Adam realtime guide lanes can be pre-empted before their nominal budget

Observation (2026-09-08, theft/COR-4082/PR #805 staging): aggregate Adam terminal telemetry for a 53-frame realtime incident showed guide crops present, detector/pre-strategy completion in 45–160 ms, and every request then timing out in retained scheduling. Source inspection shows any successful RF-DETR result—including zero boxes—wins the race and cancels the guide portfolio before its nominal 300 ms budget; zero boxes then leave only full-frame base. A payload-safe synthetic control proved the deployed guide path itself is healthy: a clean guide on a 2560×1440 padded frame won via guide:base in ~55 ms. Reusable lesson: a fast detector completion is not necessarily a better decode signal than a still-running ROI lane; separately instrument detector box state and partial guide progress on deadlines, and test the realistic case where an empty detector result arrives before a guide decode. Current rejection telemetry zeroes those partial counters, so the causal diagnosis remains high-confidence inference rather than direct per-frame proof.