Add an opt-in WebSocket transport for Codex Responses

[mjlbach]

Summary

• add src/gateway/websocket.zig, an RFC 6455 client codec (handshake keys, masked frames, fragmentation, control-frame rules) over std.Io reader/writer streams
• add src/gateway/openai_codex_websocket.zig: upgrade the Codex Responses endpoint to a WebSocket (OpenAI-Beta: responses_websockets=2026-02-06), send the request body as one response.create text frame, and feed each received frame into the same Responses reducer the SSE transport uses, with identical stream limits
• cache the connection per (session, account) with a five-minute idle and 55-minute age limit, so later turns skip DNS, TCP, and TLS setup; a busy or key-mismatched entry gets a one-off connection, and a reused connection that fails before any model output is replaced by a fresh one once
• opt in with FX_OPENAI_CODEX_TRANSPORT=websocket; the default transport is unchanged
• a WebSocket failure before any model output falls back to SSE under the original invocation admission and latches the process to SSE so a broken proxy costs one failed handshake, not one per turn; after output has been emitted the failure is surfaced instead — a replay can never duplicate tool calls
• request.deadline is respected end to end: it bounds the handshake and the stream phase (via the bounded cancel watcher), surfacing error.Timeout
• masked server frames are rejected (RFC 6455 5.1: only clients mask)

The handshake is std.http.Client's existing 101/informational path, so TLS, proxies, and the cancel watcher are inherited rather than reimplemented. Protocol and policy mirror Codex (codex-rs/core/src/client.rs) and pi (openai-codex-responses.ts); pi's transport selection, fallback-before-output rule, and cache limits are followed exactly. Prewarm, previous_response_id continuation, and zstd compression on the SSE path are follow-ups.

Testing

• zig fmt --check on touched files
• zig build test — full suite passes; new tests cover RFC 6455 vectors and frame round trips, a loopback WebSocket server proving end-to-end streaming plus connection reuse (one TCP accept serving two requests), fallback-safe handshake rejection, and the fallback error classification
• bun test acp.test.ts (116 tests) and tui-auth-source-selection.test.ts (56 tests) pass, including two new e2e tests that drive the built binary against a Bun WebSocket loopback server: a completion streamed over one upgraded connection with zero SSE requests, and pre-output fallback to SSE when the upgrade is rejected
• live against chatgpt.com: single-shot and tool-using turns stream over the WebSocket; a saved session logs codex_ws_connection reused=true on the second step, removing the ~400 ms connect observed on fresh connections

🤖 Generated with Claude Code

[mjlbach]

This PR has been folded into #521