Research spike: bring-your-own-model serving on Core AI (sister-tool feasibility)

[Arthur-Ficial]

Part of epic #189.

Core AI's genuinely new capability is running non-Apple model weights on Apple silicon from an
.aimodel file (AIModel / InferenceFunction / NDArray), with compute-unit and caching control.
This is interesting but is a large, different project from apfel core.

Why this is NOT apfel core

• apfel ships and downloads nothing ("no downloads" is a selling point). BYO weights breaks that.
• Core AI gives raw tensor inference. An LLM serving stack still needs tokenizer, sampling, KV cache,
  and chat templating - i.e. reimplementing what FoundationModels gives apfel for free.
• It fits the apfel-family pattern (apfel-tag, apfel-spot, apfel-mcp, apfel-server-kit): a separate
  repo, not apfel core.

Spike deliverables (timebox, no production code)

• Smallest viable path: load an .aimodel, run one inference function on an NDArray, measure.
• What it takes to serve a real LLM on top (tokenizer + sampling + KV cache + templating).
• Could it reuse apfel's OpenAI-compatible server + ApfelCore types? (likely yes - that is the
  reuse story.)
• Recommendation: separate repo name (e.g. apfel-coreai / aimodel-serve), or drop.

Gating

Needs macOS 27 + Metal Toolchain. Pure research - decision only, no merge to apfel.

[Arthur-Ficial]

Update (2026-06-09): there is now an official Core AI <-> FoundationModels bridge - this spike just got much shorter.

The June 2026 Foundation Models updates add:

• a LanguageModel protocol - "Adopt the LanguageModel protocol to use any large language model - server or on-device"
• open-source CoreAILanguageModel and MLXLanguageModel backends.

That means a Core AI .aimodel (or an MLX model) can be driven through the existing FoundationModels session API (prompts, tool calling, structured generation) instead of reimplementing tokenizer/sampling/KV-cache/templating on raw NDArrays.

Revised spike: wire a CoreAILanguageModel into a LanguageModel-backed LanguageModelSession, then point apfel's existing OpenAI-compatible server at it. Confirm exactly how much the bridge gives for free. Still a separate repo, but a much shorter project than originally scoped.