Commit 28b44ef replaced the Stage 0 BC sort with an order-preserving
dedup that std::abort()s unless the input BC table is already
globalBC-sorted. That contradicts the tool's purpose (PURPOSE (a):
repairing non-monotonic fGlobalBC in merged AO2Ds), so it aborted on
exactly the files it exists to fix -- the "doesn't run to completion"
regression.

Revert only that Stage 0 change (restore stage0_sortBCs, drop the
abort). Keep 28b44ef's Stage 1b track regrouping and the fIndexTracks*
remaps: that mechanism is the correct fix for the -1-group split from
b11cd3d (track fIndexCollisions values were remapped but rows left in
input order, so per-collision groups -- including the -1 ambiguous
group -- were no longer contiguous, tripping
ArrowTableSlicingCache::validateOrder downstream). It only ever failed
because the aborting Stage 0 stopped it running. Sorting BCs implies a
BC -> collisions (Stage 1) -> tracks (Stage 1b) reorder cascade, which
those stages now handle completely.

Add a validator check (checkCollisionGroupContiguity) mirroring O2's
slicing invariant: every fIndexCollisions group must be one contiguous
run. Added a history note in Stage 0 against re-introducing the
order-preserving abort.

Not yet validated against a real merged AO2D / analysis task.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>