fix(chrombpnet): drop the 33 mouse (mm10) models — chorus is hg38-only

[lucapinello]

The defect

ENCODE publishes a mouse developmental atlas of ChromBPNet models (embryonic forebrain / midbrain / hindbrain / limb / liver / heart / neural tube / facial prominence, stages E11.5–E14.5). 33 of the registry's 42 ATAC/DNASE entries were those mouse models, and every code path around them assumes hg38:

• scripts/build_backgrounds_chrombpnet.py opens genomes/hg38.fa and draws its DHS-anchored positions from the hg38 DHS vocabulary. So the per-track CDFs shipped for those 33 rows were built by pushing human sequence through mouse models.
• The registry recorded no organism, so there was no field to filter or assert on. That is what let the mismatch ship.

Why 9 is the right surviving set

The 9 human models kept here are exactly the set scripts/generate_catv1_defaults.py already listed as CHROMBPNET_HUMAN_ANNOTATIONS, with the comment "CATv1 is GRCh38-only, so those have no counterpart and are omitted." The filter already existed in one place and had simply never propagated to the registry or the builder — so this is an independent corroboration, not a fresh judgement call.

The background

chrombpnet_pertrack.npz goes 786 → 753 rows (9 human ATAC/DNASE + 744 human CHIP) by pure row subsetting — no model was run and no CDF value recomputed.

Validated (validate_753.py, all checks passed):

check	result
rows	753 = 744 CHIP + 9 ATAC/DNASE
the 9 are exactly the human allow-list	pass
rows removed	exactly 33, zero of them CHIP
9 surviving accessibility rows bit-identical	`max
sample counts unchanged	18,672 / 34,004 / 1,088,128
CHIP counts	37,344 / 68,008 / 2,176,256
CDF health	all finite, all monotone, no all-zero rows

The name-collision trap

The mouse tissue names collide with human ENCODE CHIP biosamples — liver, heart, brain and forebrain are all human CHIP contexts in chrombpnet_JASPAR_metadata.tsv. A name-only filter would silently delete human CHIP rows.

So the removal keys on (assay, cell_type), never on the bare tissue name; the validator asserts no CHIP row is dropped, and 16 human CHIP rows with colliding names were correctly kept. The registry header now carries this warning for the next person.

Scope: what is not changed

Enformer (1,643 mouse tracks) and Borzoi (2,608) stay unexposed. Both are human-only today only because a *_human_targets.txt was chosen — nothing asserts that choice against the FASTA the builders open, and nothing would catch a future *_mouse_targets.txt. Also unfixed here: AlphaGenomeOracle(organism="mouse") is accepted, stored, and never read. Those get a per-row genome field and a hard assert in follow-up work.

Re-adding mouse properly needs an mm10 FASTA and an mm10 region set for background construction (the cCRE Registry has an mm10 build; the DHS vocabulary does not), so it is tracked separately rather than as a registry edit.

Blast radius

No consumer breaks — chrombpnet.py, discovery.py, mcp/server.py, _setup_prefetch.py and the builder all read the registry dynamically, and no test pins its contents. Setup now prefetches 9 model tars instead of 42.

Two stale comments updated (discovery.py's alias-dedup note, generate_catv1_defaults.py's "also ships mouse models") and one misleading test docstring.

507 passed, 4 skipped, 0 failed

🤖 Generated with Claude Code