From Proposal Recovery to Reliable Recognition: A Provenance-Aware Measurement Framework for Open-World 3D Perception

We audit the object--evidence identity and stage denominators of a frozen LiDAR--camera detector with learned post-processing heads. Motorcycle and bicycle annotations are withheld during eight-class adaptation, although both classes occur in official detector pretraining and may occur in CLIP pretraining. On 500 nuScenes samples with 776 LiDAR-valid target objects, a coordinate origin audit corrects proposal recovery from the historically reported 176 to 92 (11.86%). Of 82 recovered objects with valid semantic evidence, 45 are correctly classified (54.88\%). Frozen recognized novel routing remains at 43 true positives, 747 false positives, and 733 false negatives. An object ledger attributes false negatives to 684 recovery misses, 10 missing evidence cases, 37 semantic errors, and two margin failures. Temperature scaling, fitted only on known calibration labels, improves the binary Brier score from 0.3610 to 0.1093 and NLL from 0.9600 to 0.3581; their paired component bootstrap differences exclude zero. The frozen M4 head has training/evaluation feature and target mismatches; an offline aligned M4 control with shared M1 correctness features improves the point Brier/NLL relative to frozen M4, but paired intervals versus M1 still include zero, and the historical router is unchanged. Joint geometric coverage on eligible unknown rows is 0.4516 for PCGU and a global residual baseline and 0.6452 for a standardized residual baseline, without a category shift guarantee. Genuine view deletion is evaluated. Historical canonical regenerative timing remains unrestored (256 versus 253 unknowns); a disclosed current-stack session timing protocol accepts 100 measured frames. A locked scene-ID-excluded P5 dump (5504 tokens/137 scenes) is reported separately under migration evaluation (RN 103/2909/3338). An independent CQR baseline has lower unknown joint 7D coverage than PCGU on the historical, v3-500, and P5 cohorts. Aligned M4 and temperature scaling substitutions leave v3-500 and P5 recognized novel counts and assignments unchanged at the frozen 0.05 gate. The evidence supports a limited, reproducible audit rather than deployment readiness or broad unknown-category generalization.