Define a provider-neutral AI client and capability contract #35

Goal

Replace LM-Studio-specific assumptions at the package boundary with a Lucid-owned provider contract.

Depends on

• Measure AI latency, token use, repairs, and edit isolation #34 for baseline measurements.

Scope

Define the smallest interface needed by Lucid's existing lanes:

• streaming and cancellation;
• structured-output mode;
• model discovery;
• usage reporting;
• optional reasoning controls;
• optional local model load/status;
• optional custom base URL.

Run a compatibility spike against the current Input, Command, and Advisor pipelines. Evaluate the Vercel AI SDK as an implementation detail, but keep Lucid's own interface so the app is not coupled to one SDK.

Acceptance criteria

• Existing pipelines depend on the Lucid interface rather than an LM Studio concrete type.
• Capability differences are explicit and testable.
• The spike documents whether the SDK preserves current streaming, cancellation, and JSON-schema behavior.
• No provider settings UI or new provider adapter is added here.