Add experimental AI/LLM provider abstraction with local Ollama support

[RedRockerSE]

Summary

This PR introduces an experimental, pluggable LLM provider abstraction to YAFT with a working local backend (Ollama and OpenAI-compatible servers) and full CLI/configuration wiring. AI features are disabled by default and default to local-only execution to ensure forensic case data never leaves the examiner's machine unless explicitly configured.

Key Changes

• AI Configuration System (src/yaft/core/ai_config.py): Pydantic models for AI provider configuration with validation. Supports Ollama (default, local), Anthropic, and OpenAI (placeholders, not yet implemented). Configuration loaded from config/ai.toml with sensible defaults.

• Provider Factory & Base Types (src/yaft/ai/factory.py, src/yaft/ai/providers/base.py): Factory function to build configured providers. Base protocol defines is_available() and summarize() interface.

• OpenAI-Compatible Provider (src/yaft/ai/providers/openai_compatible.py): Working implementation for Ollama, llama.cpp, LM Studio, LocalAI, and vLLM. Includes minimal audit logging (timestamps, latency, success/failure only—never logs prompt/response content).

• Exception Hierarchy (src/yaft/ai/exceptions.py): Clear error types for disabled features, unimplemented providers, and provider failures.

• CLI Commands (src/yaft/cli.py):

  • ai-configure: Configure provider, model, base URL, and enable/disable AI features
  • ai-status: Check configuration and test connectivity (no case data sent)

• CoreAPI Integration (src/yaft/core/api.py):

  • _load_ai_config(): Loads and merges TOML configuration with defaults
  • get_llm_provider(): Public API to retrieve configured provider instance
  • API documentation categorization for AI/LLM methods

• Configuration File (config/ai.toml): Example configuration with detailed comments explaining the local-first, opt-in design.

• Comprehensive Tests (tests/test_ai_config.py, tests/test_ai_providers.py):

  • Configuration validation, defaults, and TOML loading
  • Provider availability checks and summarization with mocked HTTP
  • Error paths (disabled features, unimplemented providers)
  • Audit log verification

• Documentation (docs/YaFT.md, README.md): Usage examples and design rationale.

Notable Implementation Details

• Local-first design: Ollama is the default provider; cloud providers (Anthropic, OpenAI) have config placeholders but raise AIProviderNotImplementedError if selected, with a clear message directing users to the local backend.
• Audit logging: Minimal, privacy-preserving logs record only metadata (timestamp, provider, model, latency, success flag)—never prompt or response content.
• Graceful degradation: Invalid AI config files fall back to defaults with a warning rather than crashing.
• No case data sent yet: This is plumbing only; no plugin currently sends forensic data to a model.

https://claude.ai/code/session_018MrbNYDqa5eLQdKyvYuTjD

[RedRockerSE]

Experimental AI/ML pluggable provider