fix(python): stop teardown test flaking on unrelated thread reaps

[claude]

What changed

test_threads_joined_deterministically_after_drop no longer asserts an exact process-wide thread count. The check is now _assert_no_thread_growth(baseline) — one-directional, matching what the sibling test_fds_stable_across_tool_churn in the same file already did with <= baseline.

Two tests pin the direction: an inflated baseline must pass, growth must still fail.

Why

The test failed intermittently and blocked unrelated PRs — assert 13 == 14 on #2206, assert 15 == 16 on main at 082447e3 (a dependabot actions bump). Note the direction: the observed count was below the baseline. Nothing leaked; a thread went away.

The cause is tokio reaping idle blocking-pool threads on a 10 s timer. A pytest session that ran async work before this module sheds an unrelated thread part-way through the churn loop. Measured directly rather than guessed:

$ # sample /proc/self/task every 500ms for 20s after an async workload
thread-count transitions over 20s after an async workload:
  t= 0.0s  threads=6
  t=10.0s  threads=5      <- idle blocking-pool thread reaped

If baseline is captured inside that window, the reap lands mid-loop and exact equality reports a defect that did not occur.

TM-PY-030 is about what a drop leaves behind, and the process-wide count is not a single tool's to own. A genuine leak compounds across the five churn iterations and still pushes the count above the baseline, so leak detection is unchanged.

Before / After

The new tests discriminate — the same tests against the old == form vs. the fix:

# old (== baseline)
$ pytest test_teardown_determinism.py -k thread_check
1 failed, 1 passed, 6 deselected in 0.03s

# new (<= baseline)
$ pytest test_teardown_determinism.py -k thread_check
2 passed, 6 deselected in 0.02s

Full Python suite, four consecutive runs:

run 1: 742 passed, 2 skipped in 13.02s
run 2: 742 passed, 2 skipped in 13.45s
run 3: 742 passed, 2 skipped in 12.88s
run 4: 742 passed, 2 skipped in 12.33s

just pre-pr green, including cargo vet. ruff check and ruff format --check clean.

Risk

• Low
• Test-only change plus a knowledge note; no library, binding, or runtime code is touched.
• The assertion is weaker in one direction by design. The direction it drops — a decrease in process-wide threads — cannot indicate a teardown defect, and the retained direction is the one TM-PY-030 cares about.
• knowledge/python-package.md records the measurement and an explicit "do not re-tighten", so the equality check does not come back.

Checklist

• Tests added or updated — two new tests pinning the invariant's direction, both verified to fail against the previous assertion
• Backward compatibility considered — test-only; no public surface

---

Generated by Claude Code

[cloudflare-workers-and-pages]

Deploying with    Cloudflare Workers

The latest updates on your project. Learn more about integrating Git with Workers.

Status	Name	Latest Commit	Preview URL	Updated (UTC)
✅ Deployment successful! View logs	bashkit	5ed0c54	Commit Preview URL Branch Preview URL	Jul 26 2026, 05:40 AM