Transport proof method for #15622 (not a committed test)

Pinned production head: fad9b847a7e6660b7babbfaeaf35bff8d5d02256.
Real ephemeral Node HTTP listener, authenticated JSON-RPC initialize and tools/call over MCP HTTP/SSE, real OrchestratorMcpService, real ThreadManagementService, real orchestrator/event sink/effect worker, migrated in-memory SQLite. Synthetic bearer scope and deterministic external-provider adapter only; no actual provider executable, subscription, live home, scheduled task, or remote deployment.

Used the existing integration fixture helpers in a temporary version of the existing toolkit test file. The Python launcher restores its exact bytes in finally, including on failure/interruption. Nothing from the one-off harness is committed. Wall-clock budgets have 7-second overhead tolerance; exact one-millisecond boundaries are separately proven with TestClock.

Earlier exploratory runs were interrupted during startup/import and one finished its six timing checks but failed due to the harness using an outdated TCP address tag. The final run uses the Effect HttpServer address formatter. These harness failures did not modify production.

Final recovery proof aborts an accepted HTTP wait, lists subagents, retries the identical mode=wait call with the same request key in the same active run/session, and verifies the same task/child, five durable tasks, and persisted completionWake=always. After releasing child gates it awaits a Deferred marking five continuation offers, checks that the recovered child belongs to the persisted delivery cohort, and matches a continuation offer to that cohort's parent run/generation/message. Only then does it acknowledge terminal tasks via task_status. This proves the cohort was offered at the probe; it does not prove delivery by a real provider CLI.

An intermediate repeat incorrectly inspected a notification field rather than the delegated completion cohort reference; it was stopped by its captured session and restored the test file. The corrected final repeat passed. Disconnect-triggered wake upgrades without an identical wait retry remain outside this change.