2026-04-28 - E5 Binder Timing Controls Outcome

Intent: Pin down when private CoreML E5 port bindings become fixed, following the validation-first notes Dragon Book call-hoisting/strength-reduction discipline for removing shard-boundary materialization without relying on post-hoc host mutation.

Setup: Updated local artifacts to call MLE5InputPortBinder bindMemoryObjectForFeatureValue, operation-level _bindInputFeaturesAndWaitEvents / _bindOutputPortsWithOptions, stream _prepareForInputFeatures, and raw _executeStream on toy E5 control models.

Result: bindMemoryObjectForFeatureValue returned YES with the stage A output feature, but did not create a true chain. With toy_b_h and an explicit h provider [10,20,30,40], stage B output was [20,40,60,80], proving the provider input wins over the attempted A-output binding. After stream _prepareForInputFeatures, attempts to change port bindings failed with Port bindings cannot be changed while operation is in use in an execution stream. Raw _executeStream after prepare worked but used provider h; raw _executeStream without stream preparation failed with No operations have been encoded to the execution stream.

Surprise / hurdle: The binder API can accept memory objects and feature values, but stream preparation encodes the operations and locks the binding plan before any later MLFeatureValue or binder mutation can express a cross-model edge.

Lesson: True E5 cross-model chaining must be expressed before or inside setupOperationForInputFeatures or lower E5RT setup; post-prepare MLFeatureValue/binder mutation is too late.

Next: Search below the prepared stream boundary for an E5RT setup or port-link primitive that can connect output and input ports before encoding; do not apply the post-hoc binder path to Phi shards.

Refs: research/ANE_CHAIN_SCHEMA.md