test(act-pg): stop the scenario benches from measuring each other

[Rotorsoft]

The notify-vs-polling bench failed CI on an unrelated PR (#1612) with expected 10.79 to be less than 0. Nothing was wrong with that PR — the bench reported a number it had no business reporting, and this fixes both halves of that.

What happened

The bench compares commit→reaction latency two ways: a reader woken by Postgres LISTEN/NOTIFY, and a reader driven by a plain 50ms correlate → drain pump. The failing run looked like this:

notify  (n=30)  10.8 ms
polling (n=0)    0.0 ms

The polling arm recorded none of its 30 reactions. 0.0 ms is what the percentile helper returns for an empty sample set, so the comparison read the arm that never ran as infinitely fast and failed on the arm that worked.

The scenario benches run as concurrent vitest workers. In that run, a bench doing a 100,000-event per-event drain held a core for 109 seconds while a Postgres latency bench hammered the same database. A 50ms setInterval pump gets no turns under that, and every exception it might have raised was swallowed by a bare catch {}.

The part that matters more than the flake

The same contention corrupts what these benches report when they pass. That CI run measured an idle p50 of 12.5 ms and p95 of 516 ms, from only 38 samples. Run alone on the same code: 5.6–7.2 ms p50, 7.6–14.9 ms p95, from 144–221 samples. Roughly a 70x distortion in the tail, in numbers that feed PERFORMANCE.md.

So the fix is to run one bench file at a time (fileParallelism: false). Measured cost: 129s → 168s wall, about 30%. The sum of test time is unchanged at ~166s — parallelism was only overlapping it, and overlapping is exactly what invalidates a wall-clock latency measurement. Slow numbers beat contaminated ones when the output is a published benchmark.

And make the silent failure loud

Serializing removes the cause that has been observed; it does not make an empty sample set legible if some other cause produces one. Each arm now fails at the point where the cause can still be named, and the pump counts what it swallows instead of hiding it. Both messages were verified by simulation:

The poll arm recorded 0 of 30 reactions. Nothing failed, so the reader
never got to run — most likely starved of CPU by a bench running alongside it.

The poll arm recorded 0 of 30 reactions. The poll pump failed 28 times;
last error: Error: boom

The pump stays best-effort — the pool genuinely can close mid-flight at teardown — it just no longer loses the evidence.

Cost

Bench job wall time grows about 30%. No framework code changes; nothing here ships in a package.

Test plan

• pnpm bench:scenarios — 10 passed, serialized, 168s
• Notify bench run 5x standalone, polling arm collected 30/30 every time
• Both new failure messages verified by simulating a starved pump and a throwing pump
• pnpm test — 238 files, 3631 passed; coverage 100% statements / 100% branches / 100% functions / 100% lines
• pnpm typecheck
• pnpm lint
• CI green
• Review

Stability charter impact

None. A bench file and the bench-only vitest config; no package source, no public surface.

rfc-gate: exempt — no libs/*/src file is touched, so the stability snapshot is unchanged.

🤖 Generated with Claude Code

https://claude.ai/code/session_015xAdM431gFeFBZTW5kRrjg