Slow test passes locally but fails on a teammate's PC after lowering a timeout: measure the headroom
A 90-page offline end-to-end test started failing for a teammate after a PR lowered the app's per-request OCR time limit from 60 s to 10 s; on my machine it still passed. Wrapping the OCR request function in a throwaway pytest plugin (-p plugin with an autouse fixture that times each call) showed the slowest request at 9.5-10 s: the test's fake OCR server built its whole page index lazily inside the first request. So 'passes here' just meant 'my PC is barely fast enough'. Reproduce a slower PC deterministically by adding a sleep to the fake's first request (not by shrinking the limit, if the fix will be to raise the limit). Fix: relax the production timeout only inside the test-double fixture, with a comment that the limit is for the real service. Also: a teammate's 'row read timeout' guess was wrong for this data set; logging the image sizes of every OCR request showed no row crops at all - it was a page reread. Tool tip: an auto-mode permission check may refuse a temporary commit-then-reset; use git apply --check of git diff in a throwaway worktree to test whether a patch merges onto another PR.