feat(run-once): integrate planning phases and require implementation pull requests

[rochecompaan]

Summary

• route fresh and active run-once work through the locked planning-pr-v1 phase coordinator while preserving unfinished legacy recovery
• support all spec/plan review gate combinations, planning review stops, and the explicit --plan-only boundary
• force implementation pull requests and independently validate target, head, status, marker, closing reference, remote head, URL identity, and commit ancestry
• persist strict implementation and finish checkpoints before cost, handoff, cleanup, branch removal, or lifecycle effects
• add recording scenarios and rejection matrices for workflow routing, recovery, implementation proof, and destructive-effect ordering
• split prompt, URL identity, and strict-state codecs into focused modules after final maintainability review

Validation

• focused issue Integrate planning phases and require implementation pull requests #189 suites: 113/113 passed
• npm run test:run-once: 699/699 passed
• npm test: 1802/1802 passed
• npm run build: passed
• npm run lint: passed
• npm run check:types: passed
• npm run check:architecture: passed (467 modules, 1986 dependencies, no violations)
• git diff --check: passed
• dependency metadata unchanged; conditional Nix build skipped per repository policy

Reviews

• Final Codex full-worktree review: passed after correctness and recovery findings were fixed
• Final thermo-nuclear full-worktree review: passed after canonical-helper, module-boundary, and test-matrix fixes
• Final validation-readiness review: passed on exact head 68eec0ac5527a7c4ac32e8ede5aec42c394723ca with no findings

Human review is required because this is a large, security-sensitive workflow/state-machine change with new public result discriminants and destructive cleanup boundaries.

Closes #189

Patchmill run cost

Stage	Model	Prompt tokens	Output tokens	Estimated cost (USD)
Development environment	gpt-5.6-sol	429,129	3,831	$0.6737
Implementation	gpt-5.6-sol	214,406,782	721,953	$183.2252
Implementation	gpt-5.6-terra	97,766,578	222,981	$27.9399
Implementation subtotal		312,173,360	944,934	$211.1651
Total		312,602,489	948,765	$211.8388

Prompt tokens include uncached input, cache reads, and cache writes. Cost is an estimate based on Pi's recorded model pricing and includes parent and subagent sessions.