feat(mcp_readability): reconcile dropped findings on re-judged tools - #698

Third part of wiring baseline carry-forward into mcp_readability. Builds on evalbench/mcp-readability-html-provenance (#697).

When a changed tool is re-judged and a baseline finding for that tool is not repeated in the new response, runs a lightweight second pass (reconcile()) to check whether the schema change actually fixed the issue or the model simply omitted it.

What changed

• evalbench/scorers/mcp_readability/style_readability.py:
  • Add _RECONCILE_PROMPT, _reconcile_dropped(), and reconcile().
  • When cf.unreported_findings(decision, judged) is non-empty, ask the judge for a boolean still_applies verdict per omitted finding_id.
  • Keep findings where still_applies is true (preserving original wording) and drop only findings ruled false.
  • Fail closed on API or JSON parse errors by keeping all findings under review.

Testing

• pytest evalbench/test/mcp_carry_forward_scorer_test.py evalbench/test/mcp_readability_test.py: 59 passed.
• pycodestyle . --config=.pycodestyle: clean.