fix(llm): strip markdown fences from LLM response before instructor parsing (#159)

[relunctance]

Fix: LLM JSON parsing fails when response has markdown fences (#159)

Problem: Entity Name Extraction fails because qwen3 (and other reasoning models) wrap JSON responses in json ... fences. instructor's pydantic validation rejects the raw markdown-wrapped content, causing infinite retry loops with exponential backoff.

Error:

Root cause: instructor v1.x parses the raw message.content string directly without stripping markdown fences.

Solution: Wrap the instructor async client in _MarkdownStrippingClient, a thin proxy that:

1. Calls the original instructor.aclient.chat.completions.create()
2. Strips json ... fences from response.choices[0].message.content
3. Strips invisible control characters (0x00-0x1f) that can also break JSON parsing
4. Returns the cleaned response to instructor's response_model parser

Applies to all OpenAI-compatible backends (qwen3, gpt-4, etc.) via the unified OpenAIAdapter.

Closes #159