Found reviewing #378.
load_corpus()'s docstring says:
477 records read three times, 7.9s to 12.4s as checks were added (#373)
#378's own description explains at length why those numbers cannot be quoted: the machine is too noisy, interleaved A/B trials put main anywhere from 15.5s to 29.9s running identical code, and the first measurement of the fix came out slower than the baseline it beats. That is why the PR moved to a deterministic metric — yaml.safe_load calls, 1431 → 477.
So the PR body and the code comment now disagree, and the code comment is the one that persists after the PR is forgotten. Someone re-tuning this later reads 7.9s→12.4s as a measured baseline and cannot reproduce it.
Fix: quote the parse count in the docstring, which is the reproducible figure and the one the change actually guarantees.
Found reviewing #378.
load_corpus()'s docstring says:#378's own description explains at length why those numbers cannot be quoted: the machine is too noisy, interleaved A/B trials put
mainanywhere from 15.5s to 29.9s running identical code, and the first measurement of the fix came out slower than the baseline it beats. That is why the PR moved to a deterministic metric —yaml.safe_loadcalls, 1431 → 477.So the PR body and the code comment now disagree, and the code comment is the one that persists after the PR is forgotten. Someone re-tuning this later reads 7.9s→12.4s as a measured baseline and cannot reproduce it.
Fix: quote the parse count in the docstring, which is the reproducible figure and the one the change actually guarantees.