Fix variant-effect/fine-mapping reproducibility bugs (T1/T2/T3) + reproduction audit

[lucapinello]

Summary

Independent reproduction of the Chorus blog post (Analysis A: SORT1/rs12740374; Analysis B: rs9504151/CDYL) surfaced three correctness bugs in the variant-effect / fine-mapping path that broke "reproduce-from-prompt." This PR ships the validated, contained fixes plus the reproduction artifacts. Full evidence: audits/2026-06-16_blogpost_reproduction_report.md.

A separate, larger windowing bug (it requires rebuilding the per-track background CDFs) is documented but not fixed here — see audits/2026-06-17_windowing_normalization_P0.md.

Fixes

T2 — LD-proxy alleles were genome-mismatched (fine_map_causal_variant).
_extract_allele_pairs assigned each proxy (ref=SENTINEL_allele, alt=PROXY_allele) from Correlated_Alleles, so a proxy's "ref" was the sentinel's base — almost never the genome base at the proxy position (≈50 "Provided reference allele … does not match the genome" warnings for one locus; chorus then substituted the wrong ref → effects computed on wrong alleles → ranking meaningless).

• chorus/utils/ld.py: each proxy now uses its own alleles (the Alleles column).
• chorus/analysis/causal.py: prioritize_causal_variants orients ref→genome per variant before scoring.
• Verified: re-run gives 0 mismatch warnings (was ~50); rs9504151 ranks borzoi_v1 with basic prediction and examples #1 (ChromBPNet IMR-90, effect −0.985).

T1 (fine-map path) — 1 bp scoring region collapsed fixed-input-oracle effects.
prioritize_causal_variants scored each variant on a 1 bp region; fixed-input oracles (e.g. ChromBPNet) map that into an N-padded output window, collapsing the effect ~4× (all proxies' |effect| ≤ 0.15). Now scores on the oracle's full input window centered on the variant (guarded for non-integer sequence_length). Verified: ChromBPNet effects restored (rs9504151 −0.985, matching a direct measurement).

Note: the same 1 bp pattern exists in _auto_region (4 MCP tools), discovery.py, batch_scoring.py, and build_backgrounds.py. Because the background CDFs were built with the collapsed window, fixing those requires a coordinated CDF rebuild — tracked as a P0 follow-up (audits/2026-06-17_windowing_normalization_P0.md), not in this PR.

T3 — AlphaGenome all-tracks scan OOM-killed the MCP server.
analyze_variant_multilayer / discover_variant with empty assay_ids on AlphaGenome expands to all 5,731 tracks; at 1 Mb the returned arrays (tens of GB) OOM-killed the whole server. chorus/oracles/alphagenome.py _predict now estimates the allocation and raises a clear ValueError before allocating (cap = 40 % of RAM, override CHORUS_AG_MAX_PREDICT_GB). Verified: blocks the 5,731-track@1Mb case, allows real runs (545- and 1,839-track scans pass).

Tests

• tests/test_analysis.py: the old Correlated_Alleles "fanout" test asserted the buggy ref=SENTINEL contract; updated to assert the corrected own-alleles behavior. (Only one test depended on it.)
• Full suite green after the fixes.

Docs / artifacts

• audits/2026-06-16_blogpost_reproduction_report.md — full per-claim reproduction.
• audits/2026-06-16_reproduction_notes.md — raw run numbers.
• audits/2026-06-17_windowing_normalization_P0.md — the deferred windowing+CDF-rebuild bug.
• audits/2026-006_chorus_blogpost.revised.md — reviewer-corrected article (number fixes + reproducibility recipes).

🤖 Generated with Claude Code