Fix data-loss and alignment defects from the post-1.2.0 audit (plan A)

[druvus]

Implements plan A of docs/superpowers/specs/2026-10-01-post-1.2.0-audit-design.md (items A1-A10): defects that lose user data or produce a wrong alignment. The spec and the three plans (A, B, C) are the first commit of this branch.

What changes

• MAFFT backend reorients reverse-strand genomes and contigs (--adjustdirection). Before: identity about 0.40 for a reverse-complemented query, 0.97 after.
• reassort alignment-fraction filter compared in skani's percent scale (it never fired).
• fill-references aligns the last round's downloads (final.msa.fasta); --curate runs before each build, including round 1 for a supplied collection; --reference is never removed by curation.
• find-references --download <collection> --curate removes only what the run downloaded.
• reassort --scan-segments cannot write outside its output directory; names that sanitise alike get separate directories.
• MAF backends keep uncovered backbone contigs and write unplaced genomes as all-gap rows, named in a warning.
• progressiveMauve rows named from the staged label.
• Whitespace in sequence lines is cleaned at staging; malformed aligner output is reported by file.
• A run no longer clears an existing, non-empty <output>/collection that no earlier run created (found by the branch review; pre-existing).

No region caller, default or recomb output schema changes. The two shipped examples give the same regions as on main.

Checks

• ruff check src tests validation, mypy src: clean.
• pytest -m "not requires_binary": 628 passed (586 on main).
• pytest -m requires_binary with mafft 7.526 and minimap2 2.31: 5 passed.
• Each new test was run against the pre-fix source and failed there.

Not verified

• progressiveMauve: checked against a simulated XMFA only; the tool is not installed here.
• Cactus path of the MAF change: no binary.
• validation/run_reassort_benchmark.py after the filter fix, and validation/run_hybrids.py: not run (both need network). The reassort precision/recall may move now that the filter is active.
• --adjustdirection on short or highly divergent contigs and inverted repeats: only the two integration cases were measured.

Review

One independent whole-branch review; its four important findings are fixed in 351d47d:
--reference had become the sibling test's anchor (a regression of this branch: a distant reference caused every closer genome to be dropped), a silent all-gap row for genomes aligned only away from the backbone, and the unguarded removal of <output>/collection. One limit is documented, not changed: a sibling that is itself the closest genome in a supplied collection becomes the curation anchor and is kept.

Deferred minor findings from that review:

• an empty supplied --collection is seeded and then curated before round 1;
• unique_scan_dirs compares names case-sensitively (HA / ha share a directory on a case-insensitive filesystem);
• segment_scan.tsv does not record the scan directory name;
• ref_contigs lengths are not checked against the MAF src_size;
• two ValueErrors in maf_to_fasta.py still surface as "Unexpected error";
• the whitespace-cleaning warning repeats on every build of the fill loop;
• docs/aligners.md says an unplaceable genome is named in a warning for all backends; only the MAF path warns;
• a curate_collection_dir docstring overstates what happens to a download that duplicates a protected genome;
• after a final build, the last row of fill_summary.tsv describes the alignment before the last downloads.

Plan B (fix-audit-report-faithfulness) is stacked on this branch.

🤖 Generated with Claude Code