fix: keep the current plan when LLM refinement drops hunks or does not reduce violations

[vitali87]

Summary

Each LLM refinement iteration replaced the plan with whatever the model returned. There was no coverage check and no comparison with the plan being refined, so a response that dropped a hunk turned a plan that only had LOC warnings into one that fails validate_plan with COVERAGE_GAP (the run aborted), and a response that was no better was accepted anyway.

Now a candidate is kept only if it passes validate_coverage and has strictly fewer LOC-bound violations than the plan it replaces. Otherwise the current plan is kept and refinement stops, with a warning naming the reason:

• Refinement iteration N produced an invalid plan (Hunk b.py[0] not assigned to any group); keeping the current plan …
• Refinement iteration N did not reduce violations (2 -> 2); keeping the current plan

Stacked on #63 (stack #75 = #59→#63→#121); #59 provides the LLMError wrapping of malformed LLM output that this relies on.

Test plan

• test_refinement_that_drops_hunks_is_rejected and test_refinement_that_does_not_improve_is_rejected — both fail on the base
• Review probed the overlap path, a multi-iteration improve-then-stall sequence (keeps the improved plan), and the chunked path
• 451 tests pass, ruff clean
• Local review: 5/5

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on base/target branches other than the default branch.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro Plus

Run ID: 76778256-b0e7-45ad-bf36-7045b9e47cc9

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}