perf(apiStabilityLeak): reuse one analysis per file

[mattiamanzati]

Why

apiStabilityLeak built a fresh ApiStabilitySession for every export of a file, and each one allocated and regrew its maps and work queues. On packages/effect the rule allocated about 250 MB this way. The rule now keeps one session per file and resets it before each export.

What changed

• ApiStabilitySession.Reset calls a new apiStabilityAnalysis.reset. That clears every per-export map, truncates the settle and publish slices, resets the settle scope and zeroes the budgets and the materialization epoch. The allocated capacity is kept.
• A shared-cache hit in typeSurface and signatureSurface goes into the session map and the settle order, but not into the publication queue. Publication skipped it anyway.
• publishTypeSurface and publishSignatureSurface allocate a snapshot's findings map only for the first retained finding.

Scope

The settle scoping and the lazy surface findings maps from #862 and #863 are unchanged. A future per-export field on apiStabilityAnalysis must also be cleared in reset. The new test compares a reset analysis field by field with a fresh one, and it fails on a field kind it does not know.

Blast radius

Only apiStabilityLeak uses the session, and the rule is off by default. A missed field in reset would leak one export's results into the next export. The field-by-field test and the byte-identical diagnostic dump guard against that.

Verification

• Effect 4.0.5 (7f524e25) with every rule executing, packages/effect: checkApiStabilityLeaks allocation 252 MB → 140 MB (median of 5 interleaved profiled runs per side, ranges disjoint). Rule CPU moved 0.98 s → 0.92 s, inside the noise.
• A dump of every rule-returned diagnostic from a clean build of the whole workload is byte-identical on both sides, with 1,404 apiStabilityLeak diagnostics each.
• New tests in api_stability_reset_test.go check three things: a reset analysis equals a fresh one, a stale budget or result does not reach the next export, and a cache hit is not queued for publication again. Mutants that drop a clear or the work reset, or that re-queue the hit, all fail them. go test ./internal/..., pnpm lint and pnpm check pass. pnpm test did not run.
• Unprofiled clean builds of the whole workload, 8 interleaved pairs: median aggregate check time is 63.4 s before and 63.2 s after, with patched faster in 4 of 8 pairs. Wall time shows no measurable change on a shared, loaded machine.

🤖 Generated with Claude Code

[github-actions]

Effect perf vs stock tsgo: 2 flagged rows

• latest · Memory allocs over stock · 🟩 -1.2 pts
• latest · Check time over stock · 🟩 -3.5 pts

Details

Metric	base	pr	Delta	Best	Worst	p-value
latest · 7.0.2 · effect (all rules)
Errors	1	1	~	1	1	counter, n=5
Warnings	0	0	~	0	0	counter, n=5
Symbols over stock	+712,107 (±4)	+712,113 (±4)	~	+712,106	+712,113	counter, n=5
Types over stock	+340,880 (±5)	+340,884 (±2)	~	+340,882	+340,886	counter, n=5
Instantiations over stock	+1,208,629 (±11)	+1,208,645 (±11)	~	+1,208,626	+1,208,648	counter, n=5
Memory allocs over stock	+19.3% (19.2 to 19.6)	+18.1% (17.8 to 18.2)	🟩 -1.2 pts	+17.8%	+18.2%	p=0.008 n=5
Check time over stock	+41.2% (39.8 to 42.6)	+37.7% (33.1 to 39.6)	🟩 -3.5 pts	+33.1%	+39.6%	p=0.008 n=5
Total time over stock	+27.2% (25.6 to 29.3)	+22.8% (21.0 to 27.7)	~	+21.0%	+27.7%	p=0.095 n=5
next · 7.1.0-dev.20261015.1 · effect (all rules)
Errors	18	18	~	18	18	counter, n=5
Warnings	0	0	~	0	0	counter, n=5
Symbols over stock	+564,074	+564,074 (±3)	~	+564,068	+564,074	counter, n=5
Types over stock	+299,171 (±1)	+299,169 (±3)	~	+299,165	+299,171	counter, n=5
Instantiations over stock	+1,001,074 (±2)	+1,001,071 (±10)	~	+1,001,055	+1,001,074	counter, n=5
Memory allocs over stock	+16.3% (15.8 to 17.9)	+15.8% (15.0 to 16.5)	~	+15.0%	+16.5%	p=0.222 n=5
Check time over stock	+36.1% (33.4 to 39.7)	+37.7% (27.6 to 43.0)	~	+27.6%	+43.0%	p=0.690 n=5
Total time over stock	+26.0% (22.3 to 27.4)	+27.0% (22.8 to 29.3)	~	+22.8%	+29.3%	p=0.421 n=5

• Stock counters, latest: 7.0.2 symbols 18,255,289 (±1) · types 6,184,373 · instantiations 26,135,429
• Stock counters, next: 7.1.0-dev.20261015.1 symbols 17,710,777 · types 5,960,725 · instantiations 25,181,429
• Workload: https://github.com/Effect-TS/effect@7f524e25cfb49a71ac39b0f44278d08d40dda5de with every Effect rule executing (skipDisabledOptimization: true)
• Pairs per side: 5, alternating stock-first and patched-first
• Commits: base d3b53a718365b55cc11799d08e67e54869714890 · head 73dfcbfe632fa7af152e4b8151e10937100a0319
• Counters are patched − stock per pair (median ±half-range); allocs and times are patched / stock − 1 per pair, median (min to max). Time and alloc deltas are flagged only when the exact two-sided Mann-Whitney U p < 0.01.