fix(ipc): use named FIFOs for task communication on Unix

[wan9chi]

Motivation

Codex CLI's and Claude Code's default sandboxes block Unix domain sockets, so vp run fails while setting up task communication before any task code runs (#562). Named FIFOs are plain files, and both sandboxes allow them.

This PR reapplies the FIFO transport from #565 with a rename, a review pass, and new coverage:

• The crate is pipe_socket: socket-style server-client IPC, implemented on named pipes rather than Unix domain sockets. It carries no task-runner specifics, so it sits outside the vite_* prefix, like fspy and pty_terminal. It exposes Server::bind, Server::name, Server::accept, and Client::connect; FIFOs on Unix and named pipes on Windows stay implementation details.
• Client::connect now fails instead of hanging when the server is gone. The rendezvous open is nonblocking, so a missing reader turns into a connection-refused error at once. While waiting for the server's ready byte, the client watches the rendezvous write end and probes the server every 100ms; a dead server turns into an error within one probe interval. macOS needs the probe because its poll does not always report FIFO events; Linux reports the death at once. The module documents the handshake step by step.
• A new integration test pins the no-hang behavior for both cases: server gone before the connection attempt, and server dying in the middle of one.
• The Codex sandbox snapshot now records the end state this PR and fix(fspy): unify shared memory on a sparse temp file across all platforms #576 were built for: the task runs inside the sandbox, fspy traces its reads, and editing the traced file causes a cache miss on the next run. Regenerated against the real codex CLI on macOS. The Claude sandbox snapshot is carried over from fix(ipc): use named FIFOs on Unix #565 unchanged; it needs an srt binary to regenerate, and the recorded behavior does not depend on the shared-memory backend that changed underneath it.

Closes #562.

🤖 Generated with Claude Code

[github-actions]

fspy benchmark

linux

dynamic/launch             change  +0.18%  [ -5.43% ..  +7.19%]  overhead   +47.15%
dynamic/access             change  -1.10%  [ -6.42% ..  +4.96%]  overhead    +4.98%
dynamic/access-relative    change  +1.01%  [ -3.18% ..  +6.83%]  overhead   +51.79%
static/launch              change  +0.96%  [ -6.08% ..  +7.69%]  overhead  +157.09%
static/access              change  +0.93%  [ -4.92% ..  +6.71%]  overhead  +723.95%
static/access-relative     change  +0.29%  [ -6.97% ..  +5.41%]  overhead +1021.09%

macos

dynamic/launch             change  -0.16%  [ -5.56% ..  +4.83%]  overhead  +240.53%
dynamic/access             change  -1.30%  [-60.29% .. +128.57%]  overhead    +4.54%
dynamic/access-relative    change  +1.83%  [ -7.97% .. +113.28%]  overhead  +251.70%

windows

dynamic/launch             change  -0.42%  [ -8.75% ..  +8.42%]  overhead   +24.56%
dynamic/access             change  -0.36%  [ -4.98% ..  +1.80%]  overhead    +1.50%
dynamic/access-relative    change  -0.18%  [ -1.79% ..  +1.47%]  overhead    +1.64%

[wan9chi]

@codex review

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 6499a5c73c

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".