nx_align_score.nx

nx_align_score.nx -- score chained alignment via Smith-Waterman. license_tier: INDEPENDENT_REDERIVE genealogy_id: international-research-sources/li-2018-minimap2-seed-and-extend G1.4 of NISHI_GENOMICS_SUBSTRATE_ROADMAP.md. The integration primitive that closes the seed-and-extend loop: query+ref packed DNA -> nx_align_minimizer (G1.1) -> nx_align_match (G1.2) -> caller pre-sort by r_pos -> nx_align_chain (G1.3) -> THIS PRIMITIVE (G1.4) -> alignment score + region Given a chain (indices into the original seed arrays), extract the implied alignment region: q_start = seeds_q[chain[0]] q_end = seeds_q[chain[chain_n-1]] + k // last seed spans k bases r_start = seeds_r[chain[0]] r_end = seeds_r[chain[chain_n-1]] + k Then unpack the 2-bit DNA in [q_start, q_end) and [r_start, r_end) into per-byte 2-bit codes and run Smith-Waterman over them. The returned score is the local-alignment score for the implied region; it is >= chain_n * k * match_score under exact match and degrades with gaps + mismatches in the inter-seed regions. Why not score the chain seeds alone: - A chain proves co-linearity of seeds but says nothing about the inter-seed gap quality. Two seeds 3bp apart in q and 50bp apart in r implies a large insertion that should drop the alignment score. Only SW captures that geometry. - Real BWA-MEM / minimap2 use the chain to GUIDE extension (banded SW around chain diagonal) rather than re-aligning the full region; that optimization is G1.5. For G1.4 reference correctness we re-align the full region. What G1.4 does NOT do (deferred): - Banded SW restricted to chain diagonal (G1.5 perf path)