Release 0.4.0

[lucapinello]

Headline changes since 0.3.0

ChromBPNet CDFs rebuilt against chrombpnet_nobias

The 0.2.x CDFs on lucapinello/chorus-backgrounds were built against the bias-aware chrombpnet variant. After the 0.3.0 default flipped to chrombpnet_nobias, percentile lookups were doing chrombpnet_nobias predictions → chrombpnet empirical CDFs — biased.

Rebuilt on A100 in ~10h. New NPZ:

• 786 tracks (22 ATAC + 20 DNASE + 744 CHIP)
• All CDFs monotone, every reservoir filled (effect_count=9609 per track)
• sha256: be61e9e8f9b919b43c599b7fbc9deb74f8f1e6dc1da5e2cdb92036a85bf13205
• ATAC/DNase p95 shifts 13.5–29.3% (bias correction)
• CHIP/BPNet p95 ~unchanged (already nobias-equivalent)

Audit: audits/2026-04-29_chrombpnet_cdf_rebuild/report.md.

HuggingFace mirror consolidation

Enformer / Borzoi / Sei / LegNet weights now fetch from chorus-controlled HF repos with original-source fallback. Closes the TFHub deprecation hazard. Audit: audits/2026-04-29_hf_mirror_consolidation/report.md.

AlphaGenome PyTorch backend

Second AlphaGenome oracle (alphagenome_pt) shipped alongside the JAX default. Same model + same weights converted to safetensors. recommend_alphagenome_backend() helper picks the right backend per platform + window size. Both install by default. Audit trail: audits/2026-04-29_alphagenome_pytorch_spike/, audits/2026-04-29_alphagenome_pt_stress_test/.

Env-setup hardening

• Timeout-soft policy in _setup_environment: cold-NFS validation timeouts don't silently downgrade use_environment=True.
• _check_env_ready() wired into every oracle's load_pretrained_model so missing-dep failures raise EnvironmentNotReadyError up-front.
• Closes tests/test_alphagenome_backends_equivalence.py: rewrite to compare via chorus._predict (raw-head shape assertion always fails) #63, chorus.core.base: env-validation timeout silently flips use_environment=True → False, then crashes on import in wrong env #64, tests/test_alphagenome_backends_equivalence.py: subprocess.run(["mamba", ...]) fails inside "mamba run -n chorus pytest" #65.

HF-mirror direct-load fix (post-merge, PR #69)

Wired chorus mirror routing into Enformer + Borzoi _load_direct paths. Caught during the scorched-earth audit at audits/2026-04-30_overnight_full_audit/report.md.

Validation summary (from overnight scorched-earth audit)

• Fresh install from zero: ✅ all 7 oracle envs rebuild via chorus setup
• 7/7 oracles chorus health Healthy
• 7/7 oracles load + predict + return finite values
• All 3 shipped notebooks execute end-to-end (49 + 127 + 59 cells, 0 errors)
• analyze_variant_multilayer end-to-end on SORT1 rs12740374 with the new CDF: PASS
• Fast suite: 368 passed

Test plan

• Fast suite local: 368 passed
• Fresh install end-to-end: all phases green (audits/2026-04-30_overnight_full_audit/report.md)
• New CDF sha256 verified against locally-built file
• Variant analysis with new CDF: HepG2 DNase at SORT1 rs12740374 returns finite report
• Linux CI fast suite

Open follow-up issues (not blocking)

• build_backgrounds_chrombpnet: --part variants/--assay X overwrites the existing interim NPZ instead of appending #73 — build_backgrounds_chrombpnet interim NPZ overwrite (caught during the rebuild)
• build_backgrounds_chrombpnet: --gpu N overrides outer CUDA_VISIBLE_DEVICES, breaking parallel-launch pattern #74 — --gpu N overrides outer CUDA_VISIBLE_DEVICES

🤖 Generated with Claude Code