feat: extract JSON from thinking preamble in ThinkingAwareOpenAILike

[82deutschmark]

Problem

When LM Studio's Qwen 3.5-35B runs with thinking enabled, reasoning_content contains the full thinking preamble concatenated with the final JSON answer. PR #231 correctly falls back to reasoning_content when content is empty — but then passes the raw mixed text to Pydantic's model_validate_json, which fails because it sees 12,000+ chars of thinking prose before the JSON object.

This caused CreateWBSLevel3Task to fail mid-pipeline on real runs with Qwen 3.5-35B thinking mode.

Solution

Add _extract_json_from_thinking() with a 3-strategy approach:

1. Direct parse — fast path for short responses where reasoning_content is already valid JSON
2. </think> tag — extract content after the tag (DeepSeek and some other models emit this marker to cleanly separate thinking from output)
3. Right-to-left scan — find the rightmost { that yields a valid JSON parse (handles Qwen 3.5-35B which does not emit </think>)

If all strategies fail, the original text is returned unchanged (existing behavior — caller handles the failure).

Testing

Tested on Qwen 3.5-35B (Q4_K_M, LM Studio 0.3.x) with thinking always-on. The function correctly extracts JSON from reasoning_content strings of 12,000–19,000 chars. Fixed CreateWBSLevel3Task failures on Batman RICO v10 pipeline run (Hartford CT variant).

Compatibility

• Models without thinking tokens: reasoning_content is not present → code path unchanged
• Models that emit </think>: strategy 2 handles them efficiently
• Models that don't emit </think> (Qwen 3.5-35B via LM Studio): strategy 3 handles them
• If reasoning_content is already valid JSON (short non-thinking responses): strategy 1 returns immediately