fix(bookstack): suppress Qwen3 think blocks, fix citations, prevent flicker

[amrit110]

Root cause

Diagnosed by testing Qwen3_5-122B-A10B directly against the gateway. The model outputs reasoning inside a <think>...</think> block as regular text content — not a separate Anthropic thinking block. The gateway strips the opening <think> tag but passes </think> through as a plain text delta. This caused three problems:

1. Thinking text visible / flickering — thinking text streamed to the UI before being cleared by text_clear, which the client couldn't suppress fast enough
2. </think> tag appearing in the answer — the closing tag passed through to the final rendered response
3. Page citations with (page N) numbers — model saw "id": 61 in tool responses and used it as a page number; custom models ignore the system-prompt instruction less reliably than Claude

Fixes

src/aieng_bot/bookstack/agent.py — thinking-aware streaming

ask_stream now buffers text silently until </think> is found in the stream, then switches to real-time token streaming for everything after it. A skip_leading_nl flag drops the \n\n separator that Qwen3 emits as a separate chunk immediately after </think>.

• Qwen3 tool-use turns: thinking buffered → </think> → empty content → tool block → no text ever shown, no text_clear needed
• Qwen3 final answer turns: thinking buffered → </think> → answer streams live, token by token
• Non-thinking models (Claude, GPT): no </think> seen → full text buffered → burst-emitted as fast 20-char chunks after completion

bookstack_agent/ui/app/components/chat-page.tsx — RAF buffering (safety net)

text_chunk events are buffered in a ref and flushed once per animation frame. If text_clear arrives before the RAF fires, the buffer is discarded silently. Handles any edge case where a small amount of text is emitted before the model switches to a tool call.

src/aieng_bot/bookstack/tools.py — strip numeric IDs from get_page

Removed id, book_id, chapter_id, updated_at from the get_page tool response. Model now only sees name, markdown, url — can't accidentally cite a numeric ID as a page number.

src/aieng_bot/bookstack/prompts.py — stronger citation + no-preamble rules

Explicit ban on preamble phrases ("Now I have all the information…" etc.) and strengthened citation instruction: "NEVER include page numbers, numeric IDs, or any internal identifiers."

Verified live against Qwen3_5-122B-A10B

think text in chunks:   False
</think> tag in chunks: False
</think> in answer:     False

Tool-use turns emit only tool_use events (no thinking text). Final answer streams token-by-token from the first post-think token.

Test plan

• All 14 agent unit tests pass (pytest tests/bookstack/test_agent.py)
• Ask a multi-step question in the UI — no thinking text visible, no flicker
• Final answer streams smoothly token-by-token
• ## Sources section uses page titles + URLs, no (page N) annotation