Fix df.signal not propagated into child sub-orchestrations

[Copilot]

• Inspect current signal/race implementation and existing E2E tests
• Check recent GitHub Actions workflow runs/log context for this issue
• Run baseline local validation (format/build/tests) before code changes
• Implement minimal fix so df.signal reaches relevant child sub-orchestrations (race/wait_for_signal case)
• Add focused E2E coverage for wait_for_signal inside df.race
• Run targeted validation for changed paths, then broader checks as needed
• Run parallel validation (Code Review + CodeQL) and address any valid findings

[pinodeca]

Review followups from the deep-dive on this change. Filing them here so they aren't lost; all are intentionally deferred — either because they're properly solved by the upstream duroxide fix tracked in microsoft/duroxide#21, or because they're polish that doesn't gate this PR.

#	Concern	Where	Status
1	Race: signal raised before the sub-orchestration is Running (or even scheduled) is delivered to the root only, lost when the child later subscribes.	raise_external_event fan-out logic	Deferred — proper fix needs unmatched-event forwarding in duroxide (microsoft/duroxide#21).
2	Pending descendants (scheduled but not yet picked up by the worker) are filtered out by the status == "running" check and miss the signal.	list_running_descendants	Deferred — same upstream fix removes the need to inspect child state at all.
3	Fan-out is a true broadcast: every running descendant receives every signal regardless of whether it has a matching schedule_wait. Harmless today but couples signal namespaces across unrelated branches.	raise_external_event	Deferred — subsumed by microsoft/duroxide#21 (option 1 / option 3 in that issue eliminate the broadcast).
4	N+1 round-trips per df.signal() call: one get_instance_tree + one get_instance_info per descendant. Fine for the common small-fanout case, would hurt for wide JOIN/RACE trees or hot signal paths.	list_running_descendants	Deferred — not worth optimizing the workaround; goes away with upstream fix.
5	Status check uses eq_ignore_ascii_case("running") against duroxide's free-form status string. Fragile if duroxide ever renames/recases the variant.	list_running_descendants	Deferred — same lifetime as the workaround.
6	continue_as_new on a parent does not GC prior-iteration sub-orchestration rows; long-running loops with a JOIN/RACE body accumulate child instances in duroxide.instances indefinitely, slowing get_instance_tree over time.	Pre-existing duroxide behavior, surfaced by this PR's reliance on get_instance_tree.	Deferred — track separately; not introduced by this PR.

Filed microsoft/duroxide#21 to request the upstream primitive (unmatched-event forwarding or an explicit raise_event_tree) that would replace this best-effort fan-out with a race-free, single-call delivery. Once that lands, this client-side workaround should be removed.