}} is permitted (Block 14).
• audit.lock_step_pairs → typed declarations driving Block 16 (meta_template_mirror for meta; rename_propagation + code_doc_lock_step for downstream).
• audit.manifest_paths → manifest location for Blocks 15/18.

Step 0.2 — Extract changed symbols from the diff

Run git diff --name-status origin/{base}..HEAD and git log --format=%B origin/{base}..HEAD. Extract:

• Renamed paths — git rename detection (R{score} entries).
• Deleted paths — files that disappeared.
• Moved/renamed identifiers — section headings changed in .md files (-## OldName paired with +## NewName); name: field changes in skill frontmatter; function/class definitions removed/added with similar names.
• Public artefact names — Factory-* skill names, rule names (basename of .claude/rules/*.instructions.md), instruction names, command names, hook script names.

Each extracted symbol → search term for Step 0.3.

Step 0.3 — Build collateral set

For each symbol from 0.2, run git grep -l <symbol> constrained to audit.root_sets minus audit.exclusions. Files that match AND are NOT in the diff = collateral set. Read those files in full. The set is naturally bounded — typical EVOLs touch ≤ 5 symbols and ≤ 30 collateral files.

Step 0.4 — Apply detectors 13-18

Run each detector against diff ∪ collateral. Routing:

Each finding cites file/line/snippet/why/fix per the existing severity rubric (references/severity-rubric.md).

Step 0.5 — Emit marker

After Phase 0 completes (regardless of verdict count), write a marker file:

.claude/state/coherence-audit-${branch_sha}.marker

• branch_sha = git rev-parse HEAD at audit time.
• The agent MUST mkdir -p .claude/state/ before writing — the directory is .gitignored and may not exist on a fresh clone or after a state cleanup.

Marker body (JSON, single line):

{"branch_sha":"...","audited_at":"ISO-8601","governance_sensitive":true,"findings":{"blocker":N,"important":N,"nit":N,"question":N}}

The marker proves Phase 0 ran for this exact tree state (commit sha). New commit → new sha → marker invalidated → audit must re-run. Same code state across sessions reuses the marker — re-auditing identical bytes is wasted work.

Step 0.6 — Verdict

• Any 🔴 Blocker → STOP. Surface findings to user with humanised resolution path. Do NOT write marker (or write marker with "blocker": >0 and let preflight.sh refuse to proceed).
• 0 🔴, any 🟡/🟢/❓ → write marker with finding counts; advise user on the non-blocker findings; proceed to Phase 1.
• 0 findings → write marker; proceed to Phase 1 silently.

Phase 1 — Branch + base resolution

current=$(git branch --show-current)
base=${BASE:-origin/main}  # or read from .claude/rules/branching.md
git fetch origin "${base#origin/}" --quiet
git diff --name-only "$base"..HEAD

If current is empty (detached HEAD) OR equals a protected branch name (main, master, develop, bare hotfix, release/*) → exit 2 with "preflight skipped: not on a working branch".

Phase 2 — Change classification

Run scripts/detect_change_type.py --git-range "$base"..HEAD --check-secrets. The flags drive which references the agent loads when surfacing findings.

Phase 3 — Per-axis analysis

Apply this routing table (load reference + run script):

Phase 4 — Hard-block enforcement

The 19 hard blocks in § Framework-aware Hard Blocks. Each one is a deterministic pass/fail. The agent does NOT downgrade these — they are blocks by definition of the rubric.

Phase 5 — Review generation

Use assets/review_template.md for the JSON structure. In --preflight mode the review is printed locally; in --review mode it is rendered to Markdown via post_review.py.

Required structure (severity buckets):

1. Summary (2-3 sentences) — what the change does + verdict.
2. 🔴 Blockers — push-blocking. File/line/snippet/why/fix per item.
3. 🟡 Important — should be fixed; doesn't block push.
4. 🟢 Nits — author decides.
5. ❓ Questions — legitimate doubts.
6. 👏 Praise — what's well done.
7. 📋 Documentation checklist (from references/docs-sync-checklist.md).
8. 🔧 Mitigations applied (only when non-empty — see Phase 5.5).

Phase 5.5 — Mitigation hot-fix (--review mode only, optional)

When the audit catches a clear-cut blocker that the reviewer can fix mechanically, the reviewer MAY apply the fix to the PR branch and record it in mitigations_applied[] instead of leaving the finding open. This shifts the resolution from "ask the author" to "show the author what was done" while preserving the audit trail.

Eligibility (ALL must hold):

1. Mechanical fix — unambiguous resolution (replace pattern X with Y; add missing manifest entry; sync byte-identical mirror file). No design choice, no stylistic preference.
2. Clear-cut category — security regex hit; broken reference; lock-step drift; manifest↔disk drift; placeholder leakage; missing test for new logic with obvious test pattern.
3. High confidence — reviewer can verify locally (tests pass, linter clean, build green).
4. Same branch — commit lands on the PR branch, not on a separate fix-branch.

If any criterion fails, leave the finding open in Blockers/Important and let the author decide. Mitigations are contributions, not overrides — the author always retains the option to revert.

Workflow:

1. Apply the fix. Commit to the PR branch with a Conventional Commits message: fix(EVOL-NNN): {what was fixed} (review mitigation) or similar attribution.
2. Run any relevant verification locally (tests, lint, build). Capture the verification evidence.
3. Push the commit to the PR branch (the push gate Phase 0 may re-fire — that's expected; re-audit if so).
4. Add an entry to mitigations_applied[] in the review JSON: finding_ref cites the original finding (blockers[N] or important[N]); commit_sha is the resolution commit's full SHA; description summarises what was done; verified_by cites evidence.
5. The corresponding entry in blockers[] / important[] STAYS — do not delete it. The author needs to see both the finding and the resolution.
6. decision SHOULD remain comment (not approve) when mitigations were applied — author must confirm acceptance.

If mitigation breaks something else (verification fails), revert the mitigation commit, leave the finding open, and explain in description what was attempted and why it failed.

Phase 6 — Publication (--review mode only)

Use scripts/post_review.py to publish on the platform. NEVER without explicit user confirmation — an incorrect review on GitHub is publicly visible. The push-gate (--preflight) NEVER publishes.

The script is dry-run by default: invoked without flags it only renders and prints the review. Publishing requires the explicit --publish flag, which the agent adds ONLY after the user has confirmed the rendered output. The review body is sanitised (control characters stripped) and passed to the platform CLI via --body-file, never inline through argv.

When mitigations_applied[] is non-empty, post_review.py renders a "🔧 Mitigations applied during review" section after Findings, citing each mitigation's commit SHA + description + evidence. The author sees finding + resolution in a single comment.

Review principles

• Severity over volume: 3 real blockers > 30 nits.
• Approve if it improves codebase health, even if not perfect.
• Ask before stating when unsure: "Is it intentional that…?" beats "This is wrong".
• Cite evidence: line number, code snippet, command to reproduce. "I think there's a problem in X" doesn't cut it.
• Be specific with proposals: "consider using X" beats "this could be improved".
• Don't duplicate what a linter catches. If ESLint/Ruff/golangci-lint will flag it, it's not human review work.
• Don't comment style if a formatter exists (Prettier, Black, gofmt). The formatter wins.

Iron law: don't claim without verifying

Before flagging a finding as a verified blocker, the skill MUST be able to show:

• The exact file and line.
• The actual code snippet (not paraphrased).
• An explanation of why it's a problem in THIS codebase (not generic).
• A concrete fix proposal.

If you can't verify one of these four points, downgrade to "Question".

PR Persistence Contract — MANDATORY

After EVERY review cycle (preflight blocker resolution, manual --review, /loop pr review iteration — including zero-finding clean first-pass), the agent MUST post a structured comment to the PR that captures the cycle state. Chat sessions disappear; the PR is the durable audit record. No silent passes — green is a deliberate verification result that MUST surface on the PR.

When to post

Always. Per cycle, exactly one comment:

• --preflight mode (blockers resolved): post after the fix lands.
• --review {PR} mode: post per run.
• /loop pr review iteration: post per iteration regardless of finding count.
• Clean first-pass (no findings, CI green out of the gate): post a green-state confirmation comment naming every gate that passed. The reader MUST be able to see that verification ran and concluded clean — silence is indistinguishable from "nobody checked".

What to include

The comment MUST contain:

1. Exit state table — every gate state at the cycle end. Minimum rows: mergeStateStatus, CI workflow status (per check), preflight.sh findings count, Phase 0 detector summary (or skipped — docs-only fast-lane when applicable), governance-script local results (validate-governance.sh, check-applicability-frontmatter.sh, check-iteration-id-format.sh, check-adr-constitution-sync.sh, and any project-specific gates), Phase 0 marker file path (or N/A).
2. Per-iteration root-cause block — for each iteration that produced findings: symptom (what failed and where), root cause (the underlying bug, not just the surface error), fix (commit SHA + one-line description + manifest bump if any), verification (which local check + which CI run confirms green). Clean first-pass iterations skip this block (no symptoms to report) but MUST still produce the exit state table.
3. Phase 0 detectors table — explicit per-detector result (13 Reference coherence / 14 Placeholder leakage / 15 Manifest↔disk / 16-meta or 16-downstream Lock-step / 17 Commit↔diff / 18 Bump severity). When Phase 0 was skipped, render one row stating Phase 0 skipped — docs-only fast-lane per SKILL.md § Phase 0 Step 0.0. Mark false-positive questions with ❓ + brief justification.
4. Branch history — table mapping each commit SHA on the branch to its commit subject. Anchors the comment to git ground truth.
5. Stale-signal disclaimer — if earlier github-actions[bot] failure comments exist on the PR, explicitly call them out as historical so reviewers don't act on them. Omit when no stale signals exist.

Format

• Markdown.
• Top-level heading:
  • ## 🔁 PR-Review Loop — Iteration Log for /loop cycles with at least one iteration that had findings.
  • ## ✅ PR-Review Loop — Clean First-Pass for /loop cycles where every iteration was zero-finding green.
  • ## 🔍 PR-Review — Manual Audit for --review mode.
• Tables for everything that can be tabular (state, detectors, branch history).
• Link every commit SHA to ../commit/{sha}. Link every CI run to its job URL.
• Code blocks for command outputs that justify a verdict (e.g. the diff that proves lock-step is restored).
• Trailer: 🤖 Generated by /loop pr review cycle (Claude Code) (or --review variant).

Implementation

Posted via gh pr comment <PR> --body "$(cat <<'EOF' ... EOF)". The skill consumer is responsible for the heredoc; the SKILL itself does not ship a script for this (the content is per-cycle and not templatable beyond the headings).

When this rule applies

Universal — both meta and downstream contexts. Every PR that any review-cycle mode touches gets exactly one comment per cycle. No carve-outs.

Why

Without persisting the analysis on the PR, the chain "I saw a failure → I diagnosed → I fixed → I verified" — OR the chain "I verified, everything clean" — lives only in the chat transcript. Reviewers who arrive later have no way to distinguish "no findings because nobody checked" from "no findings because the cycle ran and confirmed clean". The PR comment is the framework's institutional memory for review-cycle work; explicit green-state confirmation is part of that memory.

Integration with other Factory protocols

Per-context behaviour

Framework meta repo (this repo)

• Block 11 (governance-bump miss) is active.
• Blocks 7-8-12 are inactive (no docs/spec/, no codebase_inventory.json, no protected-paths.json in the meta — the framework IS the protected code, governed by Block 11 instead).
• Docs-only fast-lane allowlist matches CLAUDE.md Generation Standards §3 verbatim.

Materialised projects (downstream)

• All 12 blocks active.
• Docs-only fast-lane only when the project's own CLAUDE.md declares it (defaults to OFF).
• config/protected-paths.json consulted for Block 12.
• docs/spec/{ID}/ artefacts drive Block 8 (CVP subset).

Hook wiring (push gate)

.claude/settings.json (and template counterpart) registers a PreToolUse hook on Bash:

{
  "hooks": {
    "PreToolUse": [
      {
        "matcher": "Bash",
        "hooks": [
          {
            "type": "command",
            "command": "bash .claude/hooks/check-push-preflight.sh"
          }
        ]
      }
    ]
  }
}

The hook script reads stdin JSON, extracts tool_input.command, matches /(^|\s|;|&&|\|\|)git\s+push(\s|$)/, and on match runs scripts/preflight.sh. Non-git push commands pass through unchanged. Hard-blockers cause the hook to exit 1 with a humanised message; the agent sees the block and reports it to the user instead of the raw script output.

Project rules and ADR precedence — MANDATORY

The skill is framework code; it ships identical to every project. Project-specific behaviour comes from project rules and ADRs, NOT from skill defaults.

Scope note. This skill is the push gate. It is NOT the inline 🔍 REVIEW hat that runs during IMPLEMENT --build (DEV ↔ REVIEW ↔ SEC). Those are governed by .claude/rules/review-policy.md and live inside the BVL cycle. The push gate is a downstream, independent concern: it runs at git push time on the cumulative branch diff, regardless of what happened during IMPLEMENT.

Hierarchy (highest to lowest, on conflict)

1. docs/constitution.md — project constitution. The Hard-Gate categories (CIP, CVP, IPP, BVL, GCRP) are constitutional and cannot be disabled.
2. ADRs and FDRs — docs/project_log/adr/*.md (project-wide ADRs that amend constitution.md via factory-adr-management Accept Procedure) and docs/spec/{ID}/fdr/*.md (feature-scoped FDRs; legacy docs/spec/{ID}/adr/*.md is read until projects migrate). They may refine, exempt, or extend the rule defaults below; they CANNOT relax constitutional gates.
3. .claude/rules/*.instructions.md — materialised at SETUP, evolved via ADR-backed updates only. These are the operational source of truth for each block category.
4. Skill defaults — last-resort fallback when none of the above is present (e.g. fresh project before SETUP --generate completes).

If a project rule disagrees with a skill default, the project rule wins. If an ADR documents a deviation from a project rule, the ADR wins. The skill NEVER imposes its own pattern over a project decision.

Project rule files consumed by the push gate

Honest, narrow surface — only files the preflight script actually reads (or that drive a hard block deterministically). Everything else lives in the NOT consumed section below to prevent scope creep.

That's it. Anything not in this list is NOT consumed. If the push gate were to start consuming a new rule file, that change ships in a new EVOL with explicit manifest entries.

NOT consumed (different reviewer, different lifecycle)

Same disambiguation pattern as review-policy.md:

• .claude/rules/review-policy.md — BVL/IMPLEMENT 🔍 REVIEW hat policy (DEV ↔ REVIEW retry loop, escalation, override, environment-driven strictness). Different reviewer, different lifecycle.
• .claude/rules/security_policy.md — high-level project security policy (allowed crypto, PII handling, OWASP stance, threat model). Consumed by 🛡️ SEC hat in IMPLEMENT --build (SAST + dependency audit), QA --verify (DAST), and AUDIT (compliance dimension). The push gate's secret-pattern regexes are NOT here — they live in scripts/detect_change_type.py and may be extended only via skill changes (manifest-tracked), not via project rules.
• .claude/rules/testing.md — testing policy (coverage thresholds, framework choice, AAA pattern). Consumed by BVL during DEV ↔ REVIEW (coverage gate), IMPLEMENT --build (TDD enforcement), QA --verify (test-plan execution). The push gate's "tests-deleted-without-justification" check (Block 5) reads the diff alone, not the policy file.
• .claude/rules/architecture.md — layer dependency rules, module boundaries. Consumed by BLUEPRINT (design.md §7.8 Mandatory Patterns), BVL REVIEW hat (Check #1), AUDIT. The push gate does NOT enforce architecture; that is a deeper, design-time concern.
• .claude/rules/contract-first-policy.md + .claude/rules/api-standards.md — contract authoring + versioning policy. Consumed by BLUEPRINT (CONTRACT-FREEZE gate), factory-coherence-validation/SKILL.md (CVP Check 14/15). The push gate's API contract checks (Blocks 1, 2) operate on the spec files themselves via oasdiff/asyncapi-cli; the policy text governs how specs are written, not how the push gate reads them.
• .claude/rules/defect-prevention.md — DC catalog. Consumed by every SDLC agent listed in each entry's applicable_to: field (see CLAUDE.md § Living Governance Catalogs). The push gate is NOT in the consumer list — defect prevention is a design-and-build-time concern, not a push-time one.
• config/codebase_inventory.json — component registry. Consumed by factory-codebase-inventory/SKILL.md (CIP Canary at IMPLEMENT --build), AUDIT (DRY dimension). Block 7 in this skill (CIP) is a policy reminder, not a deterministic check — the push gate does NOT scan the inventory; it relies on the upstream CIP Canary having run during IMPLEMENT --build. If you push code that bypassed CIP Canary (e.g. edited outside Claude), the push gate will not catch it; QA --verify is the next gate.

The pattern: each rule file has ONE primary consumer. Multiple agents may read different slices, but ONE skill or SDLC phase owns the binding-decision authority. The push gate is deliberately narrow — it covers the things that are cheap and deterministic at push time, and defers everything else to its proper owner.

ADR / FDR-driven exemptions

ADRs (project-wide) and FDRs (feature-scoped) may declare frontmatter exemptions consumed by the push gate:

---
status: accepted
date: 2026-04-28
decision-makers: [ARCH]
pr_review_overrides:
  block_1_public_endpoint_without_spec: scoped
  block_1_spec_paths: ["contracts/openapi/v3/*.yaml"]
  block_8_cvp_subset: ["0a", "0c", "1", "2", "13", "14", "15"]   # disable check 16 for this project
---

The skill reads pr_review_overrides: from every ADR under docs/project_log/adr/ (project-wide), every FDR under docs/spec/{ID}/fdr/ (feature-scoped, when the diff is feature-bounded), and any legacy docs/spec/{ID}/adr/ entries kept for backward compatibility. Project-wide ADRs override skill defaults; FDRs override project-wide rules within the feature scope only. Anything not declared in an override stays at the rule-file default.

A project that wants to disable a check entirely declares it in an ADR (or FDR for feature-scoped exemptions), which is itself reviewed and ratified. There is no per-developer escape hatch — bypassing the gate without an ADR/FDR is a governance-scope violation.

Customisation (deprecated path)

A project can also create references/local-policy.md in its copy of the skill, but this is deprecated in favour of the rule + ADR mechanism above. The rule + ADR path is governance-tracked (manifest-bumped, ADR-ratified) whereas a local-policy.md is an opaque skill-local override. Use it only for one-off experiments; promote to a rule + ADR before merging to main.

Skill structure

factory-pr-review/
├── SKILL.md                          ← this file (orchestrator + push-gate spec)
├── README.md                         ← installation + framework integration
├── references/
│   ├── severity-rubric.md            ← blocker/important/nit + Hard-Gate mapping
│   ├── code-review-criteria.md       ← per-axis criteria + DRY/CIP linkage
│   ├── docs-sync-checklist.md        ← code↔docs + framework artefact matrix
│   ├── api-contract-rules.md         ← OpenAPI/AsyncAPI + breaking changes
│   ├── adr-policy.md                 ← ADR location: docs/project_log/adr/
│   └── changelog-policy.md           ← CHANGELOG + governance_versions.json (meta)
├── scripts/
│   ├── preflight.sh                  ← orchestrator (push gate entry point)
│   ├── detect_cha

…(truncated)