Add a Core AI backend and a Private Cloud Compute model

[james-333i]

Follow-up to the discussion in #210. This adds two Foundation Models protocol backends behind the OS 27 LanguageModelSession(model:) construction.

• CoreAILanguageModel (new CoreAI trait, wired like the MLX and Llama traits) runs Core AI .aimodel bundles exported with apple/coreai-models recipes. Verified on device (iPhone 17 Pro, iOS 27 beta) with Qwen3 and SmolLM2 exports: streaming, multi-turn transcripts, and the executor's tool-calling path.
• PrivateCloudComputeLanguageModel wraps the OS 27 FoundationModels.PrivateCloudComputeLanguageModel, exposing availability, quota, and vision and tool capabilities. It requires the com.apple.developer.private-cloud-compute entitlement, which Apple grants on request for apps under two million users. Verified on device, including tool calling.

To share the session plumbing, the FM adapter's respond and streamResponse bodies move verbatim into file-scope fmRespond and fmStreamResponse helpers parameterized by an async session factory. All three backends delegate to them. Streamed and returned text also drops leading newlines, which OS 27 executors emit when they strip an empty reasoning block from the output.

Draft until the dependency is upstream. The CoreAI trait currently depends on a fork of apple/coreai-models that carries two fixes offered upstream. Links are tracked in #210.

• lower-deployment-and-simulator: availability annotations plus a simulator canImport guard (their OpenAILanguageModel + SystemLanguageModel tools calling support #49)
• fix-marker-probe-unk-fallback: marker detection is broken on unk-token tokenizers

Assuming those get approved (or similar fixes are issued by Apple), the dependency flips to apple/coreai-models and this becomes mergeable. This branch includes the fix from the standalone duplicated-trailing-prompt PR and will be rebased once that merges.

[mattt]

Thanks, @james-333i. I've marked this as a draft until the two apple/coreai-models fixes land, as your description says. I'll read through the shape in the meantime.

[james-333i]

@mattt Since filing this I have found a method that removes the dependency on apple/coreai-models entirely, so I would like to restructure the PR before you spend time on the current one.

iOS 27's FoundationModels defines a public LanguageModel protocol, and LanguageModelSession.init(model:) accepts any conformer. The Core AI models in apple/coreai-models already conform to it and so does Private Cloud Compute. Reading this PR's adapter with that in mind, the only Core AI specific lines are the construction of the model. Everything else hands it to FoundationModels.LanguageModelSession(model:) through the shared bridge.

Proposed integration: a generic wrapper for any FoundationModels.LanguageModel conformer, taking an async factory so heavy loading stays lazy and the consumer owns load and unload. It depends only on the FoundationModels SDK framework behind @available(iOS 27, macOS 27, *). No CoreAI trait, no coreai-models dependency, no fork branch, nothing hosted. Core AI becomes a documented recipe: add apple/coreai-models to your app (the source package if your deployment floor is 27, or a self-built xcframework if you support earlier OS versions, which I have verified works with library evolution and CoreAI weak-linked), construct the model, and hand it to the wrapper. The PCC adapter and the shared bridge refactor stay as they are.

That makes this mergeable without waiting on anything from Apple, and it covers any future model that adopts Apple's protocol. If that works for you I will push the restructured branch.