Tolerate flip-twin collapse in .ldFromSketch instead of erroring

[Yining97]

Problem

On an LD panel that stores a variant in both ref/alt orientations (e.g. ADSP R5: 6,365 such records on chr19; R4 has none), RAISS imputes both orientations — each is a distinct panel variant_id absent from the observed GWAS — so a region's variant set ends up carrying a variant and its own ref/alt-swapped twin. .ldFromSketch calls matchVariants(allowFlip = TRUE), which treats a variant and its swap as the same variant and merges the pair to one match. The old code then computed nMissing = length(requested) - length(matched) > 0 and stopped with "N variant id(s) not present in the LD sketch panel" — but nothing is absent; the pair merged. Two real regions died this way (APOE in the kriging prefilter; chr11 in the fine-mapping LD build), making R5 unusable end to end.

Fix

.ldFromSketch now distinguishes a collapsed flip-twin (a requested id whose (chrom, pos) IS on the panel but merged onto its sibling) from a genuine absence (position not on the panel at all). Only genuine absences count toward the onMissing = "error" stop; the LD matrix is built on the deduplicated set and the survivors are exposed via keptVariantIds (now always, not just under "drop"). The callers that use the matrix alongside z/N realign to the survivors: a new .fmAlignToKeptLd helper at both fine-mapping sites, and a df row-drop at the kriging prefilter and zMismatch QC. No panel/data change — this makes pecotmr tolerant of a panel that carries both orientations.

Preserved: the genuine-absence safety check still errors, and distinct multi-allelic ALT alleles at one position are never merged (verified) so they still enter the credible set. The dropped twin is the same variant (its z is the sign-flip of the kept one), so results are unchanged.

Scope: covers the GWAS RSS path (fine-mapping both methods, kriging, zMismatch). jointEngine/causalInference/twasWeights are out of that path with different vector shapes; their z/N realignment is deferred (the central fix already removes the misleading error there).

Validation

• flip-twin request under onMissing="error" → no stop, deduplicated LD matrix with keptVariantIds; genuine absence still errors; distinct ALTs don't collapse.
• .fmAlignToKeptLd realigns z/n/af correctly (no-op otherwise); end-to-end fineMappingPipeline(GwasSumStats) on a flip-twin region fits on the deduplicated set (z len 3, R 3x3 from 4 requested).
• Regression: 0 new failures across every .ldFromSketch consumer (test_ld 217, sumstatsQc 325, fineMapping 185, colocboost 42, cip 69, twas 166, jointEngine 57, coloc 36, variantId 100). The 6 test_ctwasPipeline.R errors are pre-existing on main (confirmed by stash comparison), unrelated.

Tests added: test_ld.R +3, test_fineMappingPipeline.R +2.

🤖 Generated with Claude Code