[MaxGhenis]

What fails

On #574, one Tests shard fails at
tests/test_track_b_b0_2_counts.py::test_a_failure_from_outside_this_script_prints_no_value.
The test asserts that capfd holds ("", "") after script.main stops, and
it finds pytest-split's header instead:

At index 0 diff: '\n\n[pytest-split] Splitting tests with algorithm: duration_based_chunks\n[pytest-split] Running group 3/4 (estimated duration: 3249.47s)\n\n' != ''

Why

• pytest-split 0.11.0 writes its header with self.writer.line(...), a
  terminal writer over sys.stdout (pytest_split/plugin.py:124 and
  :171–179). When stdout is a pipe, as in CI, that text waits unflushed in the
  buffer of sys.__stdout__.
• The B0.2 script's _flush_streams (scripts/track_b_b0_2_counts.py:2340)
  flushes sys.__stdout__ and sys.__stderr__ before _silenced redirects
  fds 1 and 2.
• With -q, as CI runs pytest, nothing flushes that buffer before the first
  test's call report (_pytest/terminal.py:699). At default verbosity the
  test passes at master with pytest-split.
• So when this test runs first in its shard, the script's flush pushes the
  header into fd 1 while capfd holds it, and the test captures text it never
  wrote. In Aggregate-table captures and readers, part 1 (D4) #574's run (job 125184609985) it ran first in shard 3 of 4: that
  shard's progress line opens with F. Which test runs first depends on the
  shard split, so another change to the test set can move this one back to the
  head of a group.

Fix

An autouse fixture in the module flushes sys.__stdout__ and sys.__stderr__
before each test starts capturing. The script and its assertions don't change.
The fixture adds no tests, so the tier files don't change.

Checks run

• Real pytest-split 0.11.0 with the test first in group 2 of 2 fails at
  master's test file with the header diff above, and passes with the fixture.
• A minimal plugin that writes to sys.stdout at collection without flushing
  fails the test at master and passes with the fixture.
• Without any plugin the test passes at both.
• Full module run, serial: 247 of 248 passed under taskpolicy -b. The one failure, test_the_host_reads_see_a_background_policy_and_a_qos_clamp[policy1-qos], comes from that wrapper: it fails the same way at master under taskpolicy -b, and it passes at both master and this branch in a plain run.

Review

An independent Claude Opus 5.5 review returned APPROVE. It reproduced the
failure and the fix under real pytest-split, found the same latent failure in
test_relabeling_a_death_and_a_move_changes_no_written_byte[ages1-before1-0-7-released]
when that test runs first, and showed that the fixture runs before capfd
starts: moved after it, the target fails again. Five injected leaks from the
script still fail the target with the fixture in place. This PR's own shard
split doesn't put the target first, so its CI doesn't exercise the fix;
#574 does once it merges this.

🤖 Generated with Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
social-security-model	Ready	Preview	Nov 2, 2026 2:52pm UTC

[MaxGhenis]

The Microcosm Dynamics coordinator merged this PR at 562d0d7087e43e07c5af4de4348261de82cabfad with --merge --match-head-commit, as dd9c7be24b7228c17123a44d1fd1150d4335312f (16:52Z, 2026-11-02).

• Review. An independent Claude Opus 5.5 review approved 562d0d70, after reproducing the failure and the fix under real pytest-split 0.11.0.
• CI. gh pr checks 596 exited 0 at the head, with 11 checks passing and deploy skipped (Tests run 41466012683). GitHub reported the PR mergeable.
• Late writes. The target test still misses a value that the script writes to sys.__stdout__ after its last flush and leaves in the buffer. That text reaches the terminal after capture ends. The review found this with an injected leak; at master, capture misses that text too.
• The slot. Flush the real streams before each B0.2 test captures output #596 landed after Federal SSI interface and source-to-input contract (SSI0) #573 (89d064b5). Aggregate-table captures and readers, part 1 (D4) #574 merges master dd9c7be2, or later, and recounts before it lands.