Phase-Wise Checklist

Create one canonical plan under plans/ and treat it as a live execution ledger.

Workflow

1. Read the named parent plan/checklist and inspect the current implementation before changing any status.
2. Use the repository graph tools for code discovery when available; otherwise use focused source search. Treat plan prose as a claim, not evidence.

3. Start with the required plan shape:

   # <Hierarchy> - <Title>
   
   > **Parent:** `<path>` - reference
   > **Status:** current state
   > **Estimated effort:** estimate
   
   ---
   
   ## Overview
   ## Executive Summary
   ## Phase 1: <Area>
   ### 1.1 <Slice>
   - [ ] Verifiable action
   ## Dependencies

4. Make every checklist row atomic: one code change or one validation result. Include the affected path, rule/issue ID, expected behavior, and required proof where useful.
5. Order phases by dependency and risk: correctness/security first, then API/data contracts, then performance/cleanup, then closure gates.
6. Use statuses consistently:
   • [ ] not started or not proven;
   • [x] implemented and validated with current evidence;
   • [~] intentionally deferred/partial, with reason, owner boundary, and next gate.
7. Update a row only after the matching source/test/benchmark check succeeds. Record the command and outcome beside closure gates; keep release measurements distinct from dev-loop measurements.
8. If a row is moved to a newer canonical ledger, rewrite the source row as [~] with an explicit pointer. Do not leave duplicate active work in multiple files.
9. Preserve unrelated worktree changes. Do not close a security, performance, or migration item merely because a plan says it was done.

Evidence Rules

• Prefer current code, tests, benchmarks, scanner output, and CI configuration over historical notes.
• State negative results precisely: e.g. "no production unsafe beyond X found in this audit", never "no bugs exist".
• Keep risk statements separate from confirmed defects. Label hypotheses and give their validation step.
• For performance items, include the exact release command, dataset/path, cold/warm cache state, and metric; successful execution is not benchmark proof.

Completion Handoff

Before declaring a phase complete, confirm its rows, run the smallest relevant validation, synchronize the checklist, and provide a concise result plus the next unchecked phase. Do not create a second status document unless it is explicitly a deferred ledger with a pointer from the canonical plan.

Required Checks

• For documentation-only changes, do not run lint or test checks.
• For every non-documentation change, run make lint and make test before marking the phase complete. Record both outcomes in the canonical checklist; leave the row unchecked if either command fails.