tests/test_streaming_stripper.py:468: in test_early_markup_is_not_slower_than_the_code_it_replaces
E   AssertionError: early markup cost 1.451s against the reference's 1.316s
E   assert 1.4506334609999385 <= (1.3157524960000728 * 1.1)

Cause

The test compares how much work two code paths do, with a 10% margin:

count = 1500
reference = min(elapsed(_reference_strip, count) for _ in range(3))
incremental = min(elapsed(StreamingMarkupStripper(ENABLED).strip, count) for _ in range(3))
assert incremental <= reference * 1.10

elapsed used time.perf_counter, which measures wall clock, which also measures whatever else the machine is doing. A 10% margin does not survive that. On a 4-vCPU runner with other pytest workers in flight it fails while the code under test has not changed.

time.process_time measures this process's CPU time. That is the quantity the assertion is actually about, and neighbours do not move it. The 10% margin is untouched and both min-of-3 samples stay.

What is and is not verified

I have not reproduced the original failure locally, so I cannot claim this is verified as a fix for that specific run. Three busy processes pinned to the same core slow both arms equally, so the ratio holds and both clocks pass (5 runs each, all green).

What is demonstrated is the mechanism, on identical work:

Wall clock inflates 3.7x under contention. CPU time does not move.

Whole file: 64 passed.

Which PR, and which side is wrong

Introduced by #8538 (2026-08-12), which added this test using time.perf_counter. The source it guards is fine and is untouched here; the test has measured the wrong quantity since it was written.

Robustification

process_time is coarser than perf_counter, and 0.0 <= 0.0 * 1.10 is true. A stopped clock, or a count someone lowers, would silently turn this into an assertion that cannot fail. Both arms now have to measure something, with a 0.05s floor that sits far below the ~1.3s each arm actually takes and far above the clock's resolution. Mutating elapsed to return 0.0:

AssertionError: reference arm measured 0.0000s; too small to compare against

Found while measuring whether the backend suite can run under pytest-xdist. This is the one failure in that set that is not an order dependency: it is a timing assertion that cannot share a machine, and I would rather fix the clock than widen the margin.