perf(server): skip empty middleware work

[Upd4ting]

Summary

• skip prefix/postfix route lookups and promise continuations when their registry is empty
• skip monitor lookup, response snapshot, and promise continuation when no monitor can run
• avoid sorting zero or one matching callback
• keep fast-path state synchronized across registration, removal, duplicate IDs, and hot reload lifecycle calls

This PR intentionally does not change controller resolution or HTTPResult construction/serialization.

Rebase on #29

Rebased commit: e8d87505a5912cd06dec1a9839224d71f35050ac on 0c212790f29a241aeb750026caea7c09de553f4a (perf(api): compile controller resolution plans (#29)).

The rebase was conflict-free. The incremental diff still contains only:

• src/server.ts
• src/test/server-fast-paths.test.ts

git merge-base --is-ancestor 0c212790 HEAD succeeds, so #29 is preserved in full. The benchmark below measures new main including #29 against new main + #30; it does not reattribute #29's controller-resolution gains to this PR.

Incremental benchmark after rebase

Method

• baseline: 0c212790; candidate: e8d8750; same orb
• Node.js 22.19.0; autocannon 8.0.0; loopback HTTP
• server pinned to CPU 2; load generator pinned to CPUs 4 and 6
• 3-second warm-up followed by 10 seconds measured, 5 repetitions
• before/after order alternated per repetition; fresh server process for every sample
• all fixtures register 1,000 static routes; no fixture-specific production fast path
• response: { "hello": "world" }; dynamic route: /users/:id; POST uses a small JSON body
• dispersion is population coefficient of variation; p99 is the worst observed p99 across repetitions

These direct-server scenarios isolate the middleware/monitor change. #29 is present in both artifacts but its controller plans are intentionally outside this incremental measurement.

Scenario	Main + #29 req/s	+ #30 req/s	Mean change	CV before → after	Range before → after	p99 max	Errors
GET, c=10	34,453	36,582	+6.2%	4.9% → 7.0%	31,791–36,287 → 32,655–39,780	1 ms → 1 ms	0
GET, c=100	34,945	37,223	+6.5%	3.7% → 3.2%	32,957–36,767 → 36,120–39,341	5 ms → 5 ms	0
Dynamic GET, c=10	31,182	34,563	+10.8%	5.8% → 4.9%	29,426–34,418 → 32,667–37,699	1 ms → 0 ms	0
POST, c=10	30,109	31,231	+3.7%	1.2% → 3.0%	29,606–30,693 → 30,273–32,593	0 ms → 0 ms	0
Last of 1,000, c=10	34,970	37,791	+8.1%	2.1% → 5.5%	34,441–36,449 → 35,731–41,660	0 ms → 0 ms	0

The five paired deltas are not uniformly positive: GET c=10 includes one -9.4% run, POST one -1.4% run, and route 1,000 one -0.1% run. Their paired median changes are respectively +9.6%, +2.7%, and +7.3%. The aggregate direction is positive, but GET c=10 and route 1,000 remain noisy.

Populated pipeline

The pipeline fixture uses two prefixes, two postfixes, and one monitor.

• means: 34,266 → 35,470 req/s (+3.5%)
• medians: 35,128 → 35,720 req/s (+1.7%)
• CV: 4.2% → 1.1%
• paired deltas: +10.7%, +9.1%, +0.3%, +0.6%, -2.0%
• p99: 0 ms; errors/timeouts/non-2xx: 0

Because three runs are near zero and one regresses, the populated-pipeline result remains directional/inconclusive. This PR claims the robust fast path only for empty or non-matching work.

Correctness

The integration test exercises an initially empty pipeline, dynamic prefix/postfix/monitor registration, priority order, monitor execution, removal, and the restored empty fast path. The existing contract suite covers HEAD, OPTIONS, errors, monitor isolation/errors/priorities, WebSocket success/error monitoring, and route lifecycle behavior. #29's controller-plan tests are also included after the rebase.

Commands run after rebase:

• pnpm install --frozen-lockfile
• pnpm build
• pnpm --package=@antelopejs/core@1.4.7 dlx ajs module test . — 126 passing
• pnpm lint

GitHub CI and GitGuardian pass on the rebased commit. No review or inline comment is currently open.

Limits and risk

• This is registry-level invalidation rather than a per-route cache, so registration/removal/hot reload remain immediate and unbounded URL caches are avoided.
• Requests with matching prefix/postfix/monitor callbacks retain the existing lookup and execution path.
• The benchmark is intentionally incremental and server-level. Full-controller integration numbers for perf(api): compile request controller resolution #29 + perf(server): skip empty middleware work #30 remain documented in the existing PR comment, but are not used to estimate perf(server): skip empty middleware work #30 alone.
• The populated-pipeline result is not strong enough to claim a gain.