fix: reject train-sampling temperature=0 (produces NaN logprobs)

[dumko2001]

Summary

• TrainSamplingConfig.temperature is Field(1.0, ge=0, le=2.0), so 0 is accepted. To a sampler 0 means greedy decoding, so it looks like a valid request, yet it silently corrupts training.
• The rollout temperature is recorded per token (orchestrator/train_sink.py) and reused by the trainer to rescale logits: scaled_logits = logits / temperatures (trainer/rl/train.py). At 0 that divide is x / 0 → inf → selective_log_softmax → NaN loss. Nothing raises; the run just stops learning.
• Fix: temperature: float = Field(1.0, gt=0, le=2.0), matching the guards already on neighbouring fields (e.g. dp: int = Field(1, ge=1)).

Verification

With the new bound, a config with temperature=0 fails Pydantic validation before the run starts. Reproducing the trainer's divide gives all-NaN logprobs at 0 and finite values at any positive temperature.

---

Note

Low Risk
Single Pydantic bound and docstring on training config; prevents a silent training failure mode with no runtime behavior change for valid configs.

Overview
TrainSamplingConfig.temperature validation changes from ge=0 to gt=0, so temperature=0 is rejected at config load time instead of being accepted.

That value is not only a rollout sampling knob: it is stored per token and reused in the trainer as logits / temperature. Zero temperature caused division by zero and non-finite log-probs/loss with no clear error. The field docstring now documents that constraint.

EvalSamplingConfig is unchanged and may still allow 0 for inference-only eval.

^{Reviewed by Cursor Bugbot for commit ecc4d8a. Bugbot is set up for automated code reviews on this repo. Configure here.}