[llb555]

test_bench_sandbox_prewarm.py::test_prewarm_reused_survives_a_coarse_clock_tie fails at random on busy machines, including in CI — it just took down the Unit Tests shard 4 run for #7505, and it reproduced 2 of 3 local runs on a tree that did not touch it.

The cause is the setup, not the assertion. The test quantizes time.monotonic() to 15 ms and then spins until a fresh bucket starts, relying on the whole overlap scenario — several awaits, a scheduled task, and two events — finishing inside that remaining tick. When the machine is slow, the tool request lands in the next bucket and the precondition assert fires before the actual regression is even exercised.

This replaces the race with a frozen clock:

    # Freeze the quantized clock instead of racing a real bucket: the scenario
    # between the acquire and the tool request takes several awaits, so waiting
    # for a fresh tick and hoping both land inside it made this test fail
    # whenever the machine was busy. A constant reading still reproduces #7044
    # exactly (acquire and tool request share one timestamp) and cannot flake.
    frozen_tick = math.floor(real_monotonic() / coarse_s) * coarse_s

    def coarse_monotonic():
        return frozen_tick

The test still pins #7044: with the original strict timestamp comparison restored in bench_prewarm.py, it fails (assert False is True on prewarm_reused), because equal timestamps never satisfy <.

Verified: pytest tests/test_bench_sandbox_prewarm.py → 19 passed, five consecutive runs (previously 1 of 3 passed on the same machine); ruff check and ruff format --check clean.

[WillemJiang]

The issue has been fixed by #7513