v0.11.0

Features:

• --rejudge on eb run. Re-score every cached Claude output with the configured judge — keeps run outputs verbatim, drops existing judgments, fires no Claude invocations. Answers "did changing the judge change the verdict?" without paying for a full re-run. Composes with --baseline-from and --current-from: stitch two saved snapshots and re-judge in one command (eb run --baseline-from base --current-from cur --rejudge --save-as merged produces a dual-variant snapshot with zero Claude calls and fresh judgments on both sides). Mutually exclusive with --force (which wipes runs) and --retry-failed (which targets failed Claude rows, not successful judgments). On an existing complete snapshot, --rejudge resumes it instead of erroring "already exists."

Fixes:

• Matrix dedup now re-judges rows where a Claude run succeeded but a judgment is missing. The third dedup branch previously skipped these rows alongside fully-done ones, so manually editing snapshot.json to clear judgments left rows unscored. With this change, the same dedup naturally implements --rejudge: any row with a successful run and no (or errored) judgment is routed through the judge, and any row with a successful run and a successful judgment is still skipped. Behavior on snapshots produced by older versions is unchanged because runBenchmark always wrote a judgment for every successful run.