v17 full audit: MCP server spec drift + Borzoi/LegNet numbers + CEBP walkthrough

[lucapinello]

Summary

A comprehensive ship-ready audit: every live doc, every walkthrough README vs its committed example_output, all 17 shipped HTML reports screenshotted via headless Chrome, the Python API, the chorus-mcp tool registry, error messages, and macOS-arm64 device detection.

Five drifts fixed

1. chorus/mcp/server.py:77 (ORACLE_SPECS) — AlphaGenome "5930 tracks" → 5,731 (matches list_oracles() output and the notebooks after v16).
2. chorus/mcp/server.py:1482 (MCP system prompt) — same "5930" → "5,731". Also added 7,611 tracks explicitly for Borzoi so a Claude session using chorus-mcp gets consistent numbers.
3. chorus/mcp/server.py:71 — LegNet input_size_bp: 230 → 200 (actual create_oracle('legnet').sequence_length == 200).
4. scripts/README.md:37, 94 — Borzoi "7,612" → 7,611 (real count summed across assay types via get_track_info()).
5. examples/walkthroughs/validation/SORT1_rs12740374_with_CEBP/README.md:21 — "CEBPB binding gain: +0.22" → +0.27 (committed example_output.md line 31 shows +0.270; 0.05 gap is beyond CPU non-det).

What else was checked and is clean

• Every other walkthrough number is either an exact match with the companion example_output.md or within ±0.006 (accept as CPU non-det).
• Every formula label: chromatin/TF/histone/TSS → log2FC; gene expression → lnFC; MPRA → Δ (alt−ref). Zero rogue labels.
• 17 HTML reports render with glossary, formula chips, consensus matrix, per-layer tables, IGV placeholder. IGV itself is client-side JS and doesn't render in headless file:// — expected.
• Error paths: create_oracle('fakeOracle') → helpful ValueError with the valid list; predict() without load → clear ModelNotLoadedError.
• MCP tool registry — 22 tools, all naming matches what the walkthroughs reference.
• Platform detection on macOS arm64 — detect_platform() returns key=macos_arm64, has_cuda=False; each oracle env handles its own backend (MPS/TF Metal/JAX-CPU) by design.
• Zero live references to the old examples/applications/ path (only audits/ snapshots still mention it, which is intentional).

Known latent issues, NOT fixed here

• Off-by-one in predict_variant_effect ref-allele check (chorus/core/base.py:322-328) — tracked in audits/2026-04-21_v16_notebook_and_html_audit.md. Needs its own focused code PR with SNV + indel + reverse-strand coverage.

Test plan

• pytest tests/ --ignore=tests/test_smoke_predict.py -q → 333 passed / 1 skipped (9m 9s)
• Screenshotted all 17 HTMLs at 1400×2800 and reviewed each for label/formula/layout drift
• python -c 'from chorus.mcp.server import mcp; import asyncio; print(len(asyncio.run(mcp._list_tools())))' → 22 tools
• python -c 'import chorus; chorus.create_oracle(\"fakeOracle\", use_environment=False)' raises the expected ValueError with the valid list

🤖 Generated with Claude Code