[aarambh-darshan]

Summary

Round 3 of the optimization campaign (tracked in the local audit backlog). This PR targets everything outside arvik-middleware: the dispatch path, static-file serving, extractors, and metrics. Follows #28 (round 2) and #30 (middleware).

Router — dispatch unboxing (C2/C3)

RouterService::call used to clone an Arc, box a large async state machine, and re-box again at the erasure boundary. Dispatch is now split into:

• route() — synchronous: radix-trie match, param interning, extension inserts, returns the leaf future + pattern
• one single Box::pin at the leaf, with layered and non-layered paths sharing it

Removes one pure-waste heap alloc per request plus two atomic RMWs.

Static files (arvik-static)

Finding	Change
C7a/b	.br/.gz probes are now direct open attempts (a miss costs one failed open instead of stat+open pairs) in both ServeDir and ServeFile; plain files validate via the already-open descriptor; canonical root cached in OnceCell
C8	Embedded serving checks O(1) A::get before the O(all-assets) directory-prefix scan
C9	relative_path = one sized String with inline traversal validation (was PathBuf + Vec + N component Strings + join); duplicate field dropped
C10	Hot MIME types return HeaderValue::from_static; table verified byte-for-byte against mime_guess 2.0.5 by unit test
C12	Fallback services behind Arc<BoxCloneService> → service clone is a refcount bump
C14	Embedded ETag/mtime cached per asset path (was hex-formatted + validated every request)

New regression tests: traversal/normalization semantics for relative_path, content-type table parity vs mime_guess.

Metrics (arvik-observe) — C11

• status label borrowed from StatusCode::as_str() (canonical zero-padded form) — kills the per-request format!
• route label refcounts the interned Arc<str> instead of to_string()
• exact-size response bodies record straight from the size hint — no CountingBody, no atomic, no callback allocation for any Content-Length response; only genuinely streamed bodies get wrapped

Core / extract

• C5: constant Content-Types use HeaderValue::from_static across into_response.rs, ResponseBuilder helpers, and JSON responses (~9 sites) — was malloc+memcpy+validation per response
• C15: JSON serialization hint 128→512 B; application/json;charset=utf-8 fast path before Mime parse; BufWriter between upload chunks and temp files; targeted multipart boundary extraction (replaces multer's full Mime parse; quoted/case/first-wins semantics covered by tests); cookie-jar memoization shared across all three jar extractors
• C1 documented in-code: the erased-handler re-box is inherent while H::Future is a GAT (elimination tracked for the future trait migration)

Verification

• ✅ cargo fmt --all -- --check
• ✅ RUSTFLAGS="-D warnings" cargo clippy --workspace --all-features --all-targets
• ✅ cargo test --workspace --all-features
• Benchmarks will be re-run after merge to quantify plain-route / static-route deltas against BENCHMARK_REPORT.md baselines.

Deferred (intentional)

• C4 (Arc'd baked stacks), C6 (single extensions map), O1 (handler trait migration) — architectural, next round.