Detection: agent frameworks (pydantic-ai / LangChain / LlamaIndex) — resolve provider from the model string, treat .run()/.invoke() as metered

[AndresL230]

Part of #142.

Summary

The primary LLM path in modern Python apps is an agent abstraction, not a direct SDK call. In sapling, every AI feature goes through pydantic-ai: classifier_agent = Agent("google-gla:gemini-2.5-flash", ...) then classifier_agent.run(...) (e.g. routes/documents.py:698, all of agents/*). None are detected — the provider (Gemini) is hidden inside the model string, and agent.run() looks like a local method call.

Root cause

• agent = Agent(...) binds the var to package pydantic_ai, which is not in PACKAGE_TO_PROVIDER and is not a registered provider → resolveProvider → dropped at core-scanner.ts:269. There is no concept of "the provider is named in a constructor argument string."
• Irony: src/scanner/python-waste-detector.ts:18-30 already enumerates langchain*, llama_index*, google.generativeai for N+1 loop detection — that awareness never reaches endpoint detection.

Proposed fix

• Recognize agent/LLM construction for the common frameworks: pydantic-ai Agent(model, ...); LangChain ChatOpenAI / ChatGoogleGenerativeAI / ChatAnthropic / init_chat_model(...); LlamaIndex LLM ctors.
• Parse the model string → provider: "google-gla:" / "google-vertex:" → gemini/vertex; "openai:" / bare gpt-* → openai; "anthropic:" / claude-* → anthropic; etc. Bind the agent/LLM var to the resolved provider so the normal resolution path picks up the call.
• Treat .run() / .run_sync() / .invoke() / .ainvoke() / .complete() / .chat() on a bound agent as a metered call. Note the tool-loop cost multiplier — orchestrators fire N model round-trips per call.
• If the model string is dynamic/unreadable → emit an outbound LLM call of unknown provider (depends on #D), not nothing.

Evidence

Manual scan of sapling found ~26 Gemini sites, the majority via pydantic-ai agent.run() across agents/document.py, agents/chat_tutor.py, agents/quiz.py, routes/documents.py:698/718/721/725, routes/learn.py:499, routes/notes.py:204/223/252. Extension detected 0.

Acceptance criteria

• Agent("google-gla:gemini-2.5-flash") + .run() → provider gemini.
• LangChain ChatOpenAI(...).invoke(...) → provider openai.
• A dynamic model string still emits an outbound (unknown) LLM call, not nothing.