2026-04-28 - Phi Long-Decode Topology Baseline Moves to 20+4+6+2

Intent: Re-check existing fused layer topologies after private E5 timing showed boundary removal was low leverage and the layer stack remained the dominant cost.

Setup: Profiled existing public CoreML manifests with the same 4-way LM head, 10 warmup calls, 100 generated tokens, and --profile. Ran strict residency on the 20+4+6+2 layer shards and numerical range golden gates for [0,20) and [20,24).

Result: 20+4+6+2 is the new measured long-decode best: 17.203 tok/s, layers_ms=53.039, head_predict_reduce_ms=5.084. The 16+8+6+2 rerun over 100 generated tokens measured 16.596 tok/s, layers_ms=55.084. Gates passed for 20+4+6+2: [0,20) residency compute_non_ane=0, golden cos_hidden=0.998546; [20,24) residency compute_non_ane=0, golden cos_hidden=0.999446; tail shards were resident in the same check and previously validated.

Surprise / hurdle: The earlier short-run 16+8+6+2 winner is not the best long-decode point. Larger front fusion to 20 layers reduces layer overhead enough to win, while [0,24) remains beyond the compiler residency cliff.

Lesson: Fused topology should be selected on long decode profiles, not only short bursts; 20+4+6+2 is the current public baseline under ANE-only gates.

Next: Use phi4mini_runtime_meta_20_4_6_2.json for further public-runtime comparisons, and keep [0,24) rejected unless compiler behavior changes.

Refs: research/ANE_CHAIN_SCHEMA.md