feat: add replay/parser performance guardrails

[Korupt-virus]

Summary

• add reproducible Mitata guardrails for 20,000-frame raw lap parsing and semantic resolver/canonical envelope transformation
• preload and decompress deterministic capture input before CPU timing; exclude SQLite and raw file I/O from measured CPU paths
• add separate isolated replay I/O benchmark for SQLite, file read, gzip, parser/cache, and canonical replay behavior
• show existing parser/pipeline microbenchmarks as judgment-free informational measurements while retaining median, p99, and allocation deltas for enforced replay/* benchmarks
• keep initial-landing bootstrap behavior when base lacks either new benchmark

Benchmark boundaries

CPU guardrails

bun run bench runs existing parser/pipeline suite plus:

• replay/parse 20,000 raw lap frames
• replay/resolve 20,000 canonical envelopes

Capture file reading and gzip decompression finish before Mitata starts. Both cases use preloaded deterministic inputs, retain 20,000-frame/output assertions, and use Mitata warmup and repeated sampling.

Legacy nanosecond/microsecond parser and pipeline rows remain report-only because shared-runner jitter produces large percentage swings from tiny absolute changes. CI enforcement filters to replay/*, gates median/allocation at 10% and p99 at 25%, and treats missing base replay rows as initial-landing bootstrap. Guardrails begin enforcing after this benchmark exists on base.

Replay I/O

bun run bench:replay-io runs in temporary isolated SQLite state and measures:

• SQLite + capture identity + cached packets
• SQLite + file + gzip + parse + canonical replay

I/O results remain report-only with 25% tolerance because filesystem and cache behavior varies more than CPU-only work. Workflow uploads both base/current reports and starts comparison after benchmark lands on base.

Review changes

• removed custom worker, result schema, comparator, local worktree runner, hard RSS budgets, and platform-sensitive maxRSS conversion
• moved parser/replay CPU work into shared Mitata harness
• changed existing comparator to median, p99, and allocation statistics with separate steady-state and tail tolerances
• added optional benchmark-prefix filtering so enforcement targets only intended guardrails
• generate and publish comparison reports before enforcement so real regressions retain diagnostics
• replace each previous benchmark bot comment with a fresh marker-tagged report so current results appear at the bottom
• split the PR report into informational microbenchmark and replay CPU guardrail tables
• added comparator coverage for tolerance enforcement, prefix filtering, and initial landing
• extracted production pure boundaries for preloaded buffer parsing and canonical replay transformation

Verification

• bun test test/tooling/bench-compare.test.ts test/tooling/changelog.test.ts --timeout 60000 — passed
• bun test test/games/shared/semantic-replay-native.test.ts test/games/ac-evo/ac-evo-mid-session.test.ts test/session-capture/parse-bin-vs-gz.test.ts --timeout 120000 — 11 passed, 141 assertions
• bun run bench
  • parser median 64.02 ms, p99 67.89 ms
  • canonical replay median 50.71 ms, p99 54.54 ms
• bun run bench:replay-io
  • cached replay median 274.73 ms, p99 309.19 ms
  • full I/O replay median 557.90 ms, p99 610.79 ms
• replayed latest failed CI artifacts with --include=replay/ --threshold=10 --p99-threshold=25 --fail-on-regression — passed bootstrap
• bun run typecheck — passed
• bun run lint — 0 warnings, 0 errors

[Snazzie]

Request changes: This benchmark should not include SQLite or raw file I/O in its measured path; those make results non-reproducible. Please move the parser and semantic replay work into the existing Mitata harness using preloaded deterministic inputs:

• raw lap-frame parsing
• semantic resolver/canonical envelope transformation

Keep frame-count assertions, but replace custom report/comparison logic with Mitata JSON output and the existing benchmark comparator. Also fix or remove custom RSS measurements: current macOS output reports hundreds of GiB because maxRSS is multiplied by 1024 despite already being bytes.

[Snazzie]

Additional issue: the custom worker performs one timed run per scenario with no warmup or repeated samples. Mitata should own warmup/sampling so comparison uses stable statistics (median/p99 and variance/tolerance), rather than a single cold-process measurement.

[Korupt-virus]

@Snazzie Addressed both benchmark review points in 5cfcaf0b.

• Raw lap parsing and semantic resolver/canonical envelope transformation now run inside the existing Mitata suite with preloaded deterministic inputs.
• File read, gzip, and SQLite work finish outside those CPU measurements.
• 20,000-frame packet/envelope assertions remain.
• Custom worker, report schema, comparator, hard RSS budgets, and platform-sensitive maxRSS conversion are removed.
• Mitata now owns warmup and repeated sampling; comparison uses median, p99, allocation, and explicit tolerance instead of one cold-process duration.
• End-to-end SQLite/file behavior remains available only as a separate isolated, report-only replay I/O benchmark with wider tolerance, so storage noise cannot fail the CPU guardrail.

Local verification passed: CPU benchmark, replay I/O benchmark, comparator tests, 11 parser/replay contract tests, typecheck, lint, and changelog contract.

[github-actions]

Informational microbenchmarks

Runtime: bun on AMD EPYC

Report-only. Small timings can vary between runs.

Bench	Baseline median / p99	Current median / p99	Baseline alloc / current alloc
ac-evo/parse	2.63 µs / 8.35 µs	2.09 µs / 6.30 µs	437 b / 0 b
ac-evo/pipeline	28.24 µs / 30.73 µs	23.34 µs / 28.91 µs	30.29 kb / 17.92 kb
acc/parse	3.16 µs / 5.20 µs	3.44 µs / 4.78 µs	437 b / 874 b
acc/pipeline	40.80 µs / 64.70 µs	23.79 µs / 28.62 µs	0 b / 16.21 kb
f1/parse	246 ns / 1.67 µs	229 ns / 893 ns	0 b / 0 b
f1/pipeline	84.32 µs / 178.18 µs	77.98 µs / 154.33 µs	68.27 kb / 11.64 kb
fm/parse	1.10 µs / 2.54 µs	1.05 µs / 3.29 µs	0 b / 0 b
fm/pipeline	56.49 µs / 81.09 µs	23.81 µs / 37.17 µs	17.07 kb / 7.68 kb

Replay CPU guardrails

Runtime: bun on AMD EPYC
Thresholds: median/allocation ±10%; p99 ±25%
Included benchmarks: replay/*

Bench	Baseline median / p99	Current median / p99	Δ median	Δ p99	Δ alloc
replay/parse 20,000 raw lap frames	—	71.30 ms / 74.64 ms	Baseline pending	—	—
replay/resolve 20,000 canonical envelopes	—	48.61 ms / 50.53 ms	Baseline pending	—	—

Baseline pending. Regression assessment starts after matching results exist on the base branch.