Add unit test to prevent future regression of preset reward function

[mujtaba1747]

Thanks to Claude review bot for the callout!

By submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of your choice.

[github-actions]

Claude finished @mujtaba1747's task in 49s —— View job

---

Code Review

• Read PR diff (/tmp/pr.diff)
• Review the new test against rlvr_trainer.py
• Post review feedback

---

Verdict: LGTM ✅ — this is a clean, test-only change and I found no issues.

The new test_train_raises_when_no_reward_signal correctly guards the regression. I verified it against the source:

• The guard in RLVRTrainer.train() (rlvr_trainer.py:417-424) is the very first statement in train(), so with wait=False the test reaches it deterministically without needing the heavier session/config mocks the other tests use.
• getattr(self.hyperparameters, "preset_reward_function", None) is exactly the line under test, and Mock(spec=["to_dict"]) is the right technique here — a bare Mock() would auto-create a truthy preset_reward_function attribute and mask the guard, so spec=[...] genuinely forces the None/falsy path. The inline docstring explaining this is a nice touch for future maintainers.
• match="requires a reward signal" matches the raised message, and the custom_reward_function default (unset) satisfies the other half of the and condition.

Minor, non-blocking observations (no action needed):

• This test proves the "neither set" → raise case. If you want fuller coverage of the same guard, a complementary positive test (e.g. setting preset_reward_function and asserting train() gets past the guard) would pin down both branches — but that's optional and the existing test_peft_*/custom_reward_function tests already exercise the pass-through path indirectly.
• mock_finetuning_options / mock_validate_group are set up but the assertion only depends on the guard firing first; that's fine and consistent with the surrounding tests.

Nothing to change from my side.