feat(code-review): add graph-aware Design Critic reviewer

[wongk]

Summary

Adds the Design Critic — an always-on, deep-tier conditional core reviewer that evaluates software-design craftsmanship (module depth & information hiding, SOLID, dependency direction & layer boundaries, project structure), drawing on A Philosophy of Software Design (Ousterhout), SOLID, and Clean Architecture (Martin). It's source: "core" (exempt from the domain-critic cap), runs on Sonnet, and emits category: "Code Quality" findings scoped to design flaws a change introduces or worsens.

Bumps code-review to v3.1.0 (MINOR).

What's in it

New reviewer

• tools/prompts/design_critic_suffix.txt — the Design Critic fleet role (7-step evaluation procedure: complexity framing → project structure → dependency graph → module depth → SOLID → code-level → testability/consistency), wired into COVERAGE_CORE_CONDITIONAL with an {"type": "always"} trigger so it runs on every --depth deep review.

Domain-critic cap collapsed 5 → 3

• Uniform DOMAIN_CRITIC_CAP = 3 across standard and deep (removed the depth-aware STANDARD_DOMAIN_CRITIC_CAP/branch). Deep's extra breadth now comes from the Design Critic + Impact Analyzer rather than a wider critic allowance.

Design Critic is graph-aware

• Spawns on code-review-worker-graph and uses the codebase-memory-mcp knowledge graph (get_architecture for module/layer layout, query_graph for dependency direction & import cycles) when the repo is indexed, falling back to grep otherwise. The graph worker's tool set gained get_architecture + query_graph. Same safety envelope as the other graph roles (blanked when the server is absent, the project is ambiguous, or review_root is set).

Removed the dead callsite_snippet_hash field

• Dropped external_impact[].callsite_snippet_hash and the coupled EvidenceCheck.snippet_hash_matched (LLMs can't compute sha256; the field had no working consumer). Impact callsites are now validated by reading the cited file and content-matching the verbatim callsite_snippet.

Fleet-summary fix + accuracy cleanup (surfaced by a deep self-review of this branch)

• _render_fleet_breakdown now derives the non-partitioned core display set from _SPAWN_CORE_ROLES, so both conditional core reviewers (Impact Analyzer + Design Critic) render on the "Reviewers:" line — previously the Impact Analyzer was silently omitted. Added a regression test.
• Comment/docstring accuracy: COVERAGE_CORE_CONDITIONAL header, resolve_coverage docstring, _SPAWN_CORE_ROLES comment, and a stale /hash prompt reference.
• Extracted duplicated _diff/_empty_critic_gates test helpers into a shared mixin.

Validation

• pytest plugins/code-review/tools/python/ — 1217 passed, 3 skipped (+2 new regression tests)
• uv run ruff check . — clean
• uv run pyright — 0 errors
• README + CHANGELOG updated via /update-documentation

🤖 Generated with Claude Code