[hatoo]

Standard user headers currently spell out their names in every reusable HTTP/2 request block. Use the RFC 7541 static-table name reference while retaining the caller's plain value, order, duplicates and literal-without-indexing semantics. Unknown names remain literal. A representative eight-header block becomes 151 bytes instead of 235, saving 84 bytes on every request without per-request encoder work.

Three alternating counted pairs against a fixed Node 20.20.2/nghttp2 1.61.0 decoding peer gain 32.91% median RPS. A separate reversed-order confirmation with twice as many requests gains 33.29%. All six pairs improve (29.61–37.86%); server CPU/request falls 24.84% in both series. Mean latency falls 24.73%/24.94% and p99 falls 26.79%/38.41%. This claim is specific to the eight-header decoding workload.

• Baseline b8766f9; Rust 1.99.1, default production features, consistent fat LTO/one codegen unit, independently frozen --bin shb builds. WSL2/Ryzen 9 3950X, client CPU 8, server CPU 0; 16 connections × 64 streams. 32,768 warmups/variant, then 131,072 requests/sample (confirmation 262,144).
• All 5,046,272 Node responses across acceptance and guard/stress workloads match the peer's decoded count exactly, with verified values, zero errors, closed sessions, and exact per-request byte savings. Fixed hfast comparisons also observe +10.19%/+6.95% for the common-header workload; it does not decode HPACK, so those are supporting wire/copy observations.
• Ordinary Node GETs have +0.06% median RPS; unknown names -2.42%, both with identical bytes and mixed pairs. Short H3 guards were unstable (-8.71%/+2.68%); a single duration correction to about eight seconds/sample gives +10.04%, all three positive. No H3 gain or consistent regression is established. All 181,665,792 comparison/guard responses including warmups have zero errors. No timestamp, sample, status, error or request-accounting code changes.
• Debug and release each pass 513 unit + 62 integration tests; strict all-target Clippy and rustfmt pass. Independent python-hpack/hyper-h2 checks 453 blocks/all 52 names and 17 live cases/544 responses per binary, including duplicate/empty/UTF-8 fields, large CONTINUATION blocks, table-size zero, POST bodies and exact transport-byte accounting. A wrong-static-index mutation is rejected.

CI now runs the independent oracle with pinned versions. Reproducible counted peer/driver, complete methodology, workload JSON, limitations and all guard results are in benchmarks/h2-static-header-names.md. Raw measurements remain in local shb/optimization-results/c178/; no large artifacts are committed. HTTP/1 and HTTP/3 encoders are unchanged; this does not resolve historical QUIC loss-recovery limitations.