Independent checker allocation correction

The corrected analysis completed in 28.918 seconds with sampled peak RSS 232,996,864 bytes. Its scientific contrasts and input hashes are unchanged. The first independent checker passed fold 0 then stopped after 4.169 seconds at 312,836,096 bytes sampled RSS. Preserve artifacts/ablation_verification_execution2_20261014. It is incomplete and does not verify the study.

Release the checker’s feature arrays and close archives at each fold boundary; retain its independent algorithms and all tolerances. No scientific result or threshold is changed. The provisional analysis output has been inspected at this point; this is a disclosed post-read implementation correction. Freeze a separate verification registration binding the actual corrected checker, analysis registration, complete analysis checkpoint and failed verification evidence. The original checker hash is checked against Git snapshot c28ecc40, rather than pretending that its mutable path still contains those old bytes. New verification uses a fresh stage artifacts/ablation_verification_execution3_20261014. Keep the unchanged 256 MiB cap, 530 s ceiling, combined checker-stage 4 MiB cap and total 1,740 s ceiling.

A second verification attempt also exceeded the same ceiling after fold 0, 4.166 seconds, sampled RSS 300,564,480 bytes. Releasing fold arrays was insufficient for the checker’s allocation pattern. Preserve execution3. The checker now also constructs one method at a time, using its own independent transform, grouped-mask encoding and direct-distance replay (no worker imports). Freeze execution4 before rerun. This is a further disclosed allocation-only correction, not a new ablation, threshold relaxation or independently unseen result. Record all attempts.

Execution5, after the numerical correction, passed all five folds and the corrected metrics/selections, then stopped at 18.589 s / 317,390,848 sampled RSS while reading the entire expanded source CSV. The final checker streams this ledger and reads compressed prediction arrays once, retaining its exact-byte row comparison. No scientific calculations or tolerances change. Preserve execution5, bind the original checker input against Git 9ac5cb76 and the actual current checker to the separate execution6 registration; verify every row before concluding. The original memory, aggregate disk and total time ceilings stay fixed. This final allocation correction is post-read and disclosed, not fresh independent physical validation.