ci: run property-based tests nightly with a large max_examples

[TaiSakuma]

As discussed in #3962 (comment): keep Hypothesis max_examples small for PR CI, and run the property-based tests on a schedule with a large max_examples. The bug fixed in #3962 was within reach of the existing property tests but was only stumbled on locally — PR-level example counts are too small for discovery, and raising them for every PR (and every matrix entry) would be too slow.

What this does

• Hypothesis settings profiles in tests/properties/conftest.py replace the hardcoded per-test @settings(max_examples=...):
  • default: max_examples=200 — used by PR CI and local runs (~4 s for the current suite, measured ~5.4 ms/example).
  • nightly: max_examples=10_000, deadline=None, print_blob=True, derandomize=False — ~1 min/test locally, an estimated ~6–9 min on a CI runner.
  • Selected via the HYPOTHESIS_PROFILE environment variable. (The --hypothesis-profile CLI flag can't be used: profiles registered in a sub-conftest aren't visible at pytest_configure time.)
  • derandomize=False is explicit because on GitHub Actions Hypothesis auto-loads its built-in ci profile (derandomize=True), which the nightly profile would otherwise inherit — and then every nightly run would test the same 10,000 examples. PR CI stays derandomized, same as before.
• New scheduled workflow property-tests.yml (daily at 3:34 UTC + workflow_dispatch):
  • Builds/caches the awkward-cpp wheel with the same steps and pinned action SHAs as reusable-test.yml, then runs pytest tests/properties at the nightly profile.
  • A second pass runs under pytest-run-parallel --parallel-threads 4 at the default profile — this is how the bug behind fix: proper comparison in ak.almost_equal between record array fields #3962 was originally found, and the property tests currently never run threaded in CI (the 3.14t matrix entries use requirements-test-nogil.txt, which doesn't include hypothesis-awkward). It runs even if the serial pass fails, so one night reports both results.
• README badge for the new workflow, so scheduled failures are visible. While editing that line: the existing "Build Test" badge has pointed at the deleted test.yml (HTTP 404) since the reusable-workflows split in ci: use reusable workflows #3969 — it now points at ci.yml.

Net effect on PR CI

test_to_from_buffers.py::test_roundtrip drops from 1000 to 200 examples in PR runs; the array_equal tests stay at 200. Deep exploration moves to the nightly run, which uses fresh random seeds each night (~3.6 M examples/test/year cumulatively).

Test plan

• pytest tests/properties passes at the default profile (3.74 s vs 7.99 s before — confirms 200/test)
• HYPOTHESIS_PROFILE=nightly runs 10,000 examples (56 s for one test)
• An unknown HYPOTHESIS_PROFILE value fails loudly (InvalidArgument)
• pre-commit green (incl. workflow schema validation and zizmor); actionlint reports only the same info-level SC2086 already present in reusable-test.yml
• After merge: smoke-test with gh workflow run property-tests.yml (the cron only fires from the default branch)

Possible follow-ups (not in this PR)

• Persist a Hypothesis example database across nightly runs so past failures are replayed first (currently database=None via the inherited ci profile; print_blob=True already gives @reproduce_failure blobs for local reproduction).
• Run the threaded pass on free-threaded 3.14t for more aggressive concurrency coverage.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 85.31%. Comparing base (b025b0b) to head (f8e1310).

Additional details and impacted files

see 1 file with indirect coverage changes

[ianna]

@TaiSakuma - Thanks! It looks good to me. Please go ahead and merge it if you are done with it. Thanks.

[ikrommyd]

I'm not particularly good with how conftest works in this case. So this does not increase the max_examples when we run the whole tests right? And only increases it in this particular property based testing CI?

[ikrommyd]

Yeah I'm pretty sure it does what my above comment says. Good