Repository navigation
Conversation
|
The diagnostic proof is complete at The pinned candidate All 60 cost records and raw outputs were independently validated against executable identities and ZIP hashes. Linux GC-loop elapsed medians: S8 191.046733 ms, candidate 193.006933 ms, no-op 189.9641895 ms; paired candidate/S8 ratio 1.0079996. Worker CPU medians: S8 168.570 ms, candidate 170.390 ms, no-op 170.584 ms; paired ratio 0.9987887. Both calibrated limits were 1.02 and passed. This PR is diagnostic scaffolding only and is being closed unmerged. The real fix remains draft in #209; its separate macOS cost cohort failed the GC-loop gate. No readiness or merge is claimed. |
Diagnostic proof only. Do not merge. This branch will be closed after its artifacts are retained.
Runs the frozen cost benchmark and unchanged worker/timer tests on native Linux x64 for #209. Candidate source and portable executable are pinned to
f08cee251d437eefb2f5c419d3b1ec0e1d0ccf85, release-preview run40174160677. The original S8 executable comes from run40153838318. Archive hashes, executable revisions, engines, architecture and source hashes are verified before measurement.Ten alternating pairs per workload retain every result. The S8 acceptance limits remain1.02 for paired GC-loop elapsed and worker CPU time. If acceptance fails, a separately recorded same-engine unpatched-source control runs; it cannot override the original failure. Then the original immediate-GC regression runs ten times and the22surrounding files run once each. No test source, deadline, retry, runtime code, or Windows job is changed here. Workflow permissions are read-only.
The earlier run on the superseded empty-helper candidate is retained, not rerun unchanged or replaced. The current real-fix helper was inspected in its portable Linux/macOS binaries: zero stores are present and profile/stripped text sections match. Python/YAML syntax, whitespace and Codex P2 scoped reviews passed. This temporary PR does not replace the fix PR's own native CI.