Fix nSample parameter bug and enforce exact N-in = N-out row counts

Root cause: the R implementation was using the GWAS sample size (e.g. 5000)
for the nSample parameter, but the original DENTIST binary always uses the
LD reference panel size (ref.N from the .fam file, e.g. 165). This parameter
controls the SVD truncation rank K = min(idx.size(), nSample) * propSVD,
so using the wrong value produces completely different imputation results
(imputed_z correlation of only 0.40 instead of 1.0).

With nSample correctly set to the reference panel size (NULL default in
dentist_from_files), R now matches the binary to machine precision on both
the 2500-variant subset and full 17421-variant toy dataset:
  - imputed_z max diff: 5e-6, correlation: 1.0
  - rsq max diff: 5e-7, correlation: 1.0
  - duplicate detection: 100% agreement

Changes:
- Update nSample docs in dentist(), dentist_single_window(), and
  dentist_from_files() to clarify it is the reference panel size, not
  the GWAS sample size
- Remove incorrect filter in merge_windows() that dropped variants with
  imputed_z == 0 & rsq == 0, which caused output to have fewer rows
  than input
- Rewrite unit tests (25 -> 43) with exact row count assertions
  (expect_equal(nrow(res), N)) instead of allowing ranges
- Add new tests: windowed mode exact row counts, outlier_stat formula
  verification, X matrix input, fill region coverage, column checks,
  original_z preservation

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Author: gaow

R/dentist_qc.R

@@ -46,8 +46,9 @@ resolve_LD_input <- function(R = NULL, X = NULL, nSample = NULL, need_nSample =
 #' @param R Square LD correlation matrix. Provide either \code{R} or \code{X}.
 #' @param X Genotype matrix (samples x SNPs). If provided, LD is computed via
 #'   \code{compute_LD(X)} and \code{nSample} defaults to \code{nrow(X)}.
-#' @param nSample The number of samples. Required when \code{R} is provided; inferred
-#'   from \code{X} when \code{X} is provided.
+#' @param nSample The number of samples in the LD reference panel (NOT the GWAS sample
+#'   size). This controls the SVD truncation rank K = min(idx_size, nSample) * propSVD.
+#'   Required when \code{R} is provided; inferred from \code{X} when \code{X} is provided.
 #' @param window_size The size of the window for dividing the genomic region. Default is 2000000.
 #' @param pValueThreshold The p-value threshold for significance. Default is 5e-8.
 #' @param propSVD The proportion of singular value decomposition (SVD) to use. Default is 0.4.
@@ -154,8 +155,9 @@ dentist <- function(sum_stat, R = NULL, X = NULL, nSample = NULL,
 #' @param R Square LD correlation matrix. Provide either \code{R} or \code{X}.
 #' @param X Genotype matrix (samples x SNPs). If provided, LD is computed via
 #'   \code{compute_LD(X)} and \code{nSample} defaults to \code{nrow(X)}.
-#' @param nSample Number of samples. Required when \code{R} is provided; inferred
-#'   from \code{X} when \code{X} is provided.
+#' @param nSample Number of samples in the LD reference panel (NOT the GWAS sample
+#'   size). Controls the SVD truncation rank. Required when \code{R} is provided;
+#'   inferred from \code{X} when \code{X} is provided.
 #' @param pValueThreshold P-value threshold for outlier detection. Default is 5e-8.
 #' @param propSVD SVD truncation proportion. Default is 0.4.
 #' @param gcControl Logical; apply genomic control. Default is FALSE.
@@ -531,10 +533,6 @@ merge_windows <- function(dentist_result_by_window, window_divided_res) {
       filter(index_global >= window_divided_res$fillStartIdx[k] & index_global < window_divided_res$fillEndIdx[k])
     merged_results <- rbind(merged_results, extracted_results)
   }
-  # Filter out un-imputed variants (imputed_z == 0 and rsq == 0) after merging,
-  # matching the original DENTIST binary which skips these when writing output.
-  merged_results <- merged_results %>%
-    filter(!(imputed_z == 0 & rsq == 0))
   return(merged_results)
 }
 
@@ -593,8 +591,10 @@ read_dentist_sumstat <- function(gwas_summary) {
 #' @param gwas_summary Path to the GWAS summary statistics file (DENTIST format:
 #'   tab-separated, 8 columns with header: SNP A1 A2 freq beta se p N). May be gzipped.
 #' @param bfile PLINK binary file prefix (expects .bed/.bim/.fam files).
-#' @param nSample Sample size. If NULL, taken from the N column of the summary file
-#'   (using the median value). Default is NULL.
+#' @param nSample Number of samples in the LD reference panel. If NULL (recommended),
+#'   uses the reference panel size from the genotype matrix. This controls the SVD
+#'   truncation rank K = min(idx_size, nSample) * propSVD. Note: this is the reference
+#'   panel size, NOT the GWAS sample size. Default is NULL.
 #' @param window_size Window size in base pairs. Default is 2000000.
 #' @param pValueThreshold P-value threshold for outlier detection. Default is 5.0369e-8.
 #' @param propSVD SVD truncation proportion. Default is 0.4.