ChromBPNet CDFs rebuilt with expm1 count inversion
Rebuilt against the corrected count-head inversion (np.exp -> np.expm1); see pinellolab/chorus#113.
Scope reproduces the previous file exactly: 786 tracks (22 ATAC + 20 DNASE + 744 CHIP), same track_ids in the same order, fold 0, chrombpnet_nobias, --assay all, n_cdf_points 10000, all sampling seeds unchanged (they are hardcoded in the builder). Built by replicating the original shard layout (one ATAC/DNASE process + CHIP --shard 0/1 --shard-of 2) so the per-model RandomState(999) bin draws land at the same stream positions and expm1 is the only intended difference.
Validation: counts uniformly effect=18672 / summary=34004 / perbin=1088128 across all 786 tracks; all rows monotone non-decreasing; no all-zero rows; effect CDF has zero negatives (as before). The summary/perbin CDFs now contain small negative entries (min -0.88) because a near-dead window gives log(count+1) < 0 hence expm1 < 0 -- left unclamped so the builder and oracle.predict() agree, matching the documented Cherimoya behaviour.
Impact: percentile rankings are preserved (matched-rank percentile delta 0.0000; effect CDF mean shift -9.8e-04). Raw counts are corrected by (c+1)/c, which is small for ATAC/DNASE (median old/new 1.0122, median count 82) but large for the 744 CHIP/BPNet tracks (median old/new 1.3049, median count 3.3) -- those were ~30% too high.
sha256 cf30e17fc3558448be6710968e43c444678ab4e22472fd63e112e4b8df767972
Previous: 526beb2ce8310f6fdb331f766eac55ce3262b67f1a43416532d8bad8f83183eb
DO NOT MERGE until pinellolab/chorus#113 is merged. The bug was self-consistent (oracle and builder made the same +1 error), so publishing these CDFs against exp-based code would inflate CHIP activity percentiles by ~30% -- worse than the status quo.
Merging: expm1 code fix is live on chorus main (#113) so code and CDFs now match.