fix(rollout): guard zero-length response against x[-0:] returning the whole sequence

[hobostay]

Problem

Python slicing makes seq[-0:] return the entire sequence, not an empty one. Two rollout post-processing paths slice with [-response_length:] and break when response_length == 0:

1. slime/rollout/sft_rollout.py — for a fully-masked SFT sample (no trainable tokens, e.g. no assistant turn or a chat-template mismatch), get_response_lengths returns 0 and loss_mask[-0:] assigns the whole prompt-length mask to sample.loss_mask while sample.response_length == 0. This violates the downstream invariant len(loss_mask) == response_length, asserted in slime/ray/rollout.py and slime/utils/types.py, so one bad sample crashes the entire rollout. On any path where the assertion is not reached, the prompt tokens would silently be treated as trainable.

2. slime/rollout/on_policy_distillation.py — for an empty generation (e.g. immediate EOS), t_log_prob[-0:] keeps the full prompt+response-length teacher log-prob tensor instead of an empty one, misaligning the OPD KL computation with response_length.

Minimal reproduction:

loss_mask = [0] * 50
response_length = 0
len(loss_mask[-response_length:])  # 50, expected 0

Fix

Guard the zero-length case explicitly in both places (seq[-n:] if n > 0 else empty). Behavior for response_length > 0 is unchanged.

Tests

Added tests/test_zero_length_response.py (CPU, @pytest.mark.unit):

• SFT rollout with a fully-masked sample → loss_mask == [], invariant holds (fails without the fix).
• SFT rollout with a normal sample → tail mask unchanged.
• OPD with response_length == 0 → empty teacher_log_probs (fails without the fix).
• OPD with a normal response → tail log-probs unchanged.

All 4 tests pass locally; the two zero-length tests fail on current main, confirming they pin the bug.