test(plugins): stop the update check reaching the live registry mid-suite

[prokopto-dev]

What was failing

Semantic Release has been failing intermittently on master with exit code 139 — the Verify release candidate step (uv run pytest) segfaults partway through tests/ui, so Calculate version, update changelog, commit, tag, and push never runs. Two of the last four runs died this way (31349283155, 31349817083). The release tooling is fine; the test suite was crashing.

Why

tests/ui/test_plugin_window_rows.py::make_ui calls build_plugin_ui with default settings, which arms the post-launch plugin update check on a 12 s QTimer. Six tests, six timers, and nothing cancels them when the test that created them ends. They fire a quarter-minute later — inside whatever test is running by then — each starting a daemon thread that fetches the real registry over HTTPS.

CI's fault traceback caught it precisely there: five live threads in socket.getaddrinfo while the main thread sat in tests/ui/test_pluginmanager.py's fixture, garbage-collecting the previous test's Qt objects.

Thread 0x…  socket.getaddrinfo → httpcore connect_tcp → httpx stream
            → install.fetch_https_bytes → registry.fetch_indexes
            → updatecheck.check_for_updates → pluginbootstrap.py:171 in work
Current thread …  Garbage-collecting
            → composition.build_backend → tests/ui/test_pluginmanager.py:44 in host

Live sockets crossing another test's Qt teardown is what segfaults. Whether it lands depends on suite timing, which is why this only started biting as the suite grew — tests/ui now runs long enough for the timers to fire inside the session rather than after it.

Reproduced locally by running test_plugin_window_rows.py followed by a test that spins the event loop for 16 s: plugin-update-check threads appear after the tests that armed them have finished. With this change, zero.

The fix (tests only)

• test_plugin_window_rows.py turns update_check off — the file is about Settings > Windows rows and never wanted the check.
• tests/conftest.py grows a session-scoped guard that refuses any fetch_https_bytes call arriving without a transport. Session-scoped because the gap between tests is exactly where a leaked timer fires, and a function-scoped patch would already be unwound. The transport seam stays open, so patch_default_transport and the redirect/size-limit tests are untouched.

Tests

• QT_QPA_PLATFORM=offscreen uv run pytest → 1994 passed, 2 deselected, and ~20 s faster (56 s vs 78 s) now that it no longer waits on live DNS.
• uv run ruff check . / ruff format --check . clean.
• Verified the guard bites: a registry fetch from inside tests/ now raises could not reach the registry: the test suite must not fetch … for real. Before this branch, that same call fetched the live index successfully — the suite really was talking to the internet.

Risk / notes

No product code changes, so test: — correctly no version bump. Merge with a merge commit so the commit survives for versioning.