Qx v0.6.117 — Focused QxAI conversation flow

• Move the active model control into the composer so it remains available in narrow windows. One compact picker now groups models by provider, prioritizes favorites, exposes vision, reasoning and context capabilities, searches long catalogues, and links directly to model management.
• Keep model changes scoped to future turns in the current conversation. Historical assistant replies retain their saved provider/model snapshots, unavailable custom selections remain visible, and model changes are locked while a response is streaming.
• Align the transcript and composer on one responsive 760px content axis. Streaming follows only while the reader is already at the bottom; detached reading stays stable, each conversation restores its own in-process scroll position, and a floating jump control resumes follow mode.
• Simplify expanded tool activity into one timeline instead of nested cards, preserving bounded scrolling, returned values and empty-output disclosure.
• Preserve desktop keyboard continuity: model search supports arrow/Home/End navigation, selection restores the composer, and the picker remains bounded and identifiable at 360px.

Validation: repository checks, TypeScript, production frontend build, Rust formatting/native compile, a local Apple Silicon .app bundle and synthetic browser interaction across English/Chinese, light/dark and narrow layouts. The tag workflow builds the Apple Silicon macOS app/DMG and Windows x64 NSIS installer. Release macOS artifacts remain ad-hoc signed and are not Developer ID notarized; live-provider and installed-app acceptance are not claimed.