fix(json_parser): strip <think> reasoning tags before parsing

[Cyriellewu]

Summary

parse_json_response() did not account for chain-of-thought <think>...</think> blocks that reasoning models (Qwen, DeepSeek, etc.) emit before the JSON payload. A brace inside the reasoning was picked up by the raw_decode scan and returned instead of the real object, silently yielding wrong or empty data (e.g. empty BookEngine proposal description / scope / rationale).

Fixes #673

Root cause

For input like:

<think>
1. Analyze user intent, maybe output {"draft": true}
</think>
{"title":"T","description":"D",...}

direct json.loads fails (the <think> prefix), so the raw_decode scan runs and stops at the first { — which is the {"draft": true} inside the reasoning — returning that instead of the real object.

Fix

Strip <think> blocks only on the direct-parse failure path, then retry. Key properties:

• No regression for valid JSON: direct json.loads runs before any stripping, so a valid payload whose string values legitimately contain the substring <think> (e.g. {"x":"a <think>b</think> c"}) is preserved exactly.
• Handles closed <think>...</think> blocks (one or many), case-insensitively.
• Drops an unclosed leading <think> prelude up to the first JSON opener.
• Returns the caller's fallback when the response is reasoning-only with no JSON payload.

Testing

New TestParseJsonResponseThinkTags class (8 cases) covering: think-before-json, brace-in-reasoning-does-not-leak, markdown-fenced payload, case-insensitivity, reasoning-only fallback, multiple blocks, literal <think> inside valid JSON preserved, and unclosed prelude.

• pytest tests/utils/ → 90 passed
• ruff check / ruff format --check → clean
• mypy deeptutor/utils/json_parser.py → clean
• Verified the brace-leak regression test fails on dev without the fix.