Feat/harness retry stabilization

[romanstark]

Summary

This PR introduces a new stabilized retry strategy for harness resume flows and improves evaluator reliability for long-running sprint evaluations.
Key outcomes:

• Prevents already-verified criteria from flapping to fail on transient/inconclusive re-checks.
• Focuses retry work on still-failing criteria to reduce accidental regressions.
• Hardens evaluator JSON parsing with an automatic recovery pass when output is non-parseable.

What changed

1) Stabilized retry strategy (new feature)

• Added retryStrategy config with:
  • stabilized (default)
  • strict (legacy behavior)
• Added hardFailUnlockStreak config (default: 2):
  • Previously passed criteria are "locked".
  • Locked criteria are only unlocked after repeated hard failures.
• Added persistent sprint stability snapshots:
  • feedback/sprint-{n}-stability.json

2) Focused retries for generator

• On retry, generator now receives explicit failed-criteria focus.
• Prompt includes scope-control guidance to minimize changes outside failing criteria.

3) Threshold semantics hardening

• Contract criterion threshold is treated as an integer score gate on 1-10 scale.
• Invalid threshold values (e.g., metric targets like 90/250/48) fall back to global pass threshold.
• Prompt guidance updated so raw metrics belong in criterion descriptions, not in score threshold fields.

4) Evaluator JSON parse reliability hardening

• If evaluator output is not parseable JSON on first pass, harness retries evaluation once with strict JSON-only instruction.
• Claude evaluator response collection now captures additional result-event text to improve recoverability.
• Increased CLAUDE_MAX_TURNS from 50 to 80 to reduce truncation/non-final responses in complex evaluations.

Why

In long sprints, fully re-evaluating all criteria each retry can cause pass/fail oscillation due to:

• environment limitations (e.g., Lighthouse/Chrome unavailable),
• partial/inconclusive checks,
• model output formatting failures (non-JSON final responses).
  This PR keeps strong quality gates while reducing flaky regressions and improving run completion reliability.

CLI / behavior notes

• New flags:
  • --retry-strategy=strict|stabilized
  • --hard-fail-unlock-streak=<int>
• Default behavior now uses stabilized retries.
• Strict mode remains available for full immediate regression enforcement.

Files of interest

• shared/evaluation.ts (new): stabilization + threshold helpers
• shared/types.ts
• shared/files.ts
• shared/config.ts
• shared/prompts.ts
• claude-harness/{index.ts,harness.ts,generator.ts,evaluator.ts}
• codex-harness/{index.ts,harness.ts,generator.ts,evaluator.ts}
• README.md