Fix undefined mask variable in TransformerTemporalModel.forward()

[Mr-Neutr0n]

Summary

• Fix a NameError caused by referencing an undefined mask variable in musev/models/temporal_transformer.py at line 301.
• When vision_conditon_frames_sample_index is None (or keep_content_condition is False) and need_temporal_weight is False, the code falls into the else branch where mask was never defined — it is only created inside the if vision_conditon_frames_sample_index is not None and self.keep_content_condition block.
• Since there is no vision conditioning in this branch, all frames are treated equally and no masking is needed. The fix changes residual + mask * hidden_states to residual + hidden_states.

Details

The four branches of the output computation should be:

vision_conditon_frames_sample_index	need_temporal_weight	Output
not None	True	residual + abs(temporal_weight) * mask * hidden_states
not None	False	residual + mask * hidden_states
None	True	residual + abs(temporal_weight) * hidden_states
None	False	residual + hidden_states (was: residual + mask * hidden_states — bug)

The first three branches are correct. The fourth incorrectly referenced mask which does not exist in that scope, causing a NameError at runtime.

Test plan

• Verify the fix by reading the surrounding control flow to confirm mask is only assigned in the if branch
• Run inference with need_temporal_weight=False and no vision conditioning frames to confirm no NameError