fix(capture): let the capture fix decide for itself whether pytest needs it

[livingstaccato]

The capture patch is a stopgap for a pytest defect, and it carried no way to stop.

Why

install_capture_swap_fix guarded only on the two methods still coming from _pytest.capture. That catches another patcher and nothing else. A pytest that fixes this upstream keeps the methods exactly where they are — so the guard passes, the patch lands on the corrected implementation, and the sys.stdout reassignment goes back into a capture that had stopped doing it. Silently.

Upstream has now said which way it is going: pytest-dev/pytest#14996 drew We should use a distinct stream instead of this ai confusion. Work on distinct streams is in progress. So the patch is unlikely to be adopted as written, and likely to be obsoleted by a reshaped capture. Both outcomes need this.

What changed

• Installing measures rather than infers. A real capture is driven through swap, suspend and resume; only a pytest that loses the swap is patched. A version floor cannot answer that question — a backport, a fork, or a vendored copy each answer for themselves. A probe that cannot be driven reports no defect: patching a class whose behaviour was never established is how a fix becomes an outage.
• The _pytest.capture import moved inside the call. It sat at module scope, and pytest_plugin imports this module while loading, so a renamed or removed class would raise ImportError inside plugin load and take every suite in the family down on a routine pytest bump. Now its absence costs the fix and nothing else.
• The install marker is read from the class's own __dict__, so a subclass cannot report a fix that is not on the class in hand.

Tests

Four added, all failing before the change:

• a pytest that already preserves the swap is left alone
• a pytest that still drops it is patched
• a pytest without the class returns False instead of raising
• the probe gives back the stream it found

The existing seven are unchanged, including the two pytester subprocess runs that drive a real CliRunner under log_cli=true — one asserting the fix works, one asserting the failure returns with -p no:provide_testkit.

391 passed, 1 skipped single-process; ruff, ruff format and mypy strict clean.