Encode the residue-only part of a batch in one pass - #124

Stacked on #120. One commit.

Encoding is a quarter of the prediction pipeline and every peptide walked through it alone,
filling four small arrays a value at a time. The part that depends only on residue identity
does not need that.

What it does

• the composition matrix and the one-hot residues become a table gather and a scatter over
  the batch;
• the positional block is the same gather at different rows;
• the global vector is a sum along the batch's position axis.

Bucketed encoding of 20,000 peptides: 0.1837 to 0.0252 ms per peptide, 7.3x, and
core.predict on CPU from 2,979 to 3,734 peptidoforms/s. End to end the encoder is a quarter
of the pipeline, so ~1.25x now — and roughly 1.9x once a cheaper trunk makes encoding the
dominant term again.

What it does not do

encode_peptidoform stays the definition of a feature. Modifications are not
reimplemented: they sit on a quarter of a typical peptide list and their placement carries a
legacy quirk a shipped model was trained against, so _apply_modifications and its terminal
counterpart are called on views into the batch arrays. The exactness comes from reusing
that code rather than from matching it.

Two cases fall back to the per-peptide encoder. Both are divergences I found by testing for
them, not hypotheticals:

• Peptides under four residues. The reference's negative positional indices wrap around
  the sequence — with one residue, seq[seq_len - 2] is seq[-1], that same residue — and
  the batched form had skipped them.
• Peptides carrying a residue the one-hot block has no slot for, such as selenocysteine,
  whose atoms the reference still reads from pyteomics.

Feature layouts the route does not cover at all are declined by supports(): the rolling-sum
matrix the four-branch model reads, and the collision cross section extras. Those keep the
per-peptide path, as does vectorised_encoding=False on the dataset, which forces it for
everything. The worst case is therefore no speedup rather than different numbers.

How it is checked

tests/test_batch_features.py compares the two routes value for value and dtype for dtype
over random batches at four windows, truncation, both termini, two modifications on one
peptide, either feature switch, the short and unusual residues above, and through the dataset
in both modes. It builds its own peptides, so it needs no data files and CI runs it.

Separately verified against the held-out corpus: 20,000 real peptidoforms at windows 20, 30
and 60, byte-identical including dtypes, and identical tensors across every bucket of a real
bucketed prediction.

A crash it surfaced

In _apply_composition_to_matrices the outer try guards KeyError and IndexError, but
the inner branch — the one that strips isotope brackets, so C[13] and N[15] as TMT and
SILAC labels carry — guarded only KeyError. A labelled modification beyond the padding
window therefore raised

IndexError: index 22 is out of bounds for axis 0 with size 20

where the unlabelled path warned and carried on. Reachable today by passing a padding_length
below your longest peptide. Fixed here with the same guard the sibling branch has, and covered
by a test.

225 tests pass, ruff clean.

🤖 Generated with Claude Code