2026-04-28 - Stream-Level Two-Model Chain Breakthrough

Intent: Prove true cross-model chaining inside one stream-level execution stream by using stream-level encode hooks instead of public host roundtrips, following the validation-first notes measurement/validation discipline and the research/ANE_CHAIN_SCHEMA.md focus on stream-level ANE behavior.

Setup: CoreML imports stream-level symbols from CoreML execution runtime, including stream-level prepare and encode hooks. local two-stage stream probe now resolves those runtime symbols. Successful sequence: bind stage A input/output via CoreML operation methods; bind stage A output feature/memory object into stage B input binder; bind stage B output; call stream-level prepare+encode for stage A and stage B operation handles into one stream; then call the CoreML stream execution path.

Result: PASS on the tiny distinct-input control toy_a(x+1) -> toy_b_h(h*2) with no h provider. Stage A hidden was [2, 3, 4, 5]; stage B hidden was [4, 6, 8, 10]. This proves stage B consumed stage A’s output buffer and is the first true two-model stream-level chain in this repo.

Surprise / hurdle: The key unlock was discovering that the public CoreML/CoreML execution runtime stack already imports the stream-level prepare/encode symbols, making it possible to encode two operation handles into one stream while wiring the intermediate through runtime binders.

Lesson: Stream-Level encode access turns cross-model chaining from a host-roundtrip problem into a stream construction problem.

Next: Validate the same path on two Phi layer-range shards against the public host-roundtrip output, then profile hidden-state copy removal.

Refs: research/ANE_CHAIN_SCHEMA.md