Persist structured quiz mistakes on submission and expose them in a
new review page accessible from the utility sidebar.

Backend:
- New wrong_answers SQLite table (cascade on session delete) with
  add/list/count/update/delete store methods
- Hook POST /api/v1/sessions/{id}/quiz-results to also insert one row
  per incorrect answer, preserving the existing [Quiz Performance]
  text message consumed by the LLM context builder
- New GET /api/v1/wrong-answers (resolved filter + pagination),
  PATCH /{id} (resolved toggle), DELETE /{id}

Frontend:
- New /wrong-answers page under the (utility) route group with
  unresolved/all filter, mark-as-resolved, delete, and session link
- Sidebar nav entry using the NotebookPen icon
- English and Chinese i18n strings for all new UI text

Tests:
- 8 store tests covering filter, cascade, resolved toggle, delete
- 6 router tests including a regression test that record_quiz_results
  still writes the [Quiz Performance] text message

This PR fixes two related issues:

1. Qwen vLLM 'System message must be at the beginning' error:
   - deeptutor/agents/chat/agentic_pipeline.py: Merge system_prompt and memory_context into single system message, filter out system messages from conversation_history
   - deeptutor/services/session/context_builder.py: Filter out system messages from DB since summary is already added as system
   - deeptutor/agents/chat/chat_agent.py: Merge system prompt and RAG context into single system message

2. History reference (@reference) empty context issue:
   - deeptutor/services/session/turn_runtime.py: Add fallback to raw record content when NotebookAnalysisAgent.analyze() returns empty string

Benefits:
- Fixes Qwen vLLM compatibility issue completely
- Enables history reference feature to work correctly
- Reduces token usage by eliminating duplicate system messages
- Does not break any functionality - system context from DB is already in compressed_summary
- Cleaner message structure with exactly one system message at position 0

Testing:
- Verified with Qwen model via vLLM - no more template errors
- Chat with memory files (PROFILE.md/SUMMARY.md) works correctly
- Chat with conversation history works correctly
- History reference (@reference) now works correctly
- Conversation continuation works without errors
- No regression with other models (GPT-4, etc.)

Closes #294

…ty-context

fix: merge system messages and add history reference fallback

Record the embedding model and dimension used to index each knowledge base
in kb_config.json. On subsequent loads, compare the stored fingerprint
against the currently configured embedding model and flag mismatched KBs
with needs_reindex=True and embedding_mismatch=True.

When a mismatched KB is searched, a warning is included in the search
result so callers know search quality may be degraded and re-indexing
is recommended.

This prevents silent search quality degradation when users switch
embedding models (e.g., from text-embedding-3-large to nomic-embed-text)
without re-indexing their existing knowledge bases.

…ch-detection

feat: detect embedding model mismatch in knowledge bases

feat: add wrong-answer note for cross-session quiz review

Question Notebook with bookmarks & categories, Mermaid diagram support
in Visualize, embedding mismatch detection refinement, system message
merging for Qwen/vLLM, LM Studio & llama.cpp providers, Glass theme,
reporting agent retry resilience, and docs migration.

Made-with: Cursor

Made-with: Cursor

# Conflicts:
#	README.md

Removed news section detailing past updates and milestones.