Multi-Channel Mechanistic Signatures of Agent WANDERING: Classification, Causal Localization, and Behavior-Legible Response to Intervention

Mechanistic characterization of agent WANDERING on N=99 Qwen3.6-27B SWE-bench Pro trajectories. 60 multi-channel features (text, tool-use, per-layer residual, temporal) classify SUCCESS/LOCKED/WANDERING at macro-F1 0.636 (p=0.001), after a transparent walk-back from a leaky 0.987 baseline. Stability selection independently recovers a mid-to-edge mechanism (LOCKED->L43, WANDERING->L11), and an LLM-judge bridge to a human taxonomy places ~60% of both LOCKED and WANDERING into one text category, matching a mechanistically weak boundary (p=0.035). Finally, the residual signature does not predict which agents flip to finish under a companion L11 injection run (LOO-AUC 0.619), but tool-entropy collapse depth does (AUC 0.768): response to intervention is residual-blind but behavior-legible. Companion to Tool-Entropy Collapse (DOI 10.5281/zenodo.20368601).