[ananthsub]

What changed and why

A participant's prepare waits for the participant to become ready before it replies.
While it waits, its checkpoint can end.
For example, the controller goes quiet, the checkpoint's lease expires, and the participant reopens admission.
The controller can then prepare a second checkpoint, which reaches prepared before the first prepare wakes.

When the first prepare woke, ParticipantControlPlane.prepare in nemo_gym/_checkpoint/control.py replied with self.phase, the phase of whichever checkpoint was open by then.
In the example above, the first prepare reported prepared for a checkpoint that had already ended.

The participant's state stayed safe.
The lease expiry had fenced the first checkpoint's ID, so a commit for it is refused as stale_checkpoint.
Only the reply was wrong, and a controller reading it could believe its own checkpoint was prepared.

A prepare whose checkpoint is no longer the open one now replies idle, together with the current readiness report.
This follows the existing reply for a prepare that a resume interrupts, which already reported idle.
The controller in nemo_gym/_checkpoint/coordination.py treats any phase other than prepared as not prepared.
Once a resume or lease expiry has ended the checkpoint, a later prepare or commit for it is refused as stale_checkpoint.

A stale_checkpoint refusal was the other candidate reply.
It was not chosen because the reply to a prepare interrupted by a resume, which is the same situation, is already idle, and a unit test pins that reply.
Using one reply for both keeps the contract the same for every way a checkpoint can end while a prepare waits.

A prepare that wakes after a concurrent prepare of the same checkpoint finished it still reports that checkpoint's phase, because that phase belongs to its own checkpoint.

Issue

None needed.
This is a small fix for a race found by a test in the checkpoint stack's fault tests (#4197).

Validation

A new unit test in tests/unit_tests/test_checkpoint_control.py holds the first prepare at a gate while it waits for readiness.
While it is held, the lease expires, admission reopens, and a second checkpoint is prepared.
The test then releases the first prepare and checks three things: the second prepare reports prepared, the first prepare reports idle rather than the second checkpoint's phase, and a commit for the first checkpoint is refused as stale_checkpoint.
The test uses events rather than fixed sleeps, apart from waiting for the lease to expire.
On main without the fix, it fails because the first prepare reports prepared.

Commands, run with PYTHONPATH set to the worktree:

python -m pytest tests/unit_tests/test_checkpoint_*.py -n 16
python -m pytest tests/unit_tests -n 16 -m "not sandbox"
ruff check nemo_gym/_checkpoint/control.py tests/unit_tests/test_checkpoint_control.py
ruff format nemo_gym/_checkpoint/control.py tests/unit_tests/test_checkpoint_control.py
pre-commit run --files nemo_gym/_checkpoint/control.py tests/unit_tests/test_checkpoint_control.py

Rollout evidence

N/A.
The change affects only the control-plane reply to a prepare whose checkpoint ended while it waited, and a deterministic unit test covers that case.
No environment, agent, or verifier behavior changes.

Compatibility

The reply to /ng-control/v1/checkpoint/prepare changes only in this race: it reports idle where it used to report another checkpoint's phase.
The reply shape is unchanged.
A controller that checked for prepared now correctly sees its checkpoint as not prepared.

[macandro96]

good catch! lgtm

[github-actions]

🔄 Merge queue validation started! You can track the progress here: https://github.com/NVIDIA-NeMo/Gym/actions/runs/38833982337