2026-04-28 - Phi E5 ObjC Sync-Point Experiment

Intent: Test whether CoreML’s MLPredictionSyncPoint route can express the stage-A completion / stage-B wait dependency without directly calling raw E5RT event bind functions.

Setup: Widened coreml_e5_class_dump.m to include event/sync classes. Found private MLPredictionSyncPoint with initWithSharedEvent:value: and hidden MLPredictionOptions accessors for completionSyncPoint and waitSyncPoints. Added --objc-sync-points to e5_two_op_stream_probe, creating an MTLSharedEvent and passing completion/wait sync points through the private bind methods. Added --objc-sync-update as a separate explicit experiment for the private future-value update hooks.

Result: --objc-sync-points is stable: raw prepare, encode, and execute all succeed, but stage B remains all zeros. --objc-sync-update segfaults at _updateCompletionEventFutureValuesWithCompletionSyncPoint: in the manual raw lifecycle, so it is off by default.

Surprise / hurdle: Sync-point options are accepted by MLPredictionOptions, but they do not repair the raw multi-op path when CoreML’s normal stream preparation/async submission lifecycle is bypassed.

Lesson: The event dependency is probably not just data stored in options; it is created and advanced by a specific CoreML operation lifecycle. Calling the leaf update hook directly is unsafe.

Next: Trace or reuse prepareForInputFeatures:options:error: / prepareAsyncSubmissionForInputFeatures:options:error: ordering to learn when CoreML binds shared events, updates future values, and raw-encodes the operation.

Refs: research/ANE_CHAIN_SCHEMA.md