CoreML delegate emits unloadable multi-method program (0x23 InvalidProgram); MLX and XNNPACK load the same graphs fine

[msluszniak]

🐛 Describe the bug

Lowering a multi-method program with CoreMLPartitioner produces a .pte that cannot be loaded: every method fails with 0x23 (Error::InvalidProgram). The same exported programs lowered with MLXPartitioner load and execute correctly, so the model and the multi-method plumbing are fine and the problem is specific to the Core ML delegate.

Model: the Kokoro TTS duration predictor, lowered as three methods (forward_32, forward_64, forward_128) differing only in token count.

programs = {f"forward_{n}": ep for n, ep in predictors.items()}   # n in {32, 64, 128}

prog = to_edge_transform_and_lower(
    programs,
    partitioner=[CoreMLPartitioner(compile_specs=specs, lower_full_graph=False)],
).to_executorch()
Path("dp.pte").write_bytes(prog.buffer)

Runtime.get().load_program("dp.pte").load_method("forward_32")
# RuntimeError: Failed to load method forward_32, error: 0x:23

to_executorch() succeeds and writes a plausible artifact (35.8 MB at fp16, 62.7 MB at fp32). The failure is at load, before any execution.

Observed

backend	export	load + run
XNNPACK	OK	OK (this is the shipping artifact)
MLX	OK	OK, all three methods, correct output shapes
Core ML	OK	fails 0x23 InvalidProgram, all three methods, fp16 and fp32

The MLX row is the useful control: identical exported programs, identical multi-method structure, same to_edge_transform_and_lower call shape, only the partitioner differs.

Reproduced at both compute_precision=FLOAT16 and FLOAT32, with minimum_deployment_target=ct.target.iOS18.

Versions

ExecuTorch 1.4.1, coremltools 9.0, torch 2.14.0.dev20260702, macOS arm64 (M4).