Retention Is Not Selection: An Audited Failure Analysis of Multi-Hypothesis Homography Release

Abstract

Repetitive planar appearance can produce coherent but spatially wrong registrations. Multi-Hypothesis Ambiguity Rejection (MHAR) retains distinct homography candidates in a fixed-capacity bank and releases only when the bank selected by its frozen rule is SINGLETON. On the predeclared 1,620- pair LinTrack registry, M-V8 contains a reference-consistent mode in 1,183 pairs versus 881 for B-R0; all 338 MHAR releases match B-TOP1 in candidate ID, matrix, and support set. At the same coverage, MHAR has 64 wrong homographies versus 34 for evaluator-only B-CONF-Match@338. Alias-positive periodic pairs yield empty valid pools. Historical robot outputs provide retrospective diagnostics; a coordinate audit identifies a unit mismatch and no corrected downstream replay is claimed. The study separates candidate retention, release quality, adapter implementation, and downstream utility.