websocket optimization and benchmarking

[stephenberry]

WebSocket Optimization and Benchmarking

Performance Optimizations

Shared receive buffers — All WebSocket connections on a given thread now share a single 512KB receive buffer (one allocation per thread instead of per-connection). Unconsumed partial-frame bytes spill to a small per-connection buffer. Deferred reclamation avoids thrashing. Enabled by default; disable with ws_recv_buffer_size(0).

Fused unmask + ASCII detection — XOR unmasking now processes 8 bytes at a time and simultaneously checks whether all bytes are ASCII. For ASCII-only text frames, the separate UTF-8 validation pass is skipped entirely.

Zero-allocation write fast path — When no write is in flight, outgoing frames are built directly in a persistent per-connection buffer (capacity reused across messages). Frames are only heap-allocated and queued when a concurrent write is already in progress.

Write queue simplification — Replaced std::deque<std::unique_ptr<std::vector<uint8_t>>> with std::deque<std::vector<uint8_t>>, removing a level of indirection.

Benchmark Suite

Added benchmarks/ws_benchmark/ comparing Glaze against uWebSockets using Boost.Beast as a neutral client. Tests cover:

• Single-connection echo at 64B, 1KB, 64KB, and JSON payloads
• Connection upgrade (new WebSocket per message)
• Concurrent echo with N clients (single-threaded and multi-threaded server)