Event-Grounded Re-Entry: Boundary-Conditioned Top-Stack Control for LLM Workflows

VERIFY$\to$COMMIT is grammar-valid; without a qualifying verifier event for the latest candidate, the realized submission is nevertheless premature. Event-Grounded Re-Entry (EGR) uses that exact failure point as a computational synchronization boundary in one incremental decoder trajectory. A hard-masked transition head selects phase-specific rank-32 adapters; the committed canonical delimiter passes once through shared layers 1–28 and then through the selected layers 29–32, while prior KV entries remain immutable. On the disjoint 480-task ToolBench test, EGR keeps declared illegal transitions at zero, reduces invalid actions from 19.6% to 18.1% and commits lacking qualifying verifier evidence from 9.0% to 8.1%, and raises task success over FSM masking from 51.4% to 53.3% (paired difference +1.9 [+1.0, +2.7]). Against a 41.96M-parameter, budget-matched all-layer routed-LoRA control, EGR gains +1.1 points [+0.2, +2.0] at lower model-side latency. The FSM-matched success gains recur on HumanEval+ Verify (+1.8), a disjoint 240-issue SWE-bench extension (+2.8), and WebArena-180 (+2.6); replacing the selected VERIFY bank with the learned ACT bank raises the missing-evidence rate by 31.2 points [26.8, 35.6]. EGR runs at 1.05–1.06$\times$ model-side latency, compared with 1.38–1.43$\times$ for full-prefix stage restarts, making externally observed events useful control points for both workflow semantics and efficient neural computation.