Multi-oracle report: add missing noise-floor/p99 scale block (6ed117f gap)

[lucapinello]

What I found

Commit 6ed117f ("IGV normalization explanation: causal report + LegNet stub parity") fixed the causal report and the LegNet stub but missed the multi-oracle report, which had its own terser intro.

Independent re-audit of the 4-part IGV contract (IGV embedded / ymax=3.0 signal tracks / scale-explanation block / assay:cell_type provenance) across all 17 shipped IGV-containing HTMLs flagged rs12740374_SORT1_multioracle_report.html at 1/5 scale-explanation markers — every other report had 5/5.

What users saw

Before this fix, the multi-oracle IGV block read only:

Signals are floor-rescaled to [0, 3.0] where 1.0 is the genome-wide p99 peak for that assay.

No mention of the p95 noise floor or the "top 1% of bins" framing the other reports ship. Users opening the multi-oracle view in isolation got a less-explicit explanation than in the per-oracle reports.

Fix

• chorus/analysis/multi_oracle_report.py:712-724 — split the one-paragraph intro into two: the window/coverage caveat stays, plus the same "noise floor (p95) and peak threshold (p99)" block the single-oracle + causal reports emit.
• Shipped rs12740374_SORT1_multioracle_report.html — hand-patched to insert the same paragraph at the matching location, preserving the existing IGV tracks. scripts/regenerate_multioracle.py --consolidate from committed JSONs silently drops the IGV section (per-oracle prediction pickles aren't persisted), so a naive regen would make things worse — hand-patch was the right call.

4-part audit result (after fix)

All 17 IGV-containing HTMLs now pass:

Check	Result
(a) IGV embedded	17/17
(b) ymax=3.0 signal / 1.0 summary	17/17 consistent
(c) scale explanation (5/5 markers)	17/17 (was 16/17)
(d) assay:cell_type track provenance	17/17 (bare SPLICE_SITES labels are correctly cell-type-agnostic per AG metadata — investigated and confirmed not drift)

Flagged follow-up (not blocking this fix)

scripts/regenerate_multioracle.py --consolidate from committed JSONs silently drops the IGV section because per-oracle prediction pickles aren't persisted. Either persist the pickles or mark IGV-rebuild as requiring the --oracle <name> runs first. Worth a separate PR; doesn't affect shipped reports today because they're regenerated with the full pipeline.

Test plan

• pytest tests/ --ignore=tests/test_smoke_predict.py -q → 339 passed / 1 skipped
• Selenium re-render: multi-oracle HTML at 1600×4500 — new scale block visible, IGV container present, 0 JS console errors
• 4-part audit across all 17 IGV-containing HTMLs — all pass
• Track provenance: all oracle · assay:cell_type or assay:cell_type patterns confirmed, SPLICE_SITES cell-type-agnostic tracks verified against AG metadata (739 SPLICE_SITES tracks total; 5 have empty cell_type by design: donor/acceptor +/− strand + padding)

🤖 Generated with Claude Code