[bvolpato]

Summary

Let applications waiting on non-storage events opt out of Linux I/O-wait accounting for blocking io_uring waits. Keep the existing default policy.

Related to #216. The opt-in, feature-gated policy follows the existing reproductions and the October 21 proposal. This implements that policy on current Zig 0.17 main.

Impact

Blocking wait	Before	After
Default options	GETEVENTS (0x1)	Unchanged
Opt-in, kernel advertises support	GETEVENTS (0x1)	`GETEVENTS
Opt-in, feature bit absent	GETEVENTS (0x1)	Unchanged

Nonblocking paths keep their existing behavior. This changes accounting policy. Application latency, throughput, and power benefits have not been established.

Changes

Add the io_uring-specific no_iowait option, defaulting to false. Gate the flag on a positive wait and IORING_FEAT_NO_IOWAIT (Linux 6.15+), covering both combined submit/wait and waits after submission overflow. No Ghostty adoption or C API change is included.

Testing

• zig build test -Demit-man-pages=false -j2 --summary all passed on Linux x86_64 with Zig 0.17.
• Independently rebuilt eventfd fixtures and syscall traces confirmed the flag behavior above for ordinary/overflow waits. Static and dynamic epoll accept the shared option.
• The committed regression covers feature gating and completion behavior. The separate trace checks observe syscall flags. Unsupported-kernel fallback was simulated by masking the feature bit. Older kernels and non-Linux runtimes remain untested.
• Zig 0.17 requests the same preexisting enum-conversion migration at base and head. Scoped Zig 0.16 formatting passes.

Upstream CI awaits maintainer approval for this fork's workflow.