[Audit/Correctness] same_objects counts many-to-one IoU overlaps as one-to-one matches #67

Evidence (Sad Catapella, Batch 1; main d7a538e)

• pair_counts, lines 317–320 sets children_matched_one_to_one and parents_matched_one_to_one by taking separate unique IDs from all edges exceeding IoU. It never constructs a one-to-one matching or checks degrees.
• _relation, lines 484–489 treats those counts as evidence of same_objects.
• Contract §15.1 expressly requires one-to-one matching.

Deterministic counterexample (code-path proof; not executed)

A parent mask has one four-pixel object labelled 1. A child mask partitions the same four pixels into two disjoint two-pixel objects labelled 10 and 11. Each child's intersection with the parent is 2, union 4, IoU = 0.5. The implementation counts 2/2 children and 1/1 parents as matched and returns same_objects, despite a 2:1 relationship. Lowering the configurable IoU threshold makes this error more prevalent.

Impact

Misleading segmentation-equivalence suggestions and incorrect same_objects evidence in mask_relationships.json, potentially interpreted by downstream bricks as equivalent object identities.

Fix / acceptance

Build a bipartite matching of child and parent instances (per plane) from IoU-eligible edges, maximising cardinality with deterministic tie-breaking; count only matched endpoints. Consider sparse matching to avoid dense K×K matrices. Add regression for the 2-child/1-parent example at IoU=0.5 and a configurable lower threshold; preserve existing genuine 1:1 tests. Benchmark sparse/dense masks separately; no benchmark has yet been run.

Classification: confirmed source-level semantic defect, with analytic counterexample. No repository tests executed.