Fix the track_indices name desync, and restore predict-and-plot
Four fixes on top of the pipeline, all tested against the checkpoint on CPU.
1. track_indices desynced the channel names, silently
track_indices narrowed the tensor but not track_ids, so out.track(id) indexed the wrong column. On chr19:6,700,000-6,716,384 asking for ENCSR056HPM and ENCSR000AKO:
sub.tracks.shape -> (6144, 2)
len(sub.track_ids) -> 7362 # still the full panel
sub.track("kai1") -> mean 0.00568903
real kai1 -> mean 0.0778583 not this
real ENCSR056HPM -> mean 0.00568902 this
The call succeeded and returned a different track than the name asked for, 14x off. A name whose index sat past the narrowed width raised IndexError instead. Since track_indices is how you avoid an 11.6 GB forward at 1 Mb, this hit exactly the people using it as intended.
Fixed by resolving the channel set in one place, _select_channels, called from preprocess. The names and the indices are produced together, so they cannot drift. _forward no longer takes them as a separate parameter.
2. Selecting tracks by id, and plot=True was silently ignored
plot and tracks_to_plot were swallowed by _sanitize_parameters: the call returned normally and drew nothing. There was no plotting in the file. This restores the one-call predict-and-plot from the 2025 Space, which is the interface people are used to:
out = predictor({"chrom": "chr19", "start": 6_700_000, "end": 6_732_768, "species": "human"},
plot=True, tracks_to_plot={"K562 RNA-seq": "ENCSR056HPM"})
Adds track_ids=["ENCSR..."] to select by id rather than by integer, and tracks_to_plot implies it, so plotting two tracks computes two channels instead of 7,362. Verified bit-comparable to slicing the full panel, max diff 1.2e-07. matplotlib is imported lazily inside the plot helper, so the pipeline still has no hard dependency on it. plot=True without tracks_to_plot now raises rather than doing nothing.
3. Three assemblies bypassed the "pass 'seq' directly" message
That message only fired on a 200 carrying no dna. An assembly UCSC does not host answers 400, or with something that is not JSON, so pig, wheat and cotton raised a bare HTTPError or JSONDecodeError. The fetch now catches those and explains.
Pig also just works now: UCSC answers to Sscrofa11.1, while track_metadata says Sscrofa11. That is the only assembly name needing a translation, so it is a one-entry UCSC_ASSEMBLY_ALIASES rather than a table. I checked all 15 against the live API: 13 resolve under the config's own name, and only IWGSC and Gossypium_hirsutum_v2.1 are genuinely not hosted, which is what the improved message is for.
4. skip_lm_head=True
The pipeline never returns the language-model logits, and at 1 Mb that head is not free.
Tested
Full panel, track_ids, track_indices, plotting, plot without tracks, and pig, wheat, cotton and arabidopsis by coordinates. Values match the plain AutoModel path. pyflakes clean.
Nothing else in the repo is touched, and config.json is unchanged.