[Jaureguy760]

Bulk Rust counting scanned BAM records before choosing a SNP, whereas archived Python processes SNPs in input order and consumes each read name once per chromosome. Paired mates could therefore contribute to a different SNP, and deletions/reference skips could incorrectly leave a name eligible for a later SNP. The biallelic SNP path now preserves Python's per-SNP fetch order and seen-name timing. Existing non-SNP counting remains outside this fix.

Also restores per-SNP pseudocount removal from reported bulk region counts, fixes single-cell preparation and differential analysis for cells in obs / variants in var, and widens differential sums to UInt64. Runtime Polars and Galaxy CI dependencies are constrained to compatible versions.

Counting comparisons execute the unmodified count_snp_alleles function from a verbatim source fixture pinned to Python commit 41354fc. Cases include reversed SNP order, paired mates, multi-SNP reads, repeated read names, deletions, reference skips, overlapping region membership, chromosome separation, one/two threads, and the repository's shared BAM/VCF fixture. This replaces a prior Python reference that reproduced Rust semantics.

Validation on this revision: 170 local Python tests passed, 82 skipped, 1 deselected, including all 26 executed counting parity checks; 117 Rust unit tests passed, 7 ignored. Ruff source/tests, formatting and diff checks passed. GitHub checks are running on the new head.

Limitations: per-SNP fetching may increase index seeks and repeated decoding; large-cohort performance has not been measured. Analysis references are synthetic. The archived linear model requires explicit scalar/array adapters and ignores the phased argument for multi-SNP tests; those four cases are not certified as parity. This does not establish complete donor statistical parity. PR #133 depends on this compatibility change. No merge or release is included.