fix(health): recognise slim HF mirror for chrombpnet + add alphagenome_pt probe

[lucapinello]

fix(health): recognise slim HF mirror for chrombpnet + add alphagenome_pt probe

Two stale checks in chorus/core/weights_probe.py were causing chorus health to falsely report healthy 0.4.0 installs as ⚠ Not installed.

Symptom

On a fresh chorus setup --oracle all against 0.4.0:

✓ alphagenome: Healthy
⚠ alphagenome_pt: Not installed — run `chorus setup alphagenome_pt`
✓ borzoi: Healthy
⚠ chrombpnet: Not installed — run `chorus setup chrombpnet`
✓ enformer: Healthy
✓ legnet: Healthy
✓ sei: Healthy

…even though both flagged oracles load and predict cleanly (oracle.load_pretrained_model() succeeds in <8 s, end-to-end oracle._predict() returns the expected outputs). I.e. the runtime path is fine; only the probe is stale.

Root cause

1. chrombpnet probe (_probe_chrombpnet) was looking at the legacy ENCODE-tarball-extracted directory:

   default = CHORUS_DOWNLOADS_DIR / "chrombpnet" / "DNASE_K562"
   if not default.exists() or not any(default.iterdir()):
       return (False, [str(default)])

   In 0.3.0+ that directory is empty by default — chrombpnet weights stream from the slim HF mirror at lucapinello/chorus-chrombpnet-slim (ChromBPNet HF slim mirror + chrombpnet_nobias default (0.3.0) #59 / fix(chrombpnet): F1 — huggingface_hub missing in env yaml + v30 scorched-earth audit #60). The legacy directory only gets populated if a user explicitly requests model_type='chrombpnet' (bias-aware) or fold ≠ 0, both of which fall through to ENCODE.

2. alphagenome_pt was missing from _ARTIFACT_PROBES entirely. The oracle was added to the registry in AlphaGenome PyTorch backend (opt-in spike) #62 but the health probe was never updated. So the probe-dispatcher never had an entry for it; it always fell through to the missing-marker code path.

Fix

1. _probe_chrombpnet: accept either the slim-mirror snapshot cache (manifest.json present under models--lucapinello--chorus-chrombpnet-slim/snapshots/<rev>/) or the legacy ENCODE directory.
2. New _probe_alphagenome_pt: checks the HF cache at models--gtca--alphagenome_pytorch/snapshots/<rev>/*.safetensors (the upstream port's published weights).
3. New _hf_cache_dir() helper: both new probes route through it so they honour HF_HOME / HF_HUB_CACHE env vars via huggingface_hub.constants.HF_HUB_CACHE, with a defensive fallback to ~/.cache/huggingface/hub.

Verification

End-to-end on macOS arm64:

$ chorus health
✓ alphagenome: Healthy
✓ alphagenome_pt: Healthy
✓ borzoi: Healthy
✓ chrombpnet: Healthy
✓ enformer: Healthy
✓ legnet: Healthy
✓ sei: Healthy

7/7 ✓, was 5/7 before.

pytest -m "not integration and not slow": 368 passed, 1 skipped, 5 deselected (no regression).

Out of scope

• A regression test for the probe logic. Would need either an integration-marked test that actually invokes chorus health after setup, or a unit test that mocks the HF cache path. Tracked as a follow-up; the fix itself is small enough to inspect by reading.

🤖 Generated with Claude Code