fix(spawn): publish owned subprocess completion before retirement checks

[steipete]

After a blocked I/O callback outlasts a retirement deadline, the kernel may already have completed a child and its pipes while a deferred observer still sees stale completion state. This publishes owned subprocess completion before that observer without polling unrelated I/O or moving global event-loop phases.

Supersedes #142 by publishing only owned subprocess completion while preserving global phase order; #142 remains open for the coordinator.

The process-only kqueue/epoll snapshot retains ready owners before task dispatch and publishes them afterward. Existing readers deliver trailing bytes and real EOF; EAGAIN is never synthetic EOF. Extra stdio pipes count toward child close. The secondary descriptor closes when empty, stays above standard fds, and preserves the primary watch if additional watcher resources are exhausted.

HTTP/2 fixtures now bind to their IPv4 clients' address, and the stale-fd test selects the actual stdout socket rather than Darwin's FIFO-shaped kqueue descriptor. Assertions, deadlines, timeouts and skip conditions are unchanged.

Validation

Gate	Result
Original OpenClaw retirement, original deadline and six-second stall	Node passes; main fails; candidate 50/50 on Darwin and Linux
Reviewer-authored Node 24 controls	IO 10/10; stdio/held-pipe 50/50 on both
General ordering / nested reentry	All seven controls exactly match main; reentry 50/50 on both
Surrounding suites	110 files plus worker destruction on both; only intended regression improvement and matching baseline failures
Previously broken critical suites	Timers 24/24, HTTP2 lifecycle including synchronous POST plus server.stop(true) 23/23, HTTPS Agent reuse 30/30, TLS Duplex teardown 68/68
75-child Linux readiness	Node 75 exits, main zero, candidate 75; kernel/cleanup preconditions pass
Broker consumer	Candidate 14/14 on both; main passes the other 13
Isolated Gateway smoke	Node/main/candidate valid; all turns accounted for, zero errors/pending, accepted shutdown, exit 0, no forced cleanup
Rust / static checks / review	Rust 12/12; Clippy, TypeScript, format/source lints, Miri and Mordant green; P2 scoped-clean

Exact-head Linux/Darwin CI is green (27/27 and 21/21 report rows). Exact-head Windows CI is green for x64 and ARM64 builds, smoke and 31/31 compatibility rows each. Native artifact commit files/checksums match this head; the CI-built Darwin executable also passes the original consumer retirement fixture. Windows dispatch was a tagless dry run and Publish was skipped.

Manual comparisons retain one matching baseline failure per platform: native-Node TLS session resumption on Darwin, and an empty SHELL in the curated Linux environment. Main additionally fails the new regression. No failure was filtered or retried into a passing result.

Earlier failures and corrections are retained. An early Darwin HTTP2 run had 21 timeouts on a shared listener without listener-identity tracing, so its exact cause remains unproven. Controlled probes reproduced overlapping IPv4/IPv6 listeners; fresh main and candidate each passed all 84 original unchanged cases. No HTTP2 runtime workaround or tolerance change is included.

Related upstream work checked: oven-sh#33614, oven-sh#43888 and oven-sh#37869. None supplies this process-only undispatched-readiness mechanism; their broader scheduling or synthetic-EOF behavior is not imported.

Ready for coordinator review; intentionally unmerged. No release, tag, runtime-pin update or live/Team Gateway operation was performed.