fix: split traced filaments at sharp bends without a junction (v2.0)

[uermel]

The v2.0 counterpart of #126; the patch is identical.

Two filaments that touch end to end skeletonize as one bent branch with no junction: a V, or one filament ending against another at an acute angle. The pairing at junctions never sees the turn, so seg2fil traced them as one hairpin filament. On dataset 10426 with easymode actin, such joins turned by up to 159 degrees.

Change

• A chain from a semantic segmentation is now split wherever it turns by more than --max-bend between its directions over the direction window before and after a point. This is the same rule and window that junctions use.
• Each piece loses one label radius at the split, as a branch does at a junction, and the new ends are not extended.
• Closed chains and instance segmentations are kept whole.
• The run summary reports how many filaments were split.
• The split happens before fitting, so each piece is stored with the default Catmull-Rom curve from feat: seg2fil stores editable Catmull-Rom curves by default (v2.0) #117 like any other filament.

Effect

• 10426 actin, run 16853 at threshold 0.2: kinks over 45 degrees drop from 4.9% to 1.5% of filaments, and length-weighted straightness rises from 0.956 to 0.987, with no change in traced length.
• 10521: none of the 102 microtubules traced on 58 tomograms changes.

Tests

• New test_two_filaments_touching_end_to_end_are_two_filaments: a V traces as two straight filaments, with one split counted.
• Checked by hand: default seg2fil on the same V stores two straight Catmull-Rom filaments.
• test_filaments.py, test_converter_signatures.py and test_run_names_option.py: 53 passed on v2.0 with the locked environment.

🤖 Generated with Claude Code