test(fleet): wait on conditions instead of fixed sleeps in TUI poll tests

[Jason Robert (jrob5756)]

Problem

Main's CI has been failing intermittently. The failures are flaky tests, not a product regression — src/ needs no change.

Two failures on main, both in the Fleet Manager TUI pilot tests:

Run	Test	Failure
32296506434 (3.13/ubuntu)	test_a_scan_failure_does_not_wedge_the_screen	_refreshing never cleared
32272384555 (3.13/windows)	test_poll_tick_removes_completed_run	assert 1 == 0

The first also failed on 32044377141 (Aug 17), in its pre-#464 single-sample form — so this is a recurring flake a previous fix attempt did not close.

Root cause

These tests assert on state that only becomes true once a realset_interval poll tick lands, but wait a fixed asyncio.sleep(0.3) or sample a transient flag. CI runs pytest under coverage tracing on a small shared runner, where a tick-plus-scan is several times slower than on a developer box, so the fixed wait encodes a guess about machine speed and asserts against pre-tick state when the guess is wrong.

The guard-recovery test is the more interesting one. Instrumenting it:

• pilot.pause() costs ~1s per call under the tests' 20 Hz timer (it waits for the app to go idle, which a 0.05s poll timer keeps deferring), so the helper's 5s deadline yielded only ~4 samples.
• _refreshing is legitimately True19.5% of the time idle and 34.8% under coverage.

So the assertion was a coin flip that a slow machine loads — it can see True on every sample and report a wedged screen. The finally in _refresh_worker that it guards is correct and present.

Fix

• Add wait_for(pilot, predicate, message=...) to tests/test_fleet/conftest.py: polls a condition to a deadline, samples with asyncio.sleep rather than the ~1s pilot.pause(), and fails with a message naming the regression instead of an opaque assert 1 == 0.
• Rewrite the two guard-recovery tests to assert a monotonic, user-visible consequence — a newly written record appears in the table — rather than sampling a transient flag. That cannot be missed by any sampling rate, is unreachable if the guard latched, and proves recovery all the way to rendering.
• Convert the affected poll sites in test_tui_runs.py and test_tui_run_detail.py.
• Fix two latent siblings of the same bug: an assert call_count == 1 that could observe 0 before the mount refresh had entered the collector, and a selection-preservation test that passed vacuously when no tick landed inside the fixed sleep.

Verification

• Mutation test — removing the finally in _refresh_worker makes both rewritten tests fail deterministically in ~2s with the intended message (no scan ran while the failure was injected -- the screen was already wedged...). The new assertions are not vacuous.
• Contention test — 2 contended CPUs + coverage (~6x slowdown, worse than CI): reworked tests 8/8 pass; the previous versions reproduced the failure 1/8 under identical conditions.
• Full CI-equivalent suite (-m "not real_api and not performance", with coverage): 7426 passed, 60 skipped.
• make lint and make typecheck clean.
• Bonus: the affected tests run ~2x faster, since they now return as soon as the condition holds instead of sleeping fixed durations and calling the slow settle().