Successor Re-grounding Audits Compositional Rollout Mismatch in Neuro-Symbolic Search

Offline latent diagnostics can fail to predict behavior when a learned transition model is composed recursively inside symbolic search. We evaluate a frozen quotient-geometry evaluation package on three generalized-planning domains and compare pure latent rollout with successor re-grounding, which re-encodes the symbolic successor that search has already computed before subsequent scoring. We do not present re-grounding as a universally available transition model; we use it as an interface audit for systems whose search procedure already enumerates symbolic or tool states but carries forward recursively rolled latent successors. In the headline paired protocol with one domain-agnostic every-step policy, valid-plan rate rises from $39.5\%$ to $99.0\%$ on VisitAll, from $28.8\%$ to $75.6\%$ on Gripper, and from $0.0\%$ to $24.5\%$ on Blocksworld, where the method still trails a stronger state-centric XGBoost frontier at $80.0\%$. Seed-level gains are positive for all VisitAll and Blocksworld seeds and for $8/10$ Gripper seeds, identifying Gripper as the heterogeneity case. Root-only re-grounding does not explain the repair, while budget sensitivity shows that widening the beam lifts Blocksworld every-step success to $38.0\%$. The finding is a compositional sanity check and partial repair, not a replacement for stronger models or broader robustness studies.