[Afloat16]

Problem

Both embedder collators can dynamically pad each sub-batch to a different sequence length. M3's list-input encoding path directly concatenates the ColBERT vectors along the batch dimension, so unified fine-tuning raises a tensor-size mismatch when those sequence lengths differ. This prevents the documented sub-batch option from working on ordinary variable-length training text.

Changes

Pad the ColBERT vectors to the longest sub-batch sequence before concatenation, using the tokenizer's padding side and zero vectors for padding. Build the corresponding list-input attention mask in the same order: remove each sub-batch's first position, as ColBERT encoding already does, then pad. A leading column preserves the existing full-mask scoring interface.

Dense-only encoding and the tensor-input sub-batch path are unchanged.

Validation

Added 20 tests with an actual local BertTokenizerFast, tiny BERT backbone, sparse/ColBERT heads, and both native collators:

• Dynamic sub-batch embeddings match the existing full-batch path with right padding.
• Supervised, teacher-distillation, and self-distillation training match full-batch losses and all parameter gradients.
• Left/right padding preserves each original vector and aligns every mask value with it, with zero vectors and zero masks in added positions.
• Dense-only and fixed-length tensor sub-batches retain their previous behavior.

The unchanged upstream base produces 14 failed, 6 passed. With the fix, the new tests and existing Trainer/import checks produce 28 passed, 1 skipped. Ruff on the new test module and git diff --check pass.

The tests preserve the existing first-position removal convention and do not claim to redefine left-padded special-token handling. This change is separate from passage MaxSim masking in #1610; it does not change that scoring reduction.