Migrate plan reviews and revisions to structured responses

[wwind123]

Fixes #155

Summary

• route plan review and plan revision through structured parsing first with strict JSON-prefix/footer validation
• normalize structured planning comments and canonical plan rendering for prompts, hashing, posting, and resume
• update planning prompts and add regression coverage for structured parsing, rendering, and resume behavior

Testing

• python3 -m pytest tests/test_agent_loop.py

[wwind123]

Implemented the structured planning migration for issue #155 and opened PR #167.

Key changes:

• routed plan_review and plan_revision through strict structured-first parsing with markdown fallback only when no top-level JSON prefix exists
• added canonical structured plan rendering plus canonical_plan metadata so hashing, reviewer prompts, public comments, and resume use deterministic plan text
• updated planning prompts to describe the structured schema first and expanded regression coverage for parser, renderer, prompt, and resume behavior

Tests: python3 -m pytest tests/test_agent_loop.py (passed)

-- OpenAI Codex

[wwind123]

{
  "schema_version": 1,
  "kind": "pr_review",
  "state": "approved",
  "summary": "Solid structured-response migration for plan reviews and revisions. Hard-fail semantics after JSON-prefix detection, canonical plan storage for deterministic resume, dedicated public-comment renderers, updated prompts, and comprehensive regression coverage. The refactored `_consume_structured_footer_and_signature` correctly parameterises the state regex and human-requirements prefix rules for each document type. CI passes; no blocking issues found.",
  "blocking_items": [],
  "same_pr_followups": [],
  "future_followups": [],
  "prior_item_dispositions": []
}

-- Anthropic Claude

[wwind123]

{
  "schema_version": 1,
  "kind": "pr_review",
  "state": "approved",
  "summary": "Solid migration of plan reviews and revisions to structured responses. The implementation provides robust JSON-prefix/footer validation with hard-fail semantics, ensuring deterministic and unambiguous agent communication. Canonical rendering and metadata-backed resume logic preserve plan integrity across rounds even when agents include additional prose in public comments. Prompts have been updated with clear guidance and examples, and the regression coverage is comprehensive. The refactoring of structured response extractors into specific helpers is a significant improvement for maintainability.",
  "blocking_items": [],
  "same_pr_followups": [],
  "future_followups": [],
  "prior_item_dispositions": []
}

-- Google Gemini