perf(tests): stabilize subscribe_user_trio so the gate stops flipping

[joeblau]

subscription/subscribe_user_trio has been producing coin-flip regression verdicts on unrelated PRs. On #84 — which changes only two files under src/signing — it reported +39.6% with a band of [-29.7%, +177.2%], a band spanning zero because the three rounds disagreed in sign.

Why it was unstable

It built one transport per sample and awaited three subscriptions, so a sample measured async scheduling and first-transport warmup rather than the cost of subscribing. Every neighbouring scenario was stable across the same measurement positions, which is what rules out runner noise:

scenario	R1	R2	R3
l2book_dispatch_50_coins	2.49 → 2.65	2.51 → 2.48	2.53 → 2.55
subscribe_200_coins	14.15 → 13.74	14.12 → 15.07	14.47 → 14.48
http_request	2.97 → 2.74	3.00 → …	2.86 → 3.18
subscribe_user_trio	57.15 → 39.01	57.75 → 40.58	57.91 → 41.49

What changed

20 iterations per sample instead of 1. Per-sample warmup and scheduling jitter now amortize across 60 subscriptions rather than landing on 3. unitsPerIteration still normalizes to a per-subscription figure, so the reported unit is unchanged — it now reports the steady-state cost instead of a number dominated by the first transport in each sample.

MockWebSocket keeps only the most recent instance instead of pushing every one into a static array. lastMockWebSocket only ever read .at(-1), but the array kept every socket — and through the listeners on it, every transport and its keep-alive timers — reachable until the scenario ended. That mattered little at one transport per sample and matters a great deal at twenty.

Result

Three full-suite runs on one machine:

	run 1	run 2	run 3
before	14.68 µs (rme 112.0%)	11.00 µs (9.6%)	13.22 µs (12.7%)
after	9.90 µs (rme 6.6%)	9.61 µs (5.6%)	9.57 µs (6.8%)

Run-to-run spread falls from 33% to 3.4% and worst-case rme from 112% to 6.8%. Since baseline rme feeds the comparator's tolerance, the tighter figure also stops this scenario inflating the gate's noise allowance for every other scenario in the report.

I tried 50 iterations too: it tightens the median further but pushes rme back to 16–26% from GC pressure at 50 transports per sample, so 20 is the better trade.

Expected: the perf check will be red

This edits tests/perf, so the suite fingerprint changes and the gate fails closed by design — "Land an intentional suite change with the gate red, or as its own PR." This is that own PR. The absolute figure moves too (~13 µs → ~9.6 µs locally), which is expected: it no longer bills first-transport warmup to every sample.

bun run check exits 0; 1652 pass, 236 skip, 0 fail.

🤖 Generated with Claude Code

[joeblau]

Superseded by #86, which absorbed this branch's work: subscribe_user_trio now runs 20 iterations per sample and MockWebSocket keeps only the latest instance. Verified byte-identical — every file this branch touched matches the merged tree exactly. Closing rather than merging to avoid landing the same change twice.