feat(app): maximize RAG context and improve chat UX

[comppaz]

Summary

• Dynamic context budget: Computes available context space from model window size, conversation history, and reserves — scales memory retrieval to fill available space instead of using a fixed 8k limit
• Smarter filtering: Widens distance thresholds proportionally when budget allows, includes AI-generated summaries as overview/fallback for each memory, and subtracts page content from budget in extension chat
• Improved system prompt: Rewrites prompt for better grounding (avoids hallucination) and natural source attribution (no raw type tags)
• Chat UX fixes: Restores bullet point markers in markdown rendering, removes hover blur from assistant messages

Changed files

File	Change
backend/app/services/embeddings/filtering.py	New compute_context_budget(), budget scaling in filter/format functions, structured logging
backend/app/services/embeddings/__init__.py	Export compute_context_budget
backend/app/routes/chat.py	Compute and pass budget in _retrieve_context()
backend/app/native_messaging.py	Same budget pattern, subtract page content from budget
backend/app/services/ai/client.py	Rewritten system prompt
app/src/index.css	Add list-style-type to .chat-prose lists
app/src/components/ChatMessage.tsx	Remove glass.hover from assistant messages

Test plan

• Chat with Ollama (small model) — verify more context is retrieved and summaries appear in logs
• Chat with OpenRouter (large model) — verify budget scales up and more memories are included
• Check Context budget: and Filter stats: log lines for correct values
• Verify bullet points render correctly in chat responses
• Verify no blur on hover over assistant messages
• Test extension chat — verify page content is subtracted from budget
• Verify LLM no longer outputs raw type tags like [web] or [video]