Persistent AI agents lose judgment — reasoning texture, eliminated hypotheses, methodology, negative knowledge — when context is compacted. Unlike factual memory, judgment cannot be retrieved because it was never discretely stored. The agent experiencing compaction cannot perceive the loss (the introspection gap). We present the Unified Cognitive Substrate (UCS), fusing a toroidal routing engine with the Emergent Judgment protocol. Engine validation: an 11-experiment program identified an architectural disconnection (v1.1) and validated a separated context energy field (v1.2), achieving 1,563× improvement in context-sensitive routing differentiation. Retrieval index benchmark (v2.1): we added an inverted index over routing snapshots (RoutingSnapshotIndex) to pre-filter judgment-node retrieval. On structured, clustered operational data, the pre-filter isolates 33 relevant nodes out of 99 (33%), reducing the scoring set by 67% (min_signals=2). On random baseline data, the pre-filter scored 79/100 candidates (79%) with min_signals=1. Paper versioning: this record hosts the current paper PDF (ucs_paper_v3.pdf) and associated updates (including the retrieval benchmark subsection).