Hook bench split — Tinybench p99 explorer

Dial the workload + load conditions to see why a single absolute wall-clock assertion fails. The same p99 that passes locally fails under coverage + parallel workers — even though throughput is unchanged. Branch: test/perf-bench-split

1 · Workload

Payload size (KB)

50 KB

Has secret embedded?

2 · Runtime environment

Coverage instrumentation

Worker contention

Thermal jitter

TODAY — absolute test assertion

Threshold

scenario	p99 (ms)	verdict

Failure here means CI blocks the merge — even if the function got no slower.

AFTER — bench + ratio signal

scenario	p99 (ms)	ratio vs baseline

Ratios cancel hardware/load variance. Catastrophic backtracking shows up as ≥10x.

Why p99, not max — quick demo

Inject backtracking? (changes regex behaviour)

signal	healthy	backtrack	caught?

max moves with every GC pause too — high false-positive rate. p99 stays put until many samples slow down — strong signal, low false positives.

Numbers are synthetic but follow the actual baselines measured in src/hooks/src/__bench__/secret-handler.bench.ts. Tinybench reporter columns: hz · min · max · mean · p75 · p99 · p995 · p999 · rme · samples.