The nightly found its build with a filtered run listing (branch=main,
status=success) and trusted the first result. GitHub serves filtered
listings from a run search index that has intermittently returned
weeks-old results, so some nights tested a build from weeks earlier and
reported its differences as regressions. The same filter also matched
fork PR builds whose branch is named main.
The build is now picked from the unfiltered listing, which stays
current, and filtered here: a successful build_all run of this
repository on the requested branch. Fork PR builds are excluded by
repository. A feature branch is normally built only for its PR, so this
repository's own PR builds stay eligible, but a PR build compiles the PR
merged into its base rather than the head commit the later jobs check
out, so a push or dispatch build of the branch is preferred when the same
page of the listing has one. A scheduled run fails instead of testing a
build more than 48 hours old, and every run names the build it tested
in the job summary.
Manual runs scan further back, so a branch that last built weeks ago
can still be tested, and a new build_run_id input pins one build_all
run, read directly rather than through a search.
Runs orca-test-repo's full override-sweep effect stage (two shards) and the GUI-vs-CLI parity harness every night against the latest successful build_all Linux AppImage, with sources checked out at that build's commit. Kept out of the per-build regression step, whose time budget it would exceed, and never gates a build.
HanifKoh