mirror of
https://github.com/stablyai/orca.git
synced 2026-10-03 16:02:11 +00:00
A path-filtered PR still posts three green meta jobs — the path classifier, the root-directory guard and the LoC counter — and the usability test was "did any check run", so two of them counted as a full corroborating witness pair. Real payloads from #17002 (docs-only) and #16917 (mobile-only), 2.2h apart in two stacks, returned `clean` for a window in which no test, typecheck or static-analysis lane had run at all. A witness now has to have run a lane that would have gone red if `main` were broken. That list is an allowlist, so an unrecognised name costs a witness and degrades the verdict to `unknown` rather than buying one and pushing it to `clean`; every unrecognised name is printed on each run. Also groups two branches that share an unlisted parent as one stack. Grouping only by listed parents counted them as two independent stacks, so the corroboration the `broken` and `upstream` verdicts rest on was one witness counted twice. The doc claimed the first property the code did not have; corrected.