v1.1.0

Highlights

• Startup splash screen (new SplashWindow): starts Ollama, picks a default model by detected VRAM, downloads it with a live progress bar, preloads it into GPU/VRAM, then opens the editor.
• Cursor V2-style dark UI: new Styles/Colors.axaml + Controls.axaml design tokens applied across the main window, editor, panels, file explorer, terminal, settings, and marketplace.
• Performance & memory pass across indexing, reversing, editor, and file-system subsystems.

Models / Provisioning

• VRAM-based default selection (threshold 20 GB):
  • High-VRAM → its_the_jisoo/qwen3.6-27B-uncensored-hauhaucs-aggressive:latest
  • Low-VRAM → fredrezones55/Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive:IQ2_M
• One-time provisioning tracked via new DefaultModelEnsured preference.
• Verify-after-pull: provisioning is marked complete only after the model is confirmed in ollama list (a completed stream is no longer treated as success).
• Self-healing: a missing model is re-downloaded even if the flag is set; stale/never-installed selections are cleared.
• Tag-tolerant matching normalizes an implicit :latest.
• GPU preload: warms the selected model (10-min keep-alive) so the first prompt has no cold-start; best-effort, never blocks launch.
• HttpClient.Timeout set to infinite so large pulls aren't interrupted.

Speed / Memory Optimizations

• RAG: parallel batch indexing, coalesced/cancelable per-file re-index, fewer chunking allocations.
• EmbeddingStore: O(1) chunk counting, pooled query buffers + SIMD cosine similarity.
• Editor: no longer copies the whole document on every keystroke (lazy dirty tracking).
• Reversing: lazy-loaded decompile/IL/WAT tabs, capped string-search results, off-thread throttled AI streaming, debug grids built once and updated by property, debounced ELF filter.
• File system: off-thread/incremental tree builds, pruned unbounded change tracking.
• Misc: StringBuilder in SearchTool (O(n²)→O(n)), capped autocomplete cache, off-UI-thread terminal shell resolution, leaner embedding deserialization, reduced Markdown render allocations.