2026-04-28 - Stream Dispatch Bridge Outcome

Intent: Record the outcome of the stream dispatch investigation after checkpointing the public Phi-4-mini ANE runtime state, following call-hoisting/strength-reduction discipline for reducing CoreML shard boundary costs.

Setup: Checkpoint/tag phi4-mini-ane-q8-fusion-17tok-2026-04-28 was created on commit f273a47 before investigation. Local probes inspected whether the public CoreML load path exposed enough stream-level structure to reduce shard-boundary overhead.

Result: The stream-level boundary was observable, but the direct probe still failed on current Phi public-CoreML shards at a legacy compiled-artifact contract. The useful lesson was that public CoreML loading already creates a program-library object that can explain why shard-boundary overhead exists.

Surprise / hurdle: The public CoreML stream runtime path already owns a usable program-library handle, while the direct stream dispatch probe is blocked by older CoreML execution runtime artifact expectations that public .mlmodelc shards do not satisfy.

Lesson: The next unsupported path should investigate the CoreML stream runtime handle/operation bridge rather than trying to synthesize legacy CoreML execution runtime artifacts first.

Next: Follow the stream runtime program-library handle toward operation/chaining surfaces for already-loaded CoreML models; keep public MLComputePlan residency and golden validation as acceptance gates before any performance claim.

Refs: research/ANE_CHAIN_SCHEMA.md