fix(args): treat the fully-async rollout path as the mode that selects it

[Shi-Dong]

Problem

--fully-async selects FullyAsyncRolloutFn, but the class can also be named directly:

--rollout-function-path miles.rollout.fully_async_rollout.FullyAsyncRolloutFn

That spelling loads the same persistent producer while args.fully_async stays False, so the run silently skips everything the mode implies:

• the mode's validation — --colocate, --partial-rollout, legacy rollout v1, --pause-generation-mode abort, multi-LoRA, and the rest are all checked inside if args.fully_async:
• the guard in train.py that refuses the sync driver, so the producer can run with no async driver at all

examples/infra_features/fully_async/run_qwen3_5_4b_fully_async_eval.py uses exactly this spelling today. The mismatch surfaces during training rather than at launch.

Change

Normalize the path spelling into the flag before any validation runs: when --rollout-function-path names FullyAsyncRolloutFn, enable --fully-async and let the flag own the selection. The two spellings are one selection, so the existing "pass only one" assertion now fires only when the path names a different function.

The path string becomes a shared constant, so the override and the normalization cannot drift apart. Only the exact class is recognized: a subclass still passes --fully-async explicitly.

Behavior

Arguments	Before	After
--fully-async	mode on	unchanged
--rollout-function-path <FullyAsyncRolloutFn>	producer on, mode off	mode on
both, agreeing	AssertionError	mode on
--fully-async + a different path	AssertionError	unchanged
any other path	plugin	unchanged

A run that was already correct is unaffected. A run that was silently wrong now either works or fails at launch with the mode's own message.

Tests

Five cases in tests/fast/utils/test_arguments.py: the path enables the mode, the mode's constraints then apply (--colocate is rejected), the agreeing pair is accepted, a conflicting pair still raises, and an ordinary plugin path is untouched.

Verified locally:

• the five new cases pass; on unmodified main the path-plus---colocate combination is accepted and returns fully_async=False
• tests/fast/rollout/test_fully_async_rollout.py and tests/fast/test_train_async.py: 34 passed
• all pre-commit hooks pass

No GPU run.

Unrelated file: scripts/ci/runner_utilization_report.py

That file landed in #3586 unformatted, so the repo-wide pre-commit job fails on every branch cut from main since — including this one, on a file it does not touch. The last commit here is black alone (the pinned 24.3.0, applied through the repo's own hook); the file's AST is unchanged. pre-commit run --all-files passes with it and fails without it. Drop that commit if main is fixed first.

[claude]

Claude Code Review

This repository is configured for manual code reviews. Comment @claude review for a one-time review, or @claude review always to subscribe this PR to a review on every future push.

_{Tip: disable this comment in your organization's Code Review settings.}

[Shi-Dong]

@claude review always

[claude]

Code review found no issues

No high-confidence issues detected in this change.

[claude]

Code review found no issues

No high-confidence issues detected in this change.

[claude]

Code review found no issues

No high-confidence issues detected in this change.