Fix 6 pre-existing bugs across LD/RSS/TWAS pipeline

[gaow]

Summary

Fix 6 pre-existing bugs found during comprehensive code review. These bugs existed before the OTTERS/lassosum work but affect the pipeline.

Fixes

1. pval_hmp() — discarded duplicate p-values (R/misc.R)

• unique(pvals) silently dropped duplicates, changing the test statistic
• Fix: use all p-values as-is

2. rss_basic_qc() — lexicographic position comparison (R/sumstats_qc.R)

• separate() produces character strings; pos > "10" uses string comparison where "9" > "10" is TRUE
• Fix: mutate(start = as.integer(start), end = as.integer(end))

3. allele_qc() — removed ALL duplicate copies (R/allele_qc.R)

• vec_duplicate_detect() marks ALL occurrences (including first) as TRUE, so duplicated variants were entirely lost
• Fix: use duplicated() which keeps the first occurrence; warn with count of removed duplicates

4. compute_LD() population method — silent NA bias (R/misc.R)

• Cross-product divides by total N but per-column stats use non-missing counts; biased when missingness varies across SNPs
• Fix: warn when max-min NA rate difference exceeds 10%, suggesting method='sample' as alternative

5. gigrnd() — clamped GIG distribution to [0,1] (src/prscs_mcmc.h)

• The GIG distribution is unbounded above; clamping inside the sampler biases the distribution
• Fix: remove internal clamp, add arma::clamp(psi, 0, 1) after the GIG call in the MCMC loop, matching original Python PRS-CS exactly

6. sdpr() — buffer overflow with M < 4 (R/regularized_regression.R)

• sample_V() accesses a[M-2] with vector size M-1; with M=1 this is out-of-bounds
• Fix: add if (M < 4) stop("M must be at least 4") in R wrapper

Test plan

• PRS-CS compiles and produces valid results after gigrnd fix
• All psi values remain <= 1 (clamping works in new location)
• devtools::test() passes
• R CMD check passes

🤖 Generated with Claude Code