Action-Catalog Shift Reveals Stable Latent-Trace Hallucination Routing

Hallucination routing requires both an accurate risk score and a remedy matched to each failure, yet that pairing can become coupled to the action vocabulary used during system design. We introduce catalog-conditioned generalization, a controlled stress test that substitutes benchmark- or operator-supplied actions while fixing prompts, graders, utility, access, training budget, and comparator families. TraceDispatch (TD) addresses this shift with validation-selected multi-depth anchors, learned $64$-dimensional compression, and an order-sensitive joint risk/action head executed within one generator pass. On public unseen-family evaluations spanning HaluEval-v1/Gemma-2-27B, BFCL-v3/Mixtral-8x22B, and LiveCodeBench-v3/DeepSeek-R1-32B, TD leads OO-Router by $+4.2$–$+4.4$ AUROC and $+6.2$–$+6.6$ corrected-output points. Benchmark-aligned substitution retains $94.9$–$96.3$ of AUROC and $93.9$–$95.7$ of intervention utility; an expanded controller frontier further places TD above output-only, matched-sketch, and richer-raw-state alternatives in every aligned catalog. TD remains top-ranked across five independently supplied catalogs spanning $K=3$–$6$; on blind human gold, it reaches $81.0$ AUROC and $61.4$ intervention utility at $38.0\mathrm{ms}$ on Llama-3.1-70B, while a human-only reconstruction of the upstream pipeline changes AUROC by just $-0.9$. The results establish catalog-conditioned generalization as a discriminating controller-selection criterion and show that budgeted ordered traces preserve actionable signal across changes in catalog source, model family, supervision, and serving stack.