Fix phantom selections at window boundaries in interleaved solve

[JoOkuma]

When the CLI solves the ILP in windows (interleaved evens then odds), the boundary slices between batches ended up with phantom selected=True rows and broken parent_id references. Result: overlap-constraint violations and dangling parents in the final DB.

Refactored SQLTracking with the two-pass design discussed on the PR:

• Each batch reads the DB to detect whether its neighbours have already committed -- that determines whether each side is anchored.
• A non-anchored side keeps the configured overlap on the solver window and commits one slice past the inner range so the next-solving neighbour finds its anchor.
• An anchored side shrinks the solver window to share exactly the anchored slice with the neighbour and commits only the inner range, leaving the neighbour's prior write intact.
• The artefact-prone outer overlap is never committed -- only the next slice in, where the optimisation is reliable.

To keep the second pass from spawning or terminating tracks at the now-interior seam, MIPSolver.add_nodes gets free_appear / free_disappear flags; SQLTracking sets them False on the anchored side(s) so the solver pays the regular appearance/disappearance penalty there.

add_solution writes the commit range with full parent_id except at the leftmost slice when that slice is also the solver's leftmost (the solver has no incoming edge there and would write NO_PARENT, clobbering the previous batch's lineage).

Synthetic test runs cleanly for window_size ∈ {2,3,4,5,10,100}, overlap_size ∈ {1,2,3}, 10-30 frames, under both interleaved and sequential execution. _validate_tracking_solution now fails on dangling parents as a regression guard.