Odu Coverage Map: Session-Level Entropy Monitoring and Fisher–Rao Escape Steering for Semantic Loop Prevention

WP-20 measured standalone TERA recall@10 at 0.017 and correctly identified the HNSW residual as the real discriminator. This result has been read as a weakness of TERA geometry. We show it is evidence that TERA is being used to answer the wrong question. For memory retrieval, TERA is weak. For generation coverage monitoring, TERA's coarseness is the feature: with only 256 Odu cells, the complete session occupancy distribution fits in 1 KB of memory and admits exact closed-form entropy computation at O(1) cost per chunk. We introduce the Odu Coverage Map (OCM), a zero-parameter runtime component that maintains per-cell occupancy counts, computes the session entropy H_session over the Odu lattice, detects per-axis collapse using WP-17's factorized entropy diagnostic, and steers generation toward underexplored cells via Fisher–Rao escape directions (WP-15). Combined with a post-generation HNSW greedy mesh check, OCM provides the two-level loop detector that WP-21 requires. Empirical results across four configurations and a topic/model ablation are reported.Revision 2 adds §10 and Appendix Option C: the escape search's cold-cell targets now fall back to persisted per-cell TERA centroids served by the WP-23 .aamt substrate (aamt_odu_profile → OduCoverageMap.seed_persistent) before resorting to abstract geometric cell centers — previous sessions' memory now anchors escape targets to real places. Session occupancy counts are deliberately never seeded: the cross-session prior refines where an escape steers, never whether one triggers, preserving the session-local entropy signal this paper exists to compute. The tree-convention/axis-major cell-index permutation is verified against the reference quantizer across 20,000+ coordinates including every quartile boundary, and the full chain is verified end-to-end with seeded/unseeded/distance controls.