Test native Windows replay descendant cleanup

[ywz65]

The registered replay launcher had native Windows timeout cleanup but no test that observed its descendants. This adds a process-handle observer for a ready parent, child and grandchild, answering the explicit Windows handoff in discussion #1472.

All three processes terminate under the real timeout path. A successful-looking parent-only kill is rejected because the two descendants remain live. Joined completion is the positive control. Retained handles, a checked taskkill result and a cleanup-duration check prevent PID reuse or eventual fixture self-exit from masquerading as successful cleanup.

Scope: ordinary descendants while the leader is alive. The separate already-exited-leader Windows gap in #2377 remains open. This changes no production launcher, scientific input, selected result or kappa.

Validation on native Windows 11 build 26300: all three tests pass on CPython 3.12.7 normal, CPython 3.12.7 -O, and CPython 3.14.5 normal, with no skips. A separate AI agent reviewed readiness, handles, teardown and failure controls with no blocking finding. Hosted CI and the full mathematical cold replay are not claimed.

python3.12 -B tests/test_research_windows_process_tree.py -v
python3.12 -O -B tests/test_research_windows_process_tree.py -v
python3.14 -B tests/test_research_windows_process_tree.py -v

Canonical base 0fdd7d82de03e43c9e5a1000dfbcd380d1b687d2, priority revision 4. Task portable-replay, explicitly changed to the named native coverage gap. Exact source hashes, overlap limits, scope, reproduction and open next questions are in docs/research/windows-replay-timeout-audit.md. The source lineage is Douglas Colkitt's launcher, Paul Romeo's merged #1484 Linux fixture and Windows handoff, and #1298's portability work. #2377's POSIX cleanup change is credited and not duplicated.

Prepared for ywz65 with substantial OpenAI Codex assistance in implementation, verification and writing. Apache-2.0; no independent human review is claimed.