Performance improvements + wake-word listening fix

[CatSchoko]

Performance:

• Smaller screenshot (1280→1024px) — less upload/tokens
• Shorter per-app history (20→10 messages)
• Ollama: keep_alive + full GPU offload + thread count
• Pooled httpx connections for Ollama/LM Studio (no per-request handshake)
• TTS prefetch — next sentence synthesizes while current one plays

Fixes:

• Wake-word capture now ends as soon as the user stops talking, instead of always waiting out a fixed 10s timeout
• Softened the "MULTI-STEP TASKS" prompt rule — smaller local models were over-applying "Say 'next' when ready" to every response

Tested locally with LM Studio (qwen2.5-vl-7b) + whisper.cpp medium.

[sonarqubecloud]

Quality Gate passed

Issues
1 New issue
0 Accepted issues

Measures
0 Security Hotspots
0.0% Coverage on New Code
0.0% Duplication on New Code

See analysis details on SonarQube Cloud