Mask the blind spots a lift between distant anchors used to drop

[Paururo]

What changed and why

With --mask_blindspots and --blindspot_liftover, the H37Rv Illumina blind spots are lifted onto each run's reference. The lift interpolates between shared unique k-mers and drops a position wherever the anchors around it are far apart. Blind spots are repeats, where anchors are scarce, so on other MTBC assemblies a third of them were never masked and went to the variant caller.

Checked against minimap2's alignment of each of the cohort's two references to H37Rv, with the MTBC-ancestor FASTA the container ships (same SHA-256):

	Now (--sample 10)	With --align-gaps
E1ASM0035: blind spots masked at minimap2's place	84,457 bp (63%)	130,854 bp (97.4%)
E1ASM0035: blind spots left unmasked	49,837 bp	3,440 bp
E1ASM0057: blind spots masked at minimap2's place	87,485 bp (65%)	131,708 bp (98.2%)
E1ASM0057: blind spots left unmasked	46,695 bp	2,472 bp
Bases masked elsewhere	~300 bp	700–1,100 bp

Sampling is not the cause: with every anchor kept, interpolation alone masks 66%. PREPARE_REFERENCE now lifts the blind spots with --align-gaps, the alignment between two anchors from #9. About 15% of the H37Rv blind spots are stretches these assemblies do not have, with nothing to mask.

Filling each blind-spot stretch between its lifted ends was also tried. It masked 1.4 points more but doubled the bases masked elsewhere, so it is not included.

Outputs and defaults

• The per-reference exclusion list gains the blind spots the lift used to drop. On an MTBC assembly that is about 45 kb more masked positions.
• The step takes about 15 s and 1 GB per reference, within PREPARE_REFERENCE's 8 GB.
• Nothing changes with --mask_blindspots false (the default) or --blindspot_liftover false.

How it was verified

tests/run_tests.sh unit   # 2,187 passed, 39 skipped
tests/run_tests.sh lint   # clean
mkdocs build --strict     # passes

The numbers above come from lifting assets/H37Rv_blindspots.bed exactly as PREPARE_REFERENCE does, onto the two real references, and comparing every position with minimap2. Lifted onto H37Rv itself, 99.4% of the blind spots land on their own position. The rest are at the two ends of the genome, where there is nothing to interpolate from, and scattered inside long repeats. --blindspot_liftover false is still the setting for a reference in H37Rv coordinates.

Checklist

• tests/run_tests.sh passes locally (unit and lint; the pipeline leg in CI)
• ruff check . is clean, and I did not restyle code the change does not touch
• New or changed behaviour has a test, and tests pin what the code does today (--align-gaps is covered in test_pathotypr_liftover.py)
• A new process in modules/ has a stub: block (not applicable: no new process)
• tests/data/ was not edited by hand (fixtures come from make_test_data.py)
• User-visible changes have a CHANGELOG.md entry and, where relevant, a docs update