[LeandrodaSilva]

Summary

Benchmark-driven zero-allocation work on the TenCore.fetch hot path. Profiled the in-process request path with a new micro-benchmark, identified the three biggest per-request allocations, and eliminated them — no public API or behavior change.

Changes

• Memoize _availableLocales — was rebuilding a Set + Object.values() + sorted array on every request. Now cached in _availableLocalesCache, invalidated wherever _i18n is reassigned (setter, updateManifest, embedded init).
• Skip middleware bind+closure in _handleRequest when no middleware is registered (the common case).
• Skip the response-hook pass when no response hooks are registered (avoids an async hop).

Benchmarks (Deno 2.8.2, in-process fetch)

benchmark	baseline	optimized	Δ (p75)
core_fetch_api	~13.3 µs	~8–10 µs	13.2 µs → ~8 µs
core_fetch_i18n	~15.4 µs	~11–13 µs	15.6 µs → ~12 µs

Roughly 25–37% faster on the plain API path and 15–30% on the i18n path (p75 is the steadier signal under run-to-run noise).

Verification

• New test in tenCore.pipeline.test.ts proves the locale cache invalidates when i18n is reassigned (404 before config → 200 with Content-Language after).
• Added bench thresholds for core_fetch_api / core_fetch_i18n.
• deno fmt / deno lint / deno check clean; full suite green; coverage 90.6% (above the 90% floor).

https://claude.ai/code/session_015AsggMbqEhejVYtbq1LubN

---

Generated by Claude Code