feat: agentic loop with OGX vector search and conversation persistence

[franciscojavierarceo]

Summary

Adds the agentic loop to the gateway, integrating OGX for vector search (file_search tool) and wiring up conversation/response persistence from the storage CRUD layer (#33, #45).

• Agentic loop in handler.rs: multi-turn tool-call loop that proxies /v1/responses through vLLM, executes file_search tool calls via OGX, and feeds results back until the model completes or hits max_iterations
• VectorSearch trait in agentic-core: async trait with OGX implementation that calls the OGX vector search API
• Conversation persistence: stores conversation turns and response metadata via ConversationStore/ResponseStore on each loop iteration
• AppState refactor: unified shared state (ProxyState + stores + vector search + config) behind Arc<AppState>
• Integration test harness: mock vLLM and OGX servers for testing passthrough, single/multi-turn tool calls, max iteration limits, and conversation state hydration
• Recorded file_search cassette: replays real vLLM openai/gpt-oss-20b Responses API output for the file_search loop while still exercising live OGX search in make integration-test
• CI integration job: installs OGX and runs make integration-test
• Architecture docs: docs/architecture/index.md describing the gateway layers
• File search hardening: validates file_search call arguments, rejects empty vector_store_ids, forwards filters/max_num_results/ranking_options to OGX, rejects mixed file_search plus user function calls until that loop supports them explicitly, and accepts vLLM function calls with status: null

Test Plan

• cargo test
• cargo clippy --all-targets -- -D warnings
• cargo fmt -- --check
• SKIP=no-commit-to-branch /Users/farceo/.local/bin/uvx pre-commit run --all-files
• PATH=/Users/farceo/.local/bin:$PATH OGX_CMD='/Users/farceo/.local/bin/uv run --project /Users/farceo/dev/ogx --python 3.12 --extra starter ogx' make integration-test
• New integration tests cover:
  • Passthrough (no tools)
  • Single file_search tool call
  • Multi-turn tool call loop
  • Max iteration limit enforcement
  • Conversation state hydration across turns
  • Previous response chaining
  • Missing or empty file_search query arguments
  • Empty vector_store_ids
  • Forwarding file_search search options to OGX
  • Mixed file_search and user function call rejection
  • Replaying recorded vLLM file_search tool-call/final responses against live OGX search

[ashwing]

@franciscojavierarceo Is this still in-progress or ready for review? Happy to take a look when you're ready — especially interested in how the OGX VectorSearch trait maps to the dispatch interface we're building.

[franciscojavierarceo]

@ashwing I'm waiting for #46 to land to incorporate it and then mark as ready for review as I need the agentic loop and to update the Responses structs.