[bbfrederick]

The measurement

One job on ubuntu-latest / Python 3.14 was 1817 s, split almost exactly in half:

step	secs	%
Test (pytest -q -n auto)	940	51.7
The .exar1 archives must be present and exercised	835	46.0
everything else	42	2.3

That second step ran pytest tests/test_exar.py tests/test_exar_patch.py serially, and the Test step then re-ran all of it in parallel. It tested nothing the Test step did not — it existed only to tee its output and grep -q skipped, because xdist workers interleave stdout.

test_exar_patch.py carries the two corpus sweeps (400 s and 199 s), so that one guard was carrying nearly the whole cost. The OCR and front-end guards do exactly the same thing for 4 s and 3 s, because their files are small.

The change

Test now writes a JUnit report, and one new step asserts from it that each guarded group ran. The report is written by the xdist controller and aggregates the workers, so it carries per-test skip status that interleaved stdout cannot — which is precisely what let the serial runs go.

Every environment-discovery check is untouched and still fails fast: tesseract must be findable, node --version must work, EXAR_FILES must be non-empty. Only the "and were then exercised" half moved to after the suite.

It is stricter than what it replaces

grep -q skipped ocr.log asks whether a file's output holds that word anywhere. The new guard names the tests that must have run, counts them, and reports which ones skipped.

Against the real full-suite report it sees 3 OCR, 12 front-end and 709 .exar1 tests ran, none skipped, and correctly ignores the suite's 26 legitimate skips, which are all in tests.test_scans and outside the guarded groups.

Verification

• xdist + --junit-xml records skips correctly — checked rather than assumed: under -n 4, the XML carries a per-test <skipped> with its reason plus a suite-level count.
• The guard, extracted verbatim from the workflow, falsified four ways: healthy → exit 0; exar skipped → exit 1 naming the tests; exar not collected at all → exit 1 with a distinct message; one OCR test skipping among several → exit 1 naming it.
• Run against the real 3878-case report from a green local suite (3852 passed, 26 skipped) → exit 0 with the counts above.
• Emits ::error:: so --log-failed surfaces the one useful line.

Expected effect

~835 s off each of the six matrix jobs — ubuntu/3.14 from 1817 s to roughly 985 s. This PR's own CI run is the measurement; a workflow change cannot be checked any other way, so compare its step timings against the numbers above before merging.

Not addressed here: the Test step's own 940 s, of which the single 400 s self-drive sweep is 42%. Both big sweeps loop EXAR_PROTOCOL_FILES internally over 26 evenly-sized pairs, so parametrising them would let xdist distribute the work — but they assert global tallies for non-vacuity, and splitting them needs a separate pin on the pair count or a corpus that shrank to zero cases would pass as 0 passed.

.gitignore swaps ocr.log, which nothing writes now, for results.xml.

🤖 Generated with Claude Code