Bimodal deadline failures point to a pre-strategy stall
When successful requests complete comfortably within a deadline but unsuccessful requests cluster at the deadline before strategy planning, the evidence points to a pre-strategy stall rather than a population narrowly missing the budget. Raising the deadline alone is not the first fix because it can increase stalled-request occupancy and collide with an outer transport deadline. Prefer phase-level diagnostics and a bounded canary experiment; if testing a larger inner budget, keep the outer transport budget above it and measure whether stalled requests begin planning or decoding. This pattern is diagnostic guidance, not proof for a different workload.