Fix from-scratch install: coolbox PyPI quarantine (P0) + epinformerseq nvidia channel (P1)

[lucapinello]

Summary

A genuine scorched-earth fresh install (every env / weight / background / genome wiped, nothing cached) on macOS arm64 surfaced two install-blocking bugs that the warm CI/pytest path misses. Both are fixed here; verified end-to-end (8/8 oracles healthy, real prediction + variant scan pass, MCP server serves tool calls).

Full write-up: audits/2026-06-16_macos_arm64_fresh_install_audit.md.

Bugs fixed

1. P0 — coolbox quarantined on PyPI breaks the documented install

coolbox>=0.4.0 is quarantined on PyPI (simple index returns pypi:project-status: quarantined with zero files), so mamba env create -f environment.yml fails for every new user. Worse, pip installs the pip: block as one batch, so the failure also drops fastmcp / huggingface_hub / oxbow — the MCP server stack — leaving a confusing error pointed only at coolbox.

Fix: install coolbox from the official upstream repo, SHA-pinned to the 0.4.0 release tag (GangCaoLab/CoolBox@651b930), which is the legit actively-maintained source (257★, last push 2026-06-08), pure-pip / requires-python >=3.7 (cross-platform). Verified on arm64: builds, imports as 0.4.0, and the exact symbols chorus uses (GTF, TabFileReaderInMemory, FMT2COLUMNS) resolve. A comment tells us to revert to coolbox>=0.4.0 once PyPI un-quarantines it.

2. P1 — new epinformerseq oracle env fails to build

chorus setup reports 7/8 oracles ready; epinformerseq dies with Failed to create environment: None. Root cause: environments/chorus-epinformerseq.yml declares the vestigial nvidia conda channel, and — unlike every other oracle — epinformerseq is not registered for platform adaptation, so the channel isn't stripped on non-CUDA hosts. conda.anaconda.org/nvidia SSL-times-out → build aborts. Nothing uses the channel (torch comes via pip: torch>=2.0, which auto-selects CUDA on Linux / MPS on macOS).

Fix: remove the dead nvidia channel. Rebuilt clean (env + weights + background).

Supporting fixes

• manager.py — env-create failures logged process.stderr, which is always None (stderr is merged into the streamed stdout), producing the useless Failed to create environment: None. Now captures the last 25 output lines + exit code. (This is why New version of chorus package #2 was opaque.)
• CLI help — --include-alternative-backends help text claimed alphagenome_pt is opt-in; it actually installs by default (_SKIP_FROM_DEFAULT_SETUP is empty). Corrected to describe the flag as a no-op. No behavior change.
• README — added a note to restart Claude Code/Desktop after upgrading (a running MCP server serves stale in-memory code; it listed only 6 oracles until restart).

Verification

• mamba env create -f environment.yml builds clean (throwaway env, coolbox + MCP deps import) ✅
• chorus health → 8/8 oracles healthy ✅
• Enformer prediction + variant scan at the β-globin locus → PASS ✅
• MCP server connected and list_oracles() returns data ✅

Open item (NOT in this PR — needs a maintainer decision)

cwd-relative data paths: genomes/ and downloads/ resolve relative to the working directory, so running setup from two clones silently duplicated ~31 GB and chorus cleanup would orphan one copy. This is a behavioral/migration change, so it's documented in the audit as a proposal rather than changed here.

🤖 Generated with Claude Code