fix: lazy polars import for Ivy Bridge CPU compatibility

[LuJiansen]

Problem

On Intel Ivy Bridge CPUs (Xeon E5-2670 v2), the polars-runtime-32 shared library crashes with Illegal instruction because it is compiled with AVX2/AVX-512 instructions that these CPUs lack. This affects both wasp2-map make-reads and wasp2-count count-variants.

Clusters with mixed CPU architectures (Ivy Bridge + Haswell in the same Slurm partition) cannot use --constraint to avoid these nodes.

Solution

Make polars imports lazy — move import polars as pl from module top-level into the functions that actually use polars. Add TYPE_CHECKING guards so static analysis can still resolve pl.DataFrame annotations without triggering a runtime import.

For count-variants (the most common counting path), add a polars-free alternative make_count_df_no_polars() using pure Python csv + the existing Rust BamCounter.

Files changed (5 files, +101 −29)

File	Change
src/counting/count_alleles.py	TYPE_CHECKING guard + lazy import + new make_count_df_no_polars()
src/counting/run_counting.py	Route non-gene paths through make_count_df_no_polars; remove unused imports and dead assignments
src/counting/filter_variant_data.py	TYPE_CHECKING guard + lazy import
src/counting/parse_gene_data.py	TYPE_CHECKING guard + lazy import
src/counting/count_alleles_sc.py	TYPE_CHECKING guard + lazy import

What is NOT broken

• make_count_df() (polars path) still works on CPUs with AVX2 support
• Gene intersection mode (GTF/GFF) still uses polars
• mapping/intersect_variant_data.py was already lazy-imported — no changes needed

Testing

• ruff check: All checks passed
• ruff format --check: 5 files already formatted
• py_compile: 5/5 files pass
• Full ATAC-seq WASP pipeline (fastp → bwa → make-reads → filter → haplotag → counting) completes on Ivy Bridge nodes
• _parse_intersect_tsv unit tested
• Backward compatible: existing polars-dependent code paths unchanged