fix(workflow): harden terminal cleanup races

Queue worker results before settlement cleanup, claim terminal and death boundaries before provider callbacks, and close late-message admission.

Make child cancellation and disposal reentrancy-safe across the workflow bridge and generic subagent wrapper, with adversarial regression coverage and RFC documentation.

packages/workflow/workflow-workerthread/README.md

@@ -25,7 +25,7 @@ What the seam guarantees regardless, because benign scripts hit these constantly
 
 Inside the worker, `runWorkerSession` builds the execution core (hooks, combinators, concurrency semaphore, caps, fatal-error discipline) over a **child port**. `agent()` sends `child-start`, and the host starts the child through the holder-bound `SubagentService` handle captured synchronously by `start()`, with parent attribution, the shared per-run abort signal, and `outputSchema`/`model` pass-through. This capture is part of the seam's holder-owned lifetime: unloading the engine removes `ctx.workflows` for new calls but does not invalidate an already returned run whose worker starts another child afterward.
 
-The host observes `run.result` immediately but buffers its snapshotted wire projection until `run.started` fulfills. It then replies `child-started` with the child id before forwarding settlement, so `workflow/agent-start` always names a ready child and precedes its end. A readiness rejection replies `child-start-error`, emits no workflow agent pair, and makes the host dispose the attempt because the worker never received a handle; the worker classifies it as fatal `AGENT_START` unless cancellation already owns the run. If readiness fulfills, an infrastructure result rejection crosses as `child-failed`/`AGENT_RESULT` regardless of whether that rejection settled before or after readiness. Child disposal acknowledgements complete the RPC.
+The host observes `run.result` immediately but buffers its snapshotted wire projection until `run.started` fulfills. Provider `start()` is arbitrary code and can synchronously reenter workflow cancellation before its returned run reaches the host registry, so the host registers the run, attaches both promise observers, and re-checks admission after `start()` returns and again at readiness. A closed boundary never admits or announces the run to the worker: while the exact run remains registered, the host invokes explicit cancel once and disposes it; `child-start-error` is sent only while worker-message admission remains open. If the run was already retired, the identity guard sends no cleanup through the deleted call ID. An ordinary readiness rejection sends `child-start-error` while possible and disposes the provider attempt without adding an explicit cancellation. Otherwise the host replies `child-started` with the child id before forwarding settlement, so `workflow/agent-start` always names a ready child and precedes its end. The worker classifies a start error as fatal `AGENT_START` unless cancellation already owns the run. If readiness fulfills, an infrastructure result rejection crosses as `child-failed`/`AGENT_RESULT` regardless of whether that rejection settled before or after readiness. Child disposal acknowledgements complete the RPC.
 
 Observer narration (`phase`/`log`/`agent-start`/`agent-end`) crosses as messages and re-emits as the seam's `workflow/*` events. A **ready→go handshake** gates the body: a cancellation racing worker boot arrives before `go`, so a run cancelled before start never executes the body at all.
 
@@ -35,9 +35,15 @@ Values LEAVING the script (hook options/schemas, the script's return) are materi
 
 ## Cancellation, death, disposal
 
-Per-run limits: a concurrency semaphore (`maxConcurrentAgents`), a total-`agent()` cap (`maxTotalAgents`), and a per-call item cap (`maxItemsPerCall`), all config. `cancel()` posts the cancel to the worker (its hooks start throwing `CANCELLED`; the script dies at its next await) and cancels every host-side child NOW on **both seam channels** — the shared request signal aborts AND each registered child's explicit `cancel()` is called host-side, because the seam leaves a provider free to honor either channel and a worker wedged in a synchronous spin could not relay its own per-child cancel RPCs (those later land as idempotent no-ops). Each provider-owned cancel callback is exception-contained independently, so a broken child cannot prevent peer cancellation or workflow settlement. The caller's optional start-signal callback is retained by exact identity only while the run is live and removed at the first settlement or teardown, so a long-lived signal cannot retain completed `WorkerRun` instances. The grace then arms: a run still unsettled `disposeGraceMs` later force-settles `cancelled` and the worker is **terminated**. A cancellation that lands before the body runs (the ready→go handshake) reports `cancelled` without executing anything; a worker `result` racing an in-flight host cancellation reports `cancelled` too (first-wins settlement — the seam-visible result had not settled when cancellation was requested); post-cancel `phase`/`log` narration is suppressed host-side, while cancelled children still deliver their paired `agent-end`.
+Cancellation is bounded and host-driven. Per-run limits are a concurrency semaphore (`maxConcurrentAgents`), a total-`agent()` cap (`maxTotalAgents`), and a per-call item cap (`maxItemsPerCall`), all config. `cancel()` first records its reason, then posts to the worker (its hooks start throwing `CANCELLED`; the script dies at its next await) and cancels every host-side child NOW on **both seam channels**: the shared request signal aborts and each registered child's explicit `cancel()` runs host-side. The seam leaves a provider free to honor either channel, and a worker wedged in a synchronous spin could not relay its own per-child cancel RPCs. A host-side per-call gate turns the worker's later explicit-cancel relay into a no-op, because the seam does not require `SubagentRun.cancel()` to be idempotent. Each explicit child `cancel()` callback is exception-contained independently, post-cancel `phase`/`log` narration is suppressed host-side, and cancelled children still deliver paired `agent-end` events. The caller's optional start-signal callback is retained by exact identity only while the run is live and removed at the first settlement or teardown.
 
-A worker that dies unexpectedly (an OOM, a script reaching `process.exit` through the documented vm escape) settles the run `stopReason: 'error'` with the exit diagnostics — or `'cancelled'` when a cancel was in flight — and the host-side child registry is what winds every surviving child down. `dispose()` = cancel + immediate host-driven disposal of every registered child (a wedged worker can relay no dispose RPC, so child teardown overlaps the grace instead of starting after it; the worker's own dispose RPCs join the same per-child disposal) + bounded wait (result, then child-registry quiescence, capped by the grace) + unconditional `worker.terminate()`: the thread never outlives its run. Before an ordinary run settlement becomes observable, the host cancels every stray child on both channels too—even a fire-and-forget run still waiting on `started`, for which the worker has no handle yet—and `dispose()` then waits for their disposal (bounded by the grace) before returning. `agent-start`/`agent-end` pairing is host-guaranteed the same way: forwarded starts live in a ledger, worker-reported ends pair them on the graceful paths, and the termination paths (grace force-settle, worker death) synthesize the missing ends (outcome `cancelled`) before the run settles — a start still in flight across the force-settle can surface after `workflow/end`, immediately paired the same way.
+Terminal arbitration is first-wins at explicit host-side claim points. A cancellation before ready→go reports `cancelled` without executing the body. For a later race, the worker queues Result before its settlement-reap `ChildCancel` messages; external `cancel()` records its reason before its fanout, while Result receipt snapshots any earlier cancellation and records the terminal outcome before settlement-cleanup fanout. Same-port FIFO and those claim points mean earlier caller/signal/dispose cancellation overrides a non-cancelled report, while an arrived report cannot be rewritten by a cleanup callback. Once Result has won, a losing reentrant `cancel()` has no state, message, child-fanout, or grace-timer effect. If no earlier terminal source settles the run, the grace callback claims `cancelled`, synthesizes missing lifecycle ends, settles the result, and terminates the worker after `disposeGraceMs`.
+
+Worker death separates outcome ownership, message admission, and resource cleanup. An unexpected OOM, `error`, message failure, or premature exit claims `stopReason: 'error'` with diagnostics—or preserves an external cancellation already in flight—before reaping children or synthesizing observer events. Reentrant provider cancellation therefore cannot turn a death-first error into cancellation. The first death signal also closes worker-message admission because Node may deliver a queued `message` between `error` and `exit`; late protocol data cannot start a child, emit narration, or compete with the outcome. If Result or grace already claimed the outcome, death preserves it while still reaping promptly. The eventual `exit` then performs a final disposal-only sweep, joining any in-flight disposal without repeating explicit child cancellation. This separation lets grace settlement become observable before `worker.terminate()` reports exit without leaking the host-side registry.
+
+Disposal is the holder's bounded resource guarantee: cancel, begin host-driven disposal of every registered child immediately, wait for result plus child-registry quiescence up to the same grace, and unconditionally terminate the worker. `handle.dispose()` claims its public promise before that traversal invokes cancellation or disposal callbacks. Independently, every `disposeChild` path claims the call ID's promise before invoking the wrapped child disposer. Public-first reentry therefore returns the existing holder promise; worker-first reentry may begin holder disposal, whose child traversal joins the already-claimed call ID promise. Neither order can start a second provider disposal. A wedged worker can relay no dispose RPC, so host-driven teardown overlaps the grace; any later worker RPC joins the same per-child disposal. Before ordinary settlement becomes observable, the host also cancels every stray on both channels, including a fire-and-forget run still waiting on readiness. That work is settlement-only cleanup after the terminal claim, so provider reentry cannot rewrite the chosen result; `dispose()` then waits for its completion within the bound.
+
+Lifecycle pairing is host-guaranteed independently of outcome arbitration. Forwarded starts live in a ledger and worker-reported ends pair them on graceful paths. When death or grace is the terminal source, the host synthesizes missing ends with outcome `cancelled` before `workflow/end`. If Result settled first, later death cleanup may synthesize a survivor's end afterward; a start already crossing force-settlement may likewise surface after `workflow/end`. The same ledger still pairs every forwarded start exactly once.
 
 **Engine-specific limitations**: worker startup is paid per run; on a termination path `agentsStarted` reports the HOST-observed count (accepted `child-start`s — calls still queued worker-side for a concurrency slot are unknowable then); and a returned promise or thenable resolves per JavaScript semantics BEFORE materialization — that is what makes an un-awaited `return agent('x')` work — with the value-boundary guard applying to the resolution.