[codex] Fix boundary read evidence counts

[iskandr]

Summary

This PR fixes read-evidence undercounting at locus boundaries and adds regression coverage around the affected cases.

What changed

• allow insertion-locus reads to survive when the inserted bases are at the beginning or end of the aligned read
• add direct LocusRead regressions for insertion support at read boundaries
• add read-evidence regressions for SNVs at read starts/ends
• add read-evidence regressions for insertion counts at read boundaries
• add splice-boundary regressions for SNVs at the last base before an intron and the first base after an intron
• bump isovar to 1.4.22
• update the CLI row-count expectation to reflect the recovered real-data allele-read row
• remove a stale test comment that still described the old exon-boundary miss

Why

ReadCollector.locus_read_from_pysam_aligned_segment(...) handled insertions by requiring mapped flanks on both sides of the insertion site. That drops reads whose inserted bases occur at the start or end of the aligned query even though those reads still contain valid allele evidence. Those drops then propagate into allele-read exports and read counts.

The older exon-boundary/read-boundary bugs were also only weakly pinned down by tests. This PR makes those cases explicit at the unit level.

Impact

• insertion-supporting reads at read boundaries are no longer silently dropped
• isovar-allele-reads emits one more real-data row in the bundled fixture set because the missing boundary read is now recovered
• read-boundary and splice-boundary behavior is now covered by regressions

Validation

• ./lint.sh
• ./test.sh

Closes #49
Closes #23
Closes #24
Closes #55
Refs #14

[coveralls]

coverage: 93.441% (+0.01%) from 93.429% — fix-read-evidence-boundaries into master