M3: Add local AI provider interface and Claude adapter

[dbrosio3]

Goal

Move local AI review behind a provider interface and migrate the current Claude path into the first provider adapter.

User value

Teams can keep the existing Claude review behavior while Pushgate gains a stable boundary for GitHub Copilot and future AI providers.

Implementation notes

• Define the provider contract: review input payload, timeout behavior, exit codes, failure categories, and structured output handed to Pushgate.
• Use the provider selection/config extension points frozen in the v2 schema work.
• Implement the Claude adapter first so Claude is no longer hard-coded in the core runner path.
• Keep provider invocation isolated from local deterministic checks.
• Keep GitHub Copilot implementation in its own provider-adapter task so this interface work does not silently stop at naming Copilot in config.

Risks / tradeoffs

• Provider CLIs vary in auth, streaming, JSON support, and model naming.
• Abstracting too early can overfit to imagined providers.

Suggested priority

P1

Milestone

M3: Local AI review

TBD considerations

• Whether a custom command or OpenAI-compatible adapter follows the Copilot adapter.
• BYOK timing and expected auth model for future providers.

Acceptance criteria

• Local AI invocation flows through a provider interface.
• Existing Claude review behavior flows through a Claude adapter instead of a core-hook hard-code.
• The provider contract is exercised with tests or stubs independently from deterministic checks.
• Provider failures respect the configured local AI mode.