regen(examples): ChromBPNet expm1 CDFs + three latent staleness fixes

[lucapinello]

Phase 6 regeneration after the ChromBPNet expm1 fix (#113) and the HF background rebuild. Regenerating also surfaced three artefacts that were stale against earlier merged changes and one blocking issue that predates this work. Everything is separated below so nothing gets misattributed to expm1.

1. The intended change — expm1

ChromBPNet raw counts come down by exactly the window softmax mass w; percentile rankings are unchanged:

	ref	alt	log2FC	effect %ile	activity %ile
before	287.857	747.653	+1.37394	0.9995	0.9059
after	287.237	747.014	+1.37581	0.9995	0.9059
Δ	−0.620	−0.639	+0.0019	—	—

klf1 locus profile: max 127.333 → 127.307, mean 0.462 → 0.461. Same signature.

2–4. Latent staleness this surfaced (not caused by expm1)

• SORT1_chrombpnet was stale since 2026-05-09, predating the 2026-06-17 windowing fix (fix(regen): score short-input oracles on their native window in multi-oracle report #94). That is the big jump — ref 48.53 → 287.24, log2FC +0.318 → +1.376. The post-fix(regen): score short-input oracles on their native window in multi-oracle report #94 multioracle artefact already carried +1.37394, and the regenerated value matches it to within the expm1 residual, which is how I attributed it.
• klf1's EPInformer-seq cells were stale against a weights swap. HF commit 2654c05b7c (2026-06-04) replaced per_cell_widewin/main.pt with the "widewin-roadmap retrain (sparse H3K27ac)"; the notebook was committed alongside it but never re-executed. Hence K562 H3K27ac mean 1.771 → 0.132. Verified not device-dependent: cpu vs cuda agree to 0.006 absolute on a 3.78 range (0.16%). The mechanism that let it drift silently is epinformerseq.py:416 — hf_hub_download with no revision.
• The AlphaGenome report was stale against the CAGE minus-strand dedup (regen(multioracle): tile LegNet across the locus for the IGV browser; dedup CAGE #99/regen(multioracle): keep the CAGE minus strand (the article's +1.52) #101) and still carried both strands. all_scores 61 → 32 is entirely CAGE:HepG2 58 → 29 — precisely the duplicate regen(multioracle): keep the CAGE minus strand (the article's +1.52) #101 exists to remove. No track lost: same 4 unique descriptions before and after.

Notebooks

Re-executed on H100 with zero error outputs: klf1 14/14 cells, advanced_multi_oracle_analysis 57/57, comprehensive_oracle_showcase 38/38. The documented chorus kernelspec (examples/notebooks/README.md:45) had to be registered first — it is not registered by chorus setup. Re-execution also removed a foreign-home path leak (/Users/jieconglin) from klf1's outputs.

⚠️ Blocked, and it predates this PR — needs your decision

rs12740374_SORT1_legnet_report.html and rs12740374_SORT1_multioracle_report.html are left at their origin/main bytes on purpose.

#99 (2026-06-18) added "tile LegNet across the locus for the IGV browser", but the legnet artefact was last committed 2026-06-17 (#95, single-window). The tiling has therefore never been applied. Applying it embeds a locus-wide 1 bp array:

consolidation	size	IGV tracks
with legnet .pkl (i.e. #99 as intended)	137 MB / consolidated 145 MB	21
without legnet .pkl	9.55 MB	18 — legnet's 3 dropped
origin/main (committed)	9.93 MB	21

Both regenerated variants are wrong: the first exceeds GitHub's 100 MiB file limit (the push was rejected by the pre-receive hook), the second silently loses tracks. The committed 9.93 MB / 21-track artefact came from a pre-#99 single-window .pkl and is not reproducible with current code. So #99 has been unshippable since it merged.

Options, all needing a call from you: bin/downsample the embedded IGV arrays, revert the tiling to single-window, or move these HTMLs to Git LFS.

Consequence to be aware of while merging: the consolidated example_output.{md,json} in this PR carry the corrected ChromBPNet numbers, but the consolidated HTML does not, until the above is resolved.

Other notes

• legnet and alphagenome were re-run only to regenerate their gitignored *_variant_report.pkl, without which --consolidate silently degrades to "JSON only (no IGV predictions)".
• AlphaGenome is not deterministic on this box: two consecutive runs differ on all 64 raw values (e.g. ref_value 2573.0 → 2568.0), though quantile_score was stable in all 58 cases. Their small numeric deltas here are that noise. Worth knowing for AUDIT_CHECKLIST §13, which asserts identical predictions — that gate passed earlier only because AlphaGenome wasn't installed yet.

Fast suite: 477 passed, 4 skipped, 0 errors.

🤖 Generated with Claude Code