Warm every streaming decoder window the scheduler can emit

A windowed neural-audio decoder must warm or capture every input length its startup scheduler can actually emit, not only the first and steady-state lengths. Otherwise intermediate startup chunks silently fall back to an eager path or fail under a graph-only implementation, producing latency spikes exactly during time-to-first-audio. Derive warmup lengths as left context plus each scheduled new-code count, and test the per-slot schedule itself; configuration-only coverage is insufficient.

Applies when: a streaming decoder uses a staged startup cadence before its steady chunk size. Keywords: streaming decoder startup chunks graph warmup left context latency