Implement BaseReviewer.review(case_context) -> ReviewerResponse and register the adapter. The reviewer receives the diff and contextual evidence, never ground truth.

Patch-aware response

Responses must be valid JSON without Markdown fences. Natural-language fixes remain supported, but patch/full mode only validates repair outcomes when a unified patch is supplied.

Code
{
  "findings": [{
    "title": "specific production bug",
    "summary": "what is wrong and why it matters",
    "category": "correctness",
    "severity": "high",
    "file": "path/to/file.py",
    "line_start": 1,
    "line_end": 4,
    "evidence": "the code or behaviour that shows it",
    "confidence": 0.91,
    "suggested_fix": "natural language repair",
    "suggested_patch": "diff --git ...",
    "patch_confidence": 0.88
  }],
  "overall_risk": "high",
  "review_summary": "short summary"
}