fix: on Windows, a reaped child closes the terminal

[vyncint]

Refs #149 — the first item of step 4, found by running the migrated suite (#276) on windows-latest.

The bug

Every liveness decision reads EOF off the master: ensure_deliverable (the "child is gone" refusals in send, resize, and the mouse path) and the fast Error::Eof every wait returns instead of burning its timeout. On Unix that is the right signal, for the reason the code already states — EOF means every slave descriptor is closed, while a reaped child says nothing about a grandchild still reading. On Windows it never arrives: a pseudoconsole's output pipe stays open until the console itself is closed, whoever has exited. Five tests fail there identically, with the child gone and the harness insisting someone is listening:

a_departed_child_is_refused_identically_on_every_platform · typed_input_to_a_departed_child_is_a_typed_error · a_mouse_click_at_a_departed_child_blames_the_child · resize_after_the_child_exits_is_refused_like_send · waits_fail_fast_on_eof_instead_of_burning_the_timeout

The fix

Where EOF cannot come, the reaped child is the signal — one named constant, EXIT_CLOSES_THE_TERMINAL = cfg!(windows), with the reasoning on it.

• ExitWatch::tick runs inside each wait loop's poll (wait_until, wait_frame, wait_idle, wait_stable): a non-blocking try_wait; once the child is reaped and DRAIN_GRACE has passed with the terminal still open, state.eof = true and the wait returns exactly what it would have on Unix.
• wait_exit already drained for DRAIN_GRACE waiting on EOF; when that expires on Windows it now marks the terminal closed, so the refusals that follow a wait_exit fire.
• A no-op where EOF is real: Unix behaviour is byte-for-byte what it was, and the whole suite here says so.

Verification

Unix: full suite, three feature configurations, both clippy targets. Windows: windows.yml dispatched against this branch — the five tests above are the acceptance; result in a comment below.

[vyncint]

windows.yml against this branch — run 34183062320

All five acceptance tests pass on windows-latest:

ok  a_departed_child_is_refused_identically_on_every_platform
ok  typed_input_to_a_departed_child_is_a_typed_error
ok  a_mouse_click_at_a_departed_child_blames_the_child
ok  resize_after_the_child_exits_is_refused_like_send
ok  waits_fail_fast_on_eof_instead_of_burning_the_timeout

Suite there: 351 passed / 86 failed (345 / 92 on main). The first version of this branch got four of the five: waits_fail_fast_on_eof… still burned its timeout, because wait_until skips a wake whose generation is unchanged before it ever looks at eof — and a synthesized EOF changes no byte. The watch now bumps the generation when it marks the stream closed, which is the honest description anyway: EOF is a fact about the stream even when no byte carried it.

The macOS leg's one failure on the first push was inspect_prints_its_usage_for_help_and_for_a_missing_program with NotFound on the example binary — inspect.rs rebuilds the example from every test with no serialization, unlike fixture_bin's BUILT mutex, so two tests can race a relink. Unrelated to this change (the file is untouched), green on the re-run, filed separately.