/** * Replays the labeled test-log cases through the admission arms and reports, per arm: * required evidence kept, wrong omissions, optional noise omitted, net characters saved * (markers and pointer paid), judge calls, latency, input tokens, and what a wrong * omission would cost to recover from the archive. * * node kyrn/spikes/judge-bench/test-log-replay.ts offline: full, rules, ceiling, oracle * KYRN_JUDGE=jev node --env-file=.env kyrn/spikes/judge-bench/test-log-replay.ts adds the live judge * * Offline reference arms use mock judges and cost nothing: * - `ceiling`: a judge that always says "omit". What the protections alone guarantee, and * what a goal-blind filter (failures only, passing collapsed) would do to each goal. * - `oracle`: a judge that reads the labels. The best a correct judge could do. * * Options: SET=tuned|heldout|all WORDINGS=suffices,policy ANCHORS=on|off|both MAX_REQUESTS=150 * Only the synthetic fixtures are sent. Needs Node >= 23.6. Writes results/test-log--