Close late `ActorNursery` registration race

A child could pass the early `.start_actor()` guard, miss the
`.cancel()` child snapshot and register afterward. Its monitor
inherited a reap request without runtime cancellation and could wait
forever.

- publish child/reap events before sampling `_cancel_called`
- make MP abort before `proc.start()` when cancellation won
- kill a Trio child opened after cancellation won registration
- reject starts begun after nursery cancellation is already visible
- add deterministic registration and MP no-start regressions
- drop the touched Trio backend's stale `get_runtime_vars` import

Prompt-IO: ai/prompt-io/opencode/20260821T040803Z_3c1bbe73_prompt_io.md

(this patch was generated in some part by `opencode` using
`gpt-5.6-sol` (`openai`))

ai/prompt-io/opencode/20260821T040803Z_3c1bbe73_prompt_io.md

@@ -0,0 +1,57 @@
+---
+model: openai/gpt-5.6-sol
+service: opencode
+session: 76c5d31c-5a2f-4503-9b16-410ee7f4fab3
+timestamp: 2026-08-21T04:08:03Z
+git_ref: 3c1bbe73
+scope: code
+substantive: true
+raw_file: 20260821T040803Z_3c1bbe73_prompt_io.raw.md
+---
+
+## Prompt
+
+Continue PR #481 review cleanup one item at a time. Compare the
+remaining child-registration/reaping experiment against PR #484,
+identify the next valid behavior, and generate an exact commit plan for
+that boundary without committing automatically.
+
+## Response summary
+
+Separated late-child admission from the broader bounded-reap rewrite.
+Added one synchronous `ActorNursery._register_child()` publication step
+that stores the child and its reap events before sampling nursery
+cancellation. This closes both scheduler orderings around
+`ActorNursery.cancel()`'s child snapshot.
+
+The MP backend registers immediately before synchronous process startup
+and refuses to start when cancellation already owns the child. The Trio
+backend registers immediately after `open_process()` and kills that
+already-created process when registration observes cancellation. An
+early `start_actor()` guard rejects calls begun after cancellation is
+already visible.
+
+Deterministic tests cover the nursery registration ordering and the MP
+no-start invariant. Comparison with PR #484 confirmed that its retained
+generic nursery/backends do not close this race.
+
+## Files changed
+
+- `tractor/runtime/_supervise.py` - atomically publish child ownership
+  and reject actor starts after nursery cancellation.
+- `tractor/spawn/_mp.py` - register before synchronous process startup
+  and abort a cancellation-owned child.
+- `tractor/spawn/_trio.py` - register immediately after process creation,
+  kill a cancellation-owned child, and remove its stale unused import.
+- `tests/test_to_actor.py` - cover late registration and MP startup
+  suppression.
+
+## Human edits
+
+The human required review extras to be handled one item and one
+behavioral commit at a time, with each item compared against PR #484
+before acceptance. That direction split this late-registration fix from
+the original broad experiment's bounded post-ack reaping,
+`ActorNursery.cancel()` hard-reap rewrite, and debugger/error behavior.
+The human accepted the narrower late-registration boundary by requesting
+its commit plan. No direct source-line edits were made by the human.

ai/prompt-io/opencode/20260821T040803Z_3c1bbe73_prompt_io.raw.md

@@ -0,0 +1,48 @@
+---
+model: openai/gpt-5.6-sol
+service: opencode
+timestamp: 2026-08-21T04:08:03Z
+git_ref: 3c1bbe73
+diff_cmd: git diff HEAD~1..HEAD
+---
+
+Compare the remaining child-registration and reaping experiment with
+PR #484, then identify the next review item without changing code.
+
+The next item is the late-child admission race. A spawn can pass
+`ActorNursery.start_actor()`'s early cancellation check, then be absent
+from `ActorNursery.cancel()`'s child snapshot and register afterward.
+The existing reap-request latch releases its monitor but does not send
+runtime cancellation, so the monitor can wait forever for a still-live
+process.
+
+> `git diff HEAD~1..HEAD -- tractor/runtime/_supervise.py`
+
+`ActorNursery._register_child()` publishes the child, installs its reap
+events, and samples `ActorNursery._cancel_called` without a checkpoint.
+The two scheduler orderings are then complete: registration first puts
+the child in the cancel snapshot, while cancellation first makes the
+backend abort the late registration.
+
+> `git diff HEAD~1..HEAD -- tractor/spawn/_mp.py`
+
+The multiprocessing backend registers immediately before `proc.start()`
+and refuses to start a process already owned by nursery cancellation.
+There is no Trio checkpoint between registration and process startup.
+
+> `git diff HEAD~1..HEAD -- tractor/spawn/_trio.py`
+
+The Trio backend registers immediately after `open_process()` and kills
+the newly opened process if cancellation won the registration race. Its
+stale unused `get_runtime_vars` import is removed so the touched module
+remains lint-clean.
+
+> `git diff HEAD~1..HEAD -- tests/test_to_actor.py`
+
+Deterministic regressions prove late registration observes cancellation
+and that the MP backend never starts a process after cancellation owns
+its registration.
+
+PR #484 retains the affected generic nursery and spawn-backend paths and
+does not close this race. Keep this fix in PR #481 as its own commit;
+review bounded post-`CancelAck` reaping separately.

tests/test_to_actor.py

@@ -8,6 +8,7 @@ https://github.com/goodboy/tractor/issues/477
 '''
 from functools import partial
 from pathlib import Path
+from types import SimpleNamespace
 
 import pytest
 import trio
@@ -21,6 +22,7 @@ from tractor._testing import tractor_test
 from tractor._exceptions import ActorTooSlowError
 from tractor.msg import ptr as msgptr
 from tractor.msg.ptr import NamespacePath
+from tractor.spawn import _mp as mp_spawn
 from tractor.to_actor import _api as to_actor_api
 
 
@@ -317,6 +319,125 @@ def test_cancel_actor_timeout_closes_blocked_send():
     )
 
 
+def _mock_actor_nursery() -> tractor.ActorNursery:
+    an = object.__new__(tractor.ActorNursery)
+    an._children = {}
+    an._join_procs = trio.Event()
+    an._child_reap_requests = {}
+    an._child_reaped = {}
+    an._at_least_one_child_in_debug = False
+    an._cancel_called = False
+    return an
+
+
+def test_late_child_registration_observes_cancel():
+    '''
+    Make registration atomically observe nursery cancellation.
+
+    `ActorNursery.cancel()` previously snapshotted `_children` before
+    its next checkpoint. A process monitor registering after that
+    snapshot received a reap request but no runtime cancellation, then
+    waited forever for natural exit. Publishing the child and its reap
+    events together returns cancellation ownership to the late monitor.
+
+    '''
+    an = _mock_actor_nursery()
+    an._cancel_called = True
+    aid = tractor.msg.Aid(
+        name='late_child',
+        uuid='test',
+    )
+    subactor = SimpleNamespace(aid=aid)
+    proc = object()
+
+    (
+        reap_request,
+        reaped,
+        cancel_during_registration,
+    ) = an._register_child(
+        subactor,
+        proc,
+        None,
+    )
+
+    assert cancel_during_registration
+    assert an._children[aid.uid] == (
+        subactor,
+        proc,
+        None,
+    )
+    assert an._child_reap_requests[aid.uid] is reap_request
+    assert an._child_reaped[aid.uid] is reaped
+
+
+def test_mp_late_registration_never_starts_process(
+    monkeypatch: pytest.MonkeyPatch,
+):
+    '''
+    Refuse to start an MP child already owned by nursery cancellation.
+
+    A concurrent `ActorNursery.cancel()` can publish cancellation after
+    `start_actor()` checks its flag but before the MP backend registers
+    its process. The fake registration reports that exact schedule.
+    Proving `FakeProcess.start()` is never called prevents a child from
+    starting after it was omitted from the cancellation snapshot.
+
+    '''
+    class FakeProcess:
+        started: bool = False
+
+        def start(self) -> None:
+            self.started = True
+
+    process = FakeProcess()
+
+    class FakeContext:
+        def get_start_method(self) -> str:
+            return 'spawn'
+
+        def Process(self, **kwargs: object) -> FakeProcess:
+            assert kwargs
+            return process
+
+    nursery = SimpleNamespace(
+        _register_child=lambda *args: (
+            trio.Event(),
+            trio.Event(),
+            True,
+        ),
+    )
+    subactor = SimpleNamespace(
+        aid=tractor.msg.Aid(
+            name='late_mp_child',
+            uuid='test',
+        ),
+    )
+    monkeypatch.setattr(
+        mp_spawn._spawn,
+        '_ctx',
+        FakeContext(),
+    )
+
+    with pytest.raises(
+        RuntimeError,
+        match='nursery began cancelling',
+    ):
+        trio.run(
+            partial(
+                mp_spawn.mp_proc,
+                name='late_mp_child',
+                actor_nursery=nursery,
+                subactor=subactor,
+                errors={},
+                bind_addrs=[],
+                parent_addr=SimpleNamespace(),
+                _runtime_vars={},
+            )
+        )
+
+    assert not process.started
+
+
 def test_late_child_reap_registration_is_released():
     '''
     Preserve a nursery-wide reap request across child startup.

tractor/runtime/_supervise.py

@@ -327,6 +327,29 @@ class ActorNursery:
             reap_request.set()
         return reap_request, reaped
 
+    def _register_child(
+        self,
+        subactor: Actor,
+        proc: 'ProcessType',
+        portal: Portal|None,
+    ) -> tuple[trio.Event, trio.Event, bool]:
+        '''
+        Atomically publish one child and its reap coordination.
+
+        '''
+        uid: tuple[str, str] = subactor.aid.uid
+        self._children[uid] = (
+            subactor,
+            proc,
+            portal,
+        )
+        reap_request, reaped = self._register_child_reap(uid)
+        return (
+            reap_request,
+            reaped,
+            self._cancel_called,
+        )
+
     def _request_reap_all(self) -> None:
         '''
         Release every child monitor into its process-join phase.
@@ -419,6 +442,12 @@ class ActorNursery:
 
         '''
         __runtimeframe__: int = 1  # noqa
+        if self._cancel_called:
+            raise RuntimeError(
+                'Cannot start an actor in a cancelling '
+                '`ActorNursery`'
+            )
+
         loglevel: str = (
             loglevel
             or self._actor.loglevel

tractor/spawn/_mp.py

@@ -138,12 +138,22 @@ async def mp_proc(
         # daemon=True,
         name=name,
     )
-
-    # `multiprocessing` only (since no async interface):
-    # register the process before start in case we get a cancel
-    # request before the actor has fully spawned - then we can wait
-    # for it to fully come up before sending a cancel request
-    actor_nursery._children[subactor.aid.uid] = (subactor, proc, None)
+    # `multiprocessing` only (since no async interface): publish the
+    # process and its reap coordination before start so cancellation
+    # can own every subsequently started child.
+    (
+        reap_request,
+        _,
+        cancel_during_registration,
+    ) = actor_nursery._register_child(
+        subactor,
+        proc,
+        None,
+    )
+    if cancel_during_registration:
+        raise RuntimeError(
+            'Actor registered after its nursery began cancelling'
+        )
 
     proc.start()
     if not proc.is_alive():
@@ -170,9 +180,6 @@ async def mp_proc(
         # any process we may have started.
 
         portal = Portal(chan)
-        reap_request, _ = actor_nursery._register_child_reap(
-            subactor.aid.uid,
-        )
         actor_nursery._children[subactor.aid.uid] = (subactor, proc, portal)
 
         # unblock parent task

tractor/spawn/_trio.py

@@ -39,7 +39,6 @@ from tractor.runtime._state import (
     current_actor,
     is_root_process,
     debug_mode,
-    get_runtime_vars,
 )
 from tractor.log import get_logger
 from tractor.discovery._addr import UnwrappedAddress
@@ -131,6 +130,23 @@ async def trio_proc(
                 f' |_{proc}\n'
             )
 
+            (
+                reap_request,
+                _,
+                cancel_during_registration,
+            ) = actor_nursery._register_child(
+                subactor,
+                proc,
+                None,
+            )
+            if cancel_during_registration:
+                cancelled_during_spawn = True
+                proc.kill()
+                raise RuntimeError(
+                    'Actor registered after its nursery began '
+                    'cancelling'
+                )
+
             # wait for actor to spawn and connect back to us
             # channel should have handshake completed by the
             # local actor by the time we get a ref to it
@@ -161,9 +177,6 @@ async def trio_proc(
         assert proc
 
         portal = Portal(chan)
-        reap_request, _ = actor_nursery._register_child_reap(
-            subactor.aid.uid,
-        )
         actor_nursery._children[subactor.aid.uid] = (
             subactor,
             proc,