Constraint-Carrying Search: When Explicit Contracts Improve World-Model Planning

When does an explicit constraint representation improve world-model search over learned masking, and when is direct symbolic filtering already sufficient? We answer this as a two-axis regime question: specifications range from executable DSLs to mixtures of language and schema fields, while constraint-relevant state ranges from fully observed to latent. The Compiled Admissibility Interface (CAI) canonicalizes entities, dependencies, units, precedence, and progress once, then reuses their identifiers in bounded-continuation supervision, a horizon-conditioned graph compiler, runtime masking, and branch diagnosis; belief disagreement conservatively retains actions at uncertain constraint boundaries. Against a capacity-, protocol-, and search-budget-matched flat encoder on WebArena, WorkArena, a typed-API sandbox, and compositional gridworld, the complete CAI interface improves macro success by $+1.7$ points with paired 95% interval $[+0.8,+2.6]$, with positive point estimates in all four domains, $0.06$ fewer violation events per episode, and $2.4$ seconds lower planning time. At this operating point CAI removes $34$ of expansions relative to unpruned search while maintaining WM-FBR of at least $0.98$. The factorial phase diagram localizes the premium: it reaches $+2.5$ points for mixed specifications under partial observability, whereas the DSL-only/full-observation cell lies within the pre-specified practical-equivalence margin of symbolic filtering. In separately stratified boundary samples, method-blind executed replay records feasible-recovery false-negative rates of $2.4$ and $2.9$ for CAI, compared with $5.2$ and $5.9$ for Flat. The resulting regime map turns explicit constraint structure into a concrete choice of search interface rather than a universal default.