Fix run-once approval recovery

[rochecompaan]

Closes #137

Summary

• preserve durable spec-approved and plan-approved labels across run-once recovery transitions
• preflight approved artifacts before host mutations, including branch-only saved artifacts and workspace identity checks
• document durable approval/recovery behavior and add focused regression coverage
• include git in the Nix package test environment for real-Git regression coverage

Validation

• node --test src/cli/commands/run-once/approval-artifact-preflight.test.ts
• node --test src/cli/commands/run-once/approval-artifact-preflight.test.ts src/cli/commands/run-once/pipeline-approval-workspace-preflight.test.ts src/cli/commands/run-once/pipeline-workspace-scenarios.test.ts src/cli/commands/run-once/pipeline-planning.test.ts src/cli/commands/run-once/pipeline-development-environment.test.ts src/cli/commands/run-once/planning-artifacts.test.ts
• npm run test:run-once
• npm test
• npm run lint
• npm run build
• npm run site:build
• git diff --check
• nix build .#patchmill

Patchmill run cost

Stage	Model	Prompt tokens	Output tokens	Estimated cost (USD)
Development environment	gpt-5.5	533,324	2,267	$0.6196
Implementation	gpt-5.5	21,969,164	44,267	$17.0554
Implementation	gpt-5.6-sol	22,247,317	166,366	$25.8808
Implementation	gpt-5.6-terra	27,760,206	79,121	$11.5450
Implementation subtotal		71,976,687	289,754	$54.4813
Total		72,510,011	292,021	$55.1008

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.