Intermittent SIGABRT in terminal.search.Thread.test_0 on macOS via libxev kqueue race

[KooshaPari]

Summary

Test terminal.search.Thread.test_0 intermittently SIGABRTs on macOS / Zig 0.16.0_1 / xev kqueue backend. When the search thread is stopped via xev.Async.notify() followed by std.Thread.join(), a race between the search thread's loop.run(.once) and the test thread's loop.deinit() causes kevent() to be called on an already-closed kqueue FD. libxev's kqueue.zig:1916 then fires .BADF => unreachable (the comment "Always a race condition" at that line is correct).

Why I'm filing here

I don't have CreateIssue permission on ghostty-org/ghostty (no org membership), so this lands in Issue Triage as a starting point. Happy to convert to a proper issue or PR if a maintainer wants to take it.

Crash signature

thread N panic: reached unreachable code
zig-pkg/libxev-0.0.0-XXX/src/backend/kqueue.zig:1916:22  kevent_syscall
  .BADF => unreachable, // Always a race condition.
                    ^
zig-pkg/libxev-0.0.0-XXX/src/backend/kqueue.zig:493:48  tick
zig-pkg/libxev-0.0.0-XXX/src/backend/kqueue.zig:305:35  run
src/terminal/search/Thread.zig:197:30  threadMain_
src/terminal/search/Thread.zig:136:21  threadMain

Precursor warnings (preceding the crash)

[page] (warn): page integrity violation y=0 grapheme data but row not marked
[page] (warn): page integrity violation style ref count mismatch id=1 expected=10 actual=9
[page] (warn): page integrity violation zero row count
[page] (warn): page integrity violation zero col count
[terminal] (warn): alternate screen resize failed, replacing it err=error.OutOfMemory
[terminal] (warn): alternate screen resize failed, replacing it err=error.OutOfMemory
[terminal] (warn): alternate screen resize failed, replacing it err=error.OutOfMemory
[terminal] (warn): restoreCursor error updating style err=error.OutOfSpace

The page allocator is being corrupted during teardown before the FD race fires.

Repro

• Reliable on a heavy test root (3900+ tests, 4GB RSS peak): reproduces ~2/3 of runs.
• Hidden on the default test root (76 tests, 26MB RSS): did not reproduce across 59 attempts across 4 invocation methods including --test-filter terminal.search.Thread.test_0 and full-suite 25-iteration loops.

The race is latent; memory and threading pressure is what surfaces it.

Suggested fix

Either:

1. Guard self.loop.run(.once) in Thread.zig::threadMain_ (Thread.zig:197 and :211) with if (!self.loop.stopped()) so a stop notification that races with the search thread's loop.run doesn't trigger a kevent on a closing FD.
2. Upstream a recoverable BADF path to libxev's kqueue.zig:1916 — the .BADF => unreachable should re-arm the loop or exit cleanly when the loop is being torn down, since the comment "Always a race condition" is already admitting this is a known foot-gun, not an actual programmer error.

Environment

• macOS aarch64 (Apple Silicon)
• Zig 0.16.0_1
• libxev 0.0.0 (vendored at zig-pkg/libxev-.../)
• Fork: https://github.com/KooshaPari/Khostty