You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Flaky: plx_screens tests that build a screen read the Under posters setting without a lock #1178
watchlist::page_tests::a_remove_changes_nothing_on_the_grid_until_the_landing_starts failed in 9 of 60 full plx_screens runs (measured with a local instrumentation patch on main d03081c, so a different tree from main itself) at host load 20-130, with three different assertions: page_tests.rs:824 (the removed title's caption 2002 is not drawn, a bare 2 is), :833 (the drawn words gained a 2 mid-test) and :847 (a card moved 32 px: the caption band went from 80 to 112 px). All are a setting that changes during the test.
Cause: the caption policy is registry::captions::policy(), which in a test is the thread's TEST_UNDER_POSTERS pin, else plx_media::route::under_posters(), a process-wide cell. The Settings under-posters tests write it under testlock::serial(); this test takes neither the lock nor a pin. Deterministic forcing: a second thread holds serial() with {title: true, details: None} while this thread runs the test body unlocked: red with ("2", ..), ("2", ..) where 2002 is expected.
Scope: an audit of readers that hold neither the pin nor the lock finds 83 tests in plx_screens (watchlist 41, collection 30, library 7, home 5).
Fix: make the test-only reader in registry/captions.rs fall back to the default for such a thread (a thread that holds serial() still reads the live cell; a pinned thread reads its pin).
Credit: the structural cause was named by an independent analysis by Claude Fable 5.1; the forcing and the fix are ours.
Thanks for #1189: the flake this issue was filed for (a_remove_changes_nothing_on_the_grid_until_the_landing_starts) is closed by the PolicyPin on the Watchlist harness. About 40 other plx_screens tests (collection, home, library and a few direct WatchlistScreen::new ones) still read the process-wide Under-posters cell with neither a pin nor the lock, but none of them failed in 60 full runs, so I would close this as fixed by #1189 unless you want the whole class closed in the shared reader (I closed #1183 and can rebase the small change on request).
Activity
Bardioc1977 commentedon Nov 2, 2026
Thanks for #1189: the flake this issue was filed for (
a_remove_changes_nothing_on_the_grid_until_the_landing_starts) is closed by thePolicyPinon the Watchlist harness. About 40 otherplx_screenstests (collection, home, library and a few directWatchlistScreen::newones) still read the process-wide Under-posters cell with neither a pin nor the lock, but none of them failed in 60 full runs, so I would close this as fixed by #1189 unless you want the whole class closed in the shared reader (I closed #1183 and can rebase the small change on request).