v4.2.1

Changed

• HTTP/2 warm streaming path now borrows the cookie helpers, dropping a per-request allocation on the cookie-bearing path (commit 26d5a78). A/B measurement on a quiet host showed this neutral on TTFB and throughput; it is allocation hygiene with no measured latency or throughput effect.
• Renamed the streaming benchmark metric from ttft to ttfb (time-to-first-byte) across the benchmark harness, threshold test, JSON artifact schema keys, and binding keywords. The transport benchmark measures time-to-first-byte; the LLM-facing binding descriptions keep "TTFT" (time-to-first-token) because lower transport time-to-first-byte is what produces the faster first token when proxying a model's response stream.
• Re-baselined all published streaming, native-HTTP/3, and WebSocket benchmark numbers (README and docs/benchmarks/) to a single quiet AWS Graviton4 host so every competitor comparison is measured on the same machine, replacing the prior mix of Mac-sourced runs. Library request/response behavior is unchanged from 4.2.0; the figures move with the environment. Streaming clears every gate by wide margins with paired Wilcoxon p underflowing to zero at n=100 (H2 request-body TTFB and throughput rose versus the prior Mac figures; H2 response-body TTFB and H1 response-body throughput fell). The native HTTP/3 superiority gate passes against quiche, tokio-quiche, h3-quinn, and reqwest_h3. WebSocket loopback message-rate is parity with fastwebsockets and tokio-tungstenite inside run-to-run variance.

Removed

• Withdrew the RFC 9220 WebSocket-over-H3 tunnel superiority claim from the README. On the Graviton4 host the full-suite gate does not pass: Specter leads p50 TTFB and throughput on all three tunnel workloads and wins the slow-consumer mixed workload outright, while tokio-quiche holds a lower p95 tail on the echo and client-DATA+FIN workloads. The measured result is recorded in docs/benchmarks/native-h3-vs-rust-clients/README.md.