0.12.5

A patch release that finishes the postprocessing memory work started in 0.12.2, which turned out to help only when boxes were spread out, and fixes two correctness problems found while testing it.

🐛 Fixes

• Postprocessing memory no longer depends on how crowded the boxes are (#1417). Avoiding the N x N overlap matrix by storing one entry per intersecting pair is only a saving while boxes are spread out. Boxes piled on top of each other intersect nearly everything, so the pair count approaches N^2 and the stored list costs about as much as the matrix it replaced, which is exactly the layout that sliced inference produces on crowd scenes and small-object imagery. The merge loops read one row of matches at a time, and NMS and greedy NMM only ever read rows of boxes that survived, so rows are now answered from the STRtree on demand and peak memory is bounded by O(N + max_degree) whatever the layout. NMM reads a row per box rather than per survivor, so it keeps the stored pair list until that list is projected to be large.
• A non-positive match_threshold no longer builds a metric matrix (#1417). Both metrics are non-negative, so metric >= 0 holds for every pair and the adjacency is the complete graph. The dense path was still computing every overlap to discover this, which made the one configuration that needs no overlap computation the most expensive one and reproduced the out-of-memory failure from #1374 at 25000 boxes. The result now follows from the score order alone.
• NMM no longer merges a keeper into itself (#1416). A keeper was marked by leaving its merge_to_keep entry at -1, the same value that marks a box as unclaimed, so a box processed later could claim a keeper and append it to a merge list. With tied scores on duplicate boxes this produced a keeper listed inside its own merge list, and an index reachable as both a keeper and a merged box. Keepers now point at themselves, which makes them unclaimable without changing any other outcome.
• Zero-area boxes no longer trigger a divide-by-zero warning (#1416). The metric computed inter / denom for every pair and discarded the invalid entries afterwards, so a degenerate box emitted a RuntimeWarning. The division is now masked instead of the result.

⚡ Performance

Postprocessing on crowded and scattered layouts, IOS / 0.3:

layout	boxes	NMS before	NMS after	peak memory before	peak memory after
crowded	33337	47180 ms	135 ms	3416 MB	4 MB
crowded	20000	15931 ms	57 ms	1232 MB	2 MB
crowded	10000	2717 ms	25 ms	311 MB	1 MB
scattered	33336	539 ms	188 ms	65 MB	4 MB

• The numba backend picks its path by box density (#1418). Its NMS and greedy NMM loops ran an exhaustive pairwise scan whatever the input. Routing them by prediction count alone would regress crowded scenes, because the JIT loop skips suppressed candidates and stays competitive long after the numpy one stops being viable, so the neighbour count decides instead and is estimated from a sample of the boxes. The score sort was also an insertion sort, quadratic in its own right, and now uses the same lexsort the other backends share, which produces an identical order including ties.

avg. neighbours per box, 20000 bbox	NMS before	NMS after	greedy NMM before	greedy NMM after
0.0	909 ms	43 ms	842 ms	89 ms
0.8	781 ms	52 ms	757 ms	84 ms
5.1	495 ms	118 ms	479 ms	151 ms
20.1	241 ms	146 ms	236 ms	148 ms
78.9	141 ms	73 ms	143 ms	71 ms