Resolve native macOS stdout-capture regressions on both architectures

[LunaStev]

Observed at PR #631 head d811e0c2accc8997eecb5071f8752f88e9fb44e4. Failing CI job.

Both native macOS jobs fail the same two tests in tools.test_std_io_runtime:

• test_capture_has_no_inherited_pipe_reader_and_no_parent_leak: fixture exits 12.
• test_capture_writer_observes_consumer_closure: fixture exits 11.

capture.wave exit 12 combines child exit status, byte count and payload checks. capture_close.wave exit 11 combines wait failure and non-normal child termination. Logs contain no child status or payload, so they do not establish a leaked reader, exec failure, or SIGPIPE root cause.

Reproduce on each native host with the repository std installed:

cargo test --locked --test std_io_regressions --jobs 2 -- --nocapture

Trace fork return conventions, descriptor remapping/closure, script execution and child wait status. Compare a minimal executable child with the generated Python shebang child to distinguish runtime defects from fixture assumptions. Retain the closed-standard-descriptor and early-consumer-closure checks.

Acceptance:

• Both capture scenarios pass repeatedly on native amd64 and arm64 macOS.
• A missing child reader is demonstrated by descriptor inspection and bounded producer termination.
• Preserve child status and capture contents when a check fails.
• No libc binding workaround and no platform skip to hide failure.

Follow-up to #559 after its child-reader cleanup; related to broader Darwin provider audit #434. The two failures are intentionally grouped until evidence establishes different root causes.

Source references:

• tests/fixtures/io/capture.wave
• tests/fixtures/io/capture_close.wave
• tools/test_std_io_runtime.py
• std/process/spawn.wave

Both affected jobs: amd64, arm64.