[jlongster]

Ported from anomalyco/opencode-pty#31. Part of the stack for anomalyco/opencode-pty#25; builds directly on dev (#59700 has merged), no prerequisite.

Issue for this PR

Fixes anomalyco/opencode-pty#25 (this PR completes it)

Type of change

• Bug fix
• New feature
• Refactor / code improvement
• Documentation

What does this PR do?

• The Unix daemon loop now sleeps in poll(2) on the listener plus a private wake channel (UnixStream::pair, non-blocking, CLOEXEC) instead of a 10 ms sleep. The timeout is the minimum of the next ownership deadline, a 2 s safety net, and the accept back-off.
• Control notifies under the ownership mutex for every stop/ownership mutation. Control::tick drains the wake channel, observes, detects missed wakes, and ticks under the same lock, so a missed wake is detected deterministically (a warning in release builds; an error in debug builds, with cleanup still running).
• Handler completion wakes the loop, so finished connections close promptly.
• EMFILE-class errors (EMFILE, ENFILE, ENOBUFS, ENOMEM) from accept or cancellation now pause accept for 100 ms instead of killing the daemon.
• Windows keeps its 10 ms poll via stubs. Follow-up: WaitForMultipleObjects on the connect event plus a stop event, and replacing the per-terminal 10 ms root-exit poll.
• Enables the poll feature of nix in packages/process-daemon/Cargo.toml (no new crates; Cargo.lock is unchanged).

Ported from the opencode-pty commit; the only difference is that connection handlers are spawned with plain thread::spawn, as on dev.

Measurement (macOS, debug builds, idle owned daemon)

	Before	After
Context switches	96.9/s	1.0/s
Ping round trip	11.6 ms	0.19 ms
Owner close → daemon exit	18.9 ms	3.4 ms

Tests added

• packages/process-daemon/src/daemon/server.rs:
  • owned_handoff_expiry_keeps_daemon
  • idle_loop_does_not_spin
  • connect_does_not_wait_for_timeout
  • handler_completion_closes_socket_promptly
  • missed_wake_is_detected
  • missed_disconnect_is_detected
  • notify_storm_is_not_missed
• packages/process-daemon/src/ownership.rs: deadline_tracks_each_state
• packages/process-daemon/src/transport/unix.rs:
  • wake_is_cloexec_and_nonblocking
  • wake_notify_never_blocks_and_drain_empties
  • listener_wait_reports_each_source
  • listener_wait_rounds_timeout_up

Mutation checks (original PR)

• Removing the handler-completion (Done) notify fails handler_completion_closes_socket_promptly.
• Removing the update_ownership notify fails owned_handoff_expiry_keeps_daemon.
• Making tick notify fails idle_loop_does_not_spin, missed_wake_is_detected, and missed_disconnect_is_detected.

Flakiness runs (original PR)

• Server module ×10, full lib ×10, concurrent ×5, single-threaded ×3. All passed.

How did you verify your code works?

• Port validation (macOS arm64, Rust 1.90.0, in packages/process-daemon): cargo fmt --check, cargo clippy --locked --all-targets --all-features -- -D warnings, and cargo test --locked --lib (65 passed) pass.
• On the stack tip (daemon-windows-poll) the full macOS CI test selection and Windows/Linux cross clippy also pass, and the server module tests (13) passed 3 runs in a row.
• Original PR: the macOS CI test selection passed: cargo test --locked --lib --test service --test runtime --test ownership --test rows --test ghostty-effects (lib 59, ghostty-effects 5, ownership 8, rows 1, runtime 8 + 1 ignored, service 9), and cross clippy with -D warnings passed for x86_64-pc-windows-msvc and x86_64-unknown-linux-gnu (--all-targets --all-features, ghostty-vt link libs stubbed). Native Windows and Linux rely on CI.

Screenshots / recordings

N/A, internal daemon change.

Checklist

• I have tested my changes locally
• I have not included unrelated changes in this PR