Five commits in the fixture left a ranking test comparing string lengths

The x625.0 in that CI log is a claim about the Windows scheduler, not about the redaction routine ^[15]. The line doing the dividing is `ratio = big / max(small, 0.05)` ^[16]. On Windows, `process_time()` advances on the scheduler tick, roughly every 15.6 ms, so the 4 KiB case finishes inside a tick and reads 0.0, the floor supplies 0.05, and 31.2 divided by 0.05 is 624 ^[17]. The log printed x625.0 ^[15]. Both measurements were quantised to ticks: 31.2 is exactly 15.6 doubled ^[18]. The same input passed on Python 3.8 in the same run, and a genuine quadratic blowup does not care which interpreter runs it ^[19].

The second attempt is the more useful failure. It keyed the autorange target off `time.get_clock_info("process_time").resolution`, which returns 1e-07 on Windows, the unit `GetProcessTimes` reports values in and not the rate at which those values change ^[21]. The target would have been 0.002 ms, so the autorange would have stopped on the first iteration and the fix would have been a no-op on the only platform it exists for ^[22]. It would also have passed every local run, because the author's Mac really does resolve to a microsecond and the repeat count stays at 1 there ^[23]. The version that shipped spins on `process_time()` until the value changes, then takes the larger of the measured tick and the reported resolution; on Windows the target becomes about 312 ms ^[24]. That target is twenty ticks ^[4], and 156,000 times what the documented API would have produced ^[5].

The ranking check failed for a duller reason. Its assertion compared the raw length of two strings, and the scoped rule's body wins because it carries the Applies to: line, 41 characters ^[6]. The check had been passing on the length of the line that declares the scope it was supposed to be testing. Its fixture committed five times into a temp repository, a tenth of the fifty commits the ranking needs before it returns anything ^[5][3]. "Windows didn't find a Windows bug. Windows got unlucky with string lengths in the other direction and exposed a check that had never worked," the author wrote ^[7]. The fixture now derives its commit count from the constant, `range(_rollup.MIN_COMMITS_TO_RANK + 2)` ^[8]. With the trigger firing, the three context blocks measure 541, 306 and 300 characters, against 411, 370 and 371 before ^[9]: a real margin of 235 characters where the accidental one was 41 ^[1][2].

The middle case is the one I would expect to find in most suites. The check sets `sys.platform` to "win32", exercises the detector, restores the platform, and only then asserts that the detector stays quiet ^[12]. On Windows the real platform is the platform under test, the detector correctly fires, and the check goes red ^[13]. "I'd written a check whose final assertion was only correct on machines that are not the thing it tests," the author wrote ^[14].

For any of this to transfer, a suite needs one of three shapes: a fixture whose magnitude is compared against a constant that lives in the code under test ^[5], an assertion that runs outside the region it patched ^[12], or a ratio with a floored divisor and a clock coarser than its fast case ^[16][17]. The author does not report running mutation testing or a fixture audit. All three surfaced because CI ran the suite on a platform the author does not own ^[2]. I would expect the first two to show up as surviving mutants, since one never reaches the ranking code and the other asserts a negative. The timing check is a defect in the harness's own measurement, and no mutant of the redaction routine changes what a floored divisor prints ^[16].