v22 fresh audit post-3735ea5: setup prefetch + health + tokens all pass

[lucapinello]

Summary

Independent fresh-state re-audit after 3735ea5 ("Setup prefetch + health classification + token flow") landed on main. Every re-downloadable cache was purged before starting. No new findings — every exercised item passes.

New features verified end-to-end from a purged-cache host

chorus health on empty state — 5.5 s for 6 oracles (was 720 s pre-3735ea5)

Checking alphagenome ...  1.5 s → Not installed — run `chorus setup alphagenome`
                           (marker + HF auth hints inline)
Checking sei ...           1.0 s → Not installed          (was the 120 s hang offender)
TOTAL 6 oracles: 5.5 s

Token-halt path — bare chorus setup with no HF_TOKEN + non-TTY stdin

Exits with code 1 in <1 s. Message names all 3 alternatives (--hf-token, HF_TOKEN env, huggingface-cli login) + "Nothing was downloaded". No partial state risk.

Partial setup — --no-weights --no-genome

Pulls only the background NPZ (548 MB in ~10 s). Setup marker correctly not written (consistent with --help docstring). Subsequent chorus health keeps reporting "Not installed" — contract honoured.

README ## Tokens section clarity

Two-row table: HF_TOKEN + LDLINK_TOKEN. Columns cover "when you need it" + "how chorus setup handles it" (4-step resolution chain for HF). Halt semantics explicit. Registration URLs + AlphaGenome license page. Cross-referenced from the backgrounds section: "The backgrounds dataset is public — no HuggingFace token required" short-circuits a common misunderstanding.

Checklist re-audit — every exercised section PASS

§	Result
1 CLI + setup + health	PASS — new fast path verified
2 HF token flow	PASS — halt + 4-step resolution + 4 code paths consistent (alphagenome.py, load_template.py, README ×3, cli/_tokens.py)
4 CDF fresh pull (6 oracles)	PASS — 29.6 s total from empty cache, all monotonic, p50≤p95≤p99, signed% (0/20/0/100/100/13%)
7 HTML reports	PASS — 18/18 selenium render, 0 JS errors each
7 4-part IGV contract	PASS — 17/17 non-batch pass all 4: IGV + ymax 3.0/1.0 + 5/5 scale markers + provenance
10 Drift grep	PASS — 5,930/7,612/196 kbp/LegNet 230 bp/applications/ all empty
11 Fast test suite	PASS — 338 passed / 2 skipped (63 s; skips are integration tests correctly guarding on missing .chorus_setup_v1)
15 Offline	PASS — 0 runtime CDN fetches
16 Logging hygiene	PASS — no committed hf_… or AKIA…
18 License	PASS — LICENSE + docs/THIRD_PARTY.md

Scope deferred (as always)

• §1 full conda env recreate (80 GB / 2–4 h; destructive to ongoing work — the other agent verified separately on Linux/CUDA)
• §6 multi-oracle / advanced notebooks
• §8 MCP E2E over stdio
• §13 real-oracle determinism

Headline

The two biggest user-facing wins from 3735ea5 are both verified on a fully-purged-cache host:

• 5 s health for 6 oracles instead of 120 s per-oracle hang
• Fail-fast token halt — no mid-download 80 GB surprises

Artefacts in audits/2026-04-23_v22_fresh_audit/ (12 MB)

• report.md — this summary + timings
• screenshots/*.png (16) — selenium-rendered HTML reports at 1600×4500
• logs/00-11_*.txt + fresh notebook — pre/post-nuke state, CLI help, token-halt, partial-setup, genome download, CDF pull, device probe, selenium, pytest, consistency greps

Test plan

• ~/.chorus/backgrounds/ + genomes/hg38.* deleted, confirmed empty before any probe ran
• chorus health from empty state: 5.5 s / 6 oracles, all "Not installed" with actionable messages
• env -u HF_TOKEN bash -c 'echo | chorus setup' → exit 1 in <1 s with all 3 hints
• chorus setup --oracle enformer --no-weights --no-genome → pulls background only, no marker written
• chorus genome download hg38 → 3.1 GB fresh, indexed
• CDF auto-download (5 oracles): 29.6 s total from empty cache, all pass sanity
• Selenium: 18 HTMLs, 0 JS errors per report, 4-part IGV contract holds on all 17 non-batch
• pytest tests/ --ignore=tests/test_smoke_predict.py -q → 338 passed / 2 skipped

🤖 Generated with Claude Code