Four brain tests fail on master independent of any branch: test_alert_correctness (humidifier/heater ineffective) and test_stale_hub_alerts (live hub still rais

CLOSED 2026-09-14 20:50: not reproducible on master 1fe2a69. The two files pass in isolation (17 passed) and inside the full suite (928 passed, 172 s), and Brain CI on master is green for the #267 and #268 merges. No commit between 6cda2d3 and 1fe2a69 touched alerts.py, dash_computed.py, the two test files or conftest.py, and neither test reads the wall clock, so the earlier red was almost certainly that worktree’s environment (a stale DSC_DATA / ops DB carried in from the shell) rather than a regression. Reopen with the captured failure output if it recurs. — ORIGINAL: Found running the full suite for PR #259 (848 passed, 5 failed; the fifth was the branch’s own API-segment omission, fixed). Re-ran the four in a detached worktree of untouched master (6cda2d3): they fail there identically, so they were red before this work. They all assert binary_sensor.dsc__ineffective_suspect turns on after ~10 min of demand with no response; the flag stays off. Likely a recent change to the ineffective-appliance hold (_held_flag / _INEFFECTIVE_ON_S) or to how the tests seed demand history. Needs a bisect; CI on master is red until then, which is also why PR merges show UNSTABLE.