review_replay: optional judged recall over the credited hits #291

Problem

dokimi:review_replay credits a human review thread to the nearest finding on the same file within window lines, one to one. It checks location only, on purpose: the grader is deterministic and runs on the image's standard-library python3 (contract §4.5). So a finding that merely sits near a thread gets credit.

The dokimi report on a replay of IntelPython/dpnp#3099 (26 tasks, 156 runs, 210 credited thread hits) quoted credited threads as evidence of what the guidance catches. An audit read the pairs behind them:

• pr3098-t2: the thread asks to add a TODO. The credited findings complain that the TODO is repeated five times and has no issue URL. pr3055-t9 and pr3128-t1 look similar.
• pr3066-t1: one of its two credited findings is about something else (deprecating before removing).

Chance credit also grows with the number of findings. The guidance arm reported more findings (5.26 vs 4.37 per review), so location-only recall can favour it. A list of "threads the guidance made the agent catch" needs each pair read.

Proposal

An optional step after grading, outside the grader, for example dokimi repo hits judge <batch> --judge-model <id> --judge-provider <p>:

• Input: for every credited pair it reads the thread's comment and the finding's text. They come from grade.json details.matches, which a separate change adds.
• Judgment: a judge model labels each pair same point, related or unrelated. The call goes through dokimi.llm, as the architecture requires.
• Output: a hits_judged.json per cell, with each label and recall_judged: the threads whose pair is same point, over all gold threads. repo compare reads it as a measure, for example --metric judge:recall_judged, next to recall.
• What it can and can't fix: only credited pairs are rescored. So recall_judged ≤ recall, and a finding outside the window that does raise the thread's point stays uncredited.
• Cost on that batch: 210 pairs, roughly $0.5 with a small judge or $3 with a mid-size one.

Open questions

• Judge model: a different family from the model under test, as novelty requires? (arms run does not check.)
• Calibration: a person labels a sample of pairs, and the step reports agreement before any judged figure is quoted. How large a sample?
• Uncredited findings: should near-misses (a finding on the same file outside the window) be judged too, so judged recall can also rise?
• Ties: three labels or two?

Done when

The command runs over a batch with a fake LLM in the tests and writes hits_judged.json per cell. repo compare reads recall_judged. The docs (contract §4.5 and §9) say what judged recall is and what it is not.