Stage M6 — Static decode plan on Apple *(old 20, Apple half)*

[thanos]

Stage M6 — Static decode plan on Apple (old 20, Apple half)

Goal: make the Apple hot decode path boring.

• After warm-up: weights resident, KV preallocated to context bound,
  scratch arena fixed, pipelines compiled and cached, zero heap
  allocation per token — asserted by instrumentation, not claimed.
• Bounded in-flight work; predictable encode/submit structure per token;
  ICB/replay decision from M3 finalized here.
• Memory report: weights / KV / scratch / peak RSS, per model and
  context length.

Gate: an allocation-tracking assertion passes over a long generation;
per-token latency variance is characterized (p50/p95/p99).
Tutorial: docs/tutorials/20-static-decode-plan.md.