error_later is silently dropped when a test wraps warnings.showwarning #15139

Written by Claude Opus 5.5 via Claude Code for the pytest maintainers; I prompted it, it did the work, I read it.

Follow-up to #15135, which adds the error_later warning filter action.

If a test replaces warnings.showwarning with a wrapper that forwards to the original, as numpy.testing.suppress_warnings does, a warning matching an error_later filter is shown as a plain warning and the test passes. -W error fails the same test.

import warnings

def test_it():
    orig = warnings.showwarning

    def fwd(*a, **k):
        orig(*a, **k)

    warnings.showwarning = fwd
    try:
        warnings.warn("w", UserWarning)
    finally:
        warnings.showwarning = orig

• pytest -W error_later::UserWarning: 1 passed, 1 warning
• pytest -W error::UserWarning: 1 failed

Real-world shape: numpy 2.4.1's suppress_warnings forwarding a UserWarning behaves the same. numpy deprecated suppress_warnings in 2.4, so the exposure is shrinking.

Cause

The stdlib decides whether an error_later filter applies: a probe object in the filter's module slot records a verdict when the filter wins. To avoid leaving a verdict behind for warnings that something else takes (pytest.warns, the test's own catch_warnings), the probe records only when the warning will be shown to pytest's recording directly. A forwarding wrapper puts itself in front of that recording, so the probe records nothing, and the warning reaches pytest without its verdict.

Possible fix (untested)

Record the verdict whenever the probe matches, keyed by module and message text. pytest's recording, or its showwarning wrapper, honours the verdict only if the text matches. A probe at the front of the filter list clears stale verdicts.

• Cost: one string compare per recorded warning, plus one extra filter entry.
• Residual risk: two warnings with identical text could be mismatched.