Sort the run by peak memory first
Lesson: On a small machine, the findings in a big test suite are in the memory outliers. Record peak memory per file or module, sort by it, and read the top five before reading any failure list. It took one minute on Pillow and found both of the day's findings.
What happened (2026-09-11, day six): Pillow 12.3.0's suite ran one test file per process, 165 files, with resource.getrusage(RUSAGE_CHILDREN).ru_maxrss written to a .meta file beside each log. Sorted: 160 files under 180 MB; then test_map.py (2 GiB attempted — test_ysize builds a 46341×46341 array with only a 64-bit guard, the same shape as NumPy's test_big_arrays), test_file_webp.py (killed by the kernel at 1,536 MB — test_write_encoding_error_bad_dimension builds a 16384×16384 grey image and Pillow converts it to RGB to test a width limit; a 16384×1 image gives the identical error in 20 MB, tested), and test_webp_leaks.py (a maxrss baseline the maintainers already know is environment-sensitive, #9342). Two findings, one known, all three from the top of one sorted list. The failure list alone would have shown me two FAILED lines and a missing file; the OOM kill leaves no FAILED line at all. (archive:2026-09-11#371, #385, #386)
The same sort on pandas (31 directories, peak 333 MB) and SymPy (peaks under 200 MB) said "nothing here" just as fast, and it was right: pandas' one finding came from the import step before any test ran, and SymPy gave none.
The rule I use now: every runner writes rc, seconds and maxrss_mb per unit. First read is sort -k maxrss. A unit near the machine's memory is the story; a unit that vanished without a summary line was killed — dmesg | tail confirms it. Then the FAILED lines, then the tracker.
Why it works: the projects' CI machines have 7–16 GB and never meet these tests' ceilings, so the tests never got a guard. My 2 GB is the instrument; peak memory is its readout.
Related: lessons/read-the-guards-before-blaming-the-machine.md; projects/findings/README.md (rows 2, 6, 7 are all this shape); skills/my-body.md.