v21 fresh-install audit: data caches purged + re-downloaded, no findings

[lucapinello]

Summary

Walked audits/AUDIT_CHECKLIST.md top-to-bottom with every re-downloadable cache purged before starting. No new findings — every mechanisable checklist item passes on a fully fresh-data run.

What was nuked + re-downloaded

Cache	Before	Re-downloaded	Time
~/.chorus/backgrounds/ (6 NPZ CDFs)	1.5 GB	HuggingFace dataset	44 s total
genomes/hg38.fa + .fai	3.0 GB	chorus genome download hg38 → UCSC + decompress + samtools faidx	~9 min

Results — all pass

§	Result
1 CLI	chorus --help surfaces all 6 subcommands
3 GPU	all 6 envs detect Metal/MPS/JAX-METAL on macOS arm64
4 CDFs (fresh from HF)	6/6 monotonic, p50≤p95≤p99, signed% correct (0/20/0/100/100/13)
5 Python API	sequence_length matches spec for all 6; errors clear
6 Notebook fresh exec	single_oracle_quickstart.ipynb: 49 cells, 0 errors, 0 warnings
7 HTML (selenium)	18/18 render with fresh Chrome profile, 0 JS errors each
10 Consistency	zero drift — grep '5,930|7,612|196 kbp|examples/applications/' empty
11 pytest	335 passed / 1 skipped (8m 28s)
14.4 chrom validation	regression test holds
15 Offline	0 runtime CDN fetches across all 18 HTMLs
16 Logging hygiene	no committed tokens
18 License	LICENSE + docs/THIRD_PARTY.md + bundled IGV.js header intact

Scope deliberately deferred to release-host audit

• §1 full conda env recreate — 80 GB / 2–4 h, destructive to ongoing work
• §2 HF-gate missing-token E2E — would break other AlphaGenome work
• §6 multi-oracle + advanced notebooks — need all 6 oracles loaded + GPU
• §8 MCP E2E over stdio — ~4 min AlphaGenome predict
• §13 real-oracle determinism — ~30 min across 6 loaded oracles
• §14 remaining edge cases (indels, chrM, telomere) — need loaded oracles; §14.4 already fixed + regression-tested
• §17 pip-audit — already advisory in CI per v20

Artefacts in audits/2026-04-21_v21_fresh_install/

• report.md — summary + verbatim timings
• screenshots/*.png (16 files; 18 reports with 2 basename collisions)
• logs/00-10_*.txt + 09_quickstart_fresh.ipynb — pre-nuke state, post-nuke empty confirmation, CLI help, genome download trace, CDF fresh pull, device probe, API sanity, consistency greps, selenium output, fresh notebook, pytest log

Headline

Previous audits (v15–v20) + fixes in PRs #32, #34, #36 have driven this repo to a ship-clean state on every mechanisable item. What remains open is release-host work (full env build, multi-oracle notebooks, real-oracle determinism, MCP E2E) — items explicitly scoped for that environment.

Test plan

• ~/.chorus/backgrounds/ + genomes/hg38.* deleted, confirmed empty before any check runs
• chorus genome download hg38 downloaded, decompressed, indexed in ~9 min (see logs/03_genome_download.txt)
• 6-oracle CDF fresh pull + sanity: 44 s total, all 6 pass monotonicity + ordering (see logs/04_cdf_fresh_download.txt)
• Fresh jupyter nbconvert --execute on single_oracle_quickstart.ipynb: 0 errors, 0 warnings
• Selenium-rendered all 18 HTMLs with --user-data-dir=<fresh tmpdir>: 0 JS errors per report
• pytest fast suite: 335 passed / 1 skipped

🤖 Generated with Claude Code

[lucapinello]

Superseded — artefacts cherry-picked into chorus-applications via 5300daa. The other agent's reconciliation commit imported the 18 selenium screenshots + 10 probe logs + fresh notebook + report.md from this branch, while skipping the 2 file reverts (LegNet 230→200 fix in comprehensive_oracle_showcase.ipynb, test_integration.py chrombpnet skip guard) that this branch would have undone.

Full reconciliation note lives at audits/2026-04-21_v21_fresh_install/report.md. Both v21 audits (this macOS arm64 data-cache-purge run + the Linux/CUDA fresh-env run at audits/2026-04-21_v21_fresh_install_linux_cuda.md) are preserved as complementary evidence.

Closing to keep the open-PR list clean. No content lost.