Rotorsoft / act-root / 32610427202

100%

push

github

test(act-sqlite): re-record the drain perf baseline and document the regression (#1547)

The drain: correlate+drain 50 events baseline moves 3.684ms -> 5.645ms.
This absorbs a REAL slowdown, not measurement drift: the identical bench
script on one machine measures 1.13ms at the commit that seeded the baseline
and 1.78ms on master, a 1.57x local ratio matching CI's 1.53x.

There is no single commit to revert. About a fifth predates the
subscription-work-mark series and the rest arrived during it; each change
stayed under the 1.5x allowance and passed, and two months of that crossed
the line.

Absorbing it is defensible because this scenario drains 50 events over a
handful of streams — the shape where the work-mark rework costs and never
pays, while the same rework made claim 17-18x faster from 1k to 50k
subscribed streams. The gate currently measures only the half that got worse.

PERFORMANCE.md carries the measurements, the bisect, and three follow-ups:
a many-streams drain scenario, per-release baseline re-recording, and the
small-scale cost itself.


Claude-Session: https://claude.ai/code/session_01K5eB6CMELZJfayCYpQta8N

Co-authored-by: rotorsoft <rotorsoft@outlook.com>
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>

4237 of 4237 branches covered (100.0%)

Branch coverage included in aggregate %.

9191 of 9191 relevant lines covered (100.0%)

1330.38 hits per line