fix: detect Codex boot marker format in PTY startup gate

[khaliqgant]

Summary

• Adds find_relaycast_boot_marker() to src/pty_worker.rs that matches both Claude and Codex MCP boot message formats
• Claude: "booting mcp server: relaycast"
• Codex: "Starting MCP servers (0/2): relay, relaycast"
• Codex agents now pass the startup gate immediately instead of waiting for the 25s timeout fallback

Context

This is the pre-existing boot marker mismatch identified during the Codex spawn regression investigation. While PR #597 fixes the registration timeout (the acute regression from #591), this PR fixes the underlying reason Codex agents always hit the 25s startup timeout.

Test plan

• Spawn a Codex agent — should pass startup gate in ~2-5s (not 25s)
• Spawn a Claude agent — should still work (original marker still matched first)
• cargo test passes

🤖 Generated with Claude Code

---

[xkonjin]

Code Review

Summary: Adds detection for the Codex-style MCP boot marker format ("Starting MCP servers (0/2): relay, relaycast") alongside the existing Claude-style marker. This unblocks the PTY startup gate for Codex sessions.

Issues Found

⚠️ Hardcoded search window of 200 chars

let search_end = floor_char_boundary(lower_output, (idx + 200).min(lower_output.len()));

The 200-char window assumes "relaycast" appears close to "starting mcp server" in the output. This is reasonable for the current Codex format but fragile if future output includes more MCP servers in the list (e.g., 10+ servers with long names). Consider either:

• Making the window configurable, or
• Searching to end-of-line (\n) instead of a fixed byte count

✅ Strengths

• Clean extraction into a dedicated find_relaycast_boot_marker() function with clear doc comment.
• Returns byte offset of the end of the marker, maintaining the existing contract.
• Falls through to existing RELAYCAST_BOOT_MARKER check first (Claude path), so no regression for working setups.
• floor_char_boundary usage is correct for safe UTF-8 slicing.

Missing: No test coverage
This function is pure and testable. A unit test with both Claude-style and Codex-style input strings would catch regressions if the marker formats change again. Something like:

#[test]
fn test_find_relaycast_boot_marker_codex() {
    let input = "starting mcp servers (0/2): relay, relaycast\n";
    assert!(find_relaycast_boot_marker(input).is_some());
}

Verdict

Correct fix for a real compatibility issue. Ship-ready, but adding a unit test for the new function would be valuable.