[jmschrei]

This PR shortens the Unreleased section of docs/whats_new.rst ahead of the release. The section goes from 70 KB and 110 bullets to 22 KB and 57 bullets. Nothing outside the Unreleased section changes.

What changed

• Each entry now says what changed and, in a clause, the failure it fixes. It keeps at most one headline number. Benchmark configurations, per-model breakdowns and repeated measurements are removed.
• Related entries are merged:
  • io: from 16 bullets to 7;
  • seqlet: from 9 to 5;
  • the Claude Code skill: from 19 to 4, because most of its bullets repeated the library changes as edits to reference files;
  • Testing: from 6 to 1.
• Order: saliency and dependency_map move up next to Packed sequences, since they are the new features of the release. Claude Code skill moves from the top to near the end.

Credits

Every thanks, issue link and citation in the old section is in the new one, on the entry it credits. A script compared the two sections to check this.

Credit	Entry
@cyborgdnx	saliency, dependency_map
@avantikalal	deep_lift_shap trailing output axes, pisa reading n_outputs
@emattei, and jmschrei/cherimoya#103	deep_lift_shap drawing a batch's references in one call
@Al-Murphy	dinucleotide_shuffle(allow_N=True)
#107	the FASTA memory map in extract_loci
#111	one_hot_encode and non-ASCII characters; two entries that each linked it are merged

The citations of Nair et al. (fastISM), Majdandzic et al. (gradient correction) and Tomaz da Silva et al. (dependency maps) are kept, as is the Nature Methods paper in the README entry.

Checks

• sphinx -W passes. The rendered section has the same 22 headings.
• prepush: credits, changelog heading and commit wording pass.

The six surfaces

1. Tests: not affected, because no code changes.
2. Docs: only the changelog changes.
3. README: not touched.
4. Changelog: this is the change.
5. Bundled skill: not affected.
6. Compatibility: no code changes.

🤖 Generated with Claude Code