Classify per-species loci in fetch and always emit the scaffold in FASTA headers

[gwct]

Fixes behavior where scaffolds for non-reference species were silently dropped from FASTA headers.

• Scaffold is always in FASTA headers. It previously lived only in an internal dict key, so --fasta-dedupe most-seq or --expected-species discarded it so it can't be resolved to a position in a multi-scaffold genome. Records spanning several source sequences now comma-join them rather than silently naming one.

• Per-element classification in maf_fetch_summary.tsv (12 new columns, always written, MAF or FASTA): how many species are a clean single locus vs split / multi_scaffold / multi_strand / no_bases, plus the extent and concentration of the splits.

• Opt-in --loci-table writes per-(element, species, block) rows with each species' source scaffold, MAF-frame coordinates and srcSize, so a per-species BED is a filter away.

• Reference-side columns renamed ref.n.overlapping.blocks / ref.block.bases, and interblock.distances removed. It was structurally all zeros and invited being read as a per-species contiguity check. The check survives as a ref-coverage-gap warning.

• Crash fix: -fh species-coords-id with a 3-column BED raised UnboundLocalError.

Why classification rather than just fixing the header: MAF blocks are contiguous in the reference only, so a stitched non-reference row can be several pieces from different places, and the header's span is a bounding box. On real data 88.5% of (element, species) pairs are genuinely one locus; the rest aren't, and now say so.

Breaking: the summary schema changed (a column removed, two renamed, 12 added), so positional parsers need updating. Per-region .maf/.fa output is byte-identical to v0.5.0 across all 14 tested configurations. 137 tests passing.

🤖 Generated with Claude Code