Empty the LD sketch for zero-variant sumstats entries

[Yining97]

Problem

When a study has zero variants in the analysis window, its entry early-exits QC ("fewer than two variants after pre-harmonization QC") with variantsIn = 0 — correct — but the retained ldSketch keeps the full genome-wide panel (13,205,438 rows, ~135 MB serialized) instead of being trimmed. Cause: .subsetSketchToRange and .subsetSketchToIds build their keep-set from the entries' variant positions/ids, so with zero variants they hit if (length(...) == 0L) return(ldSketch) and return the full sketch. Version-independent (0.6.9, 0.6.11). Results/audit are unaffected — only the serialized slot. Cost scales with the panel, not the region (a study missing 100 blocks writes ~13.5 GB of overhead).

Fix

A zero-variant object references no LD, so its sketch should be empty. Added .emptySketch() (zeros @snpInfo, preserves all other slots) and made both trimmers empty on genuinely-zero variants. .subsetSketchToIds stays conservative (full sketch) only for the pathological "ranges present but SNP mcol absent" case, so a malformed-but-non-empty entry is never blanked. Populated studies are unchanged.

Validation

• Zero-variant → 0-row sketch on both trimmers; emptied handle valid, getNSamples preserved, serializes/reloads. 200k-row synthetic sketch: 1.139 MB → 0.00052 MB (2199x smaller).
• Populated study byte-identical ([min,max] span / matched-id keep-set); conservative no-mcol case; NULL-safe.
• test_sumstatsQc.R: 328 tests, 0 failed (3 new tests, 16 expectations).

No API change; no change to results or the early-exit audit.

🤖 Generated with Claude Code