Build 26358951

NEW: Text detection is lighter on every capture. The normalization step that runs before text detection now uses a precomputed table - same math, same results (verified bit-for-bit identical recognition on 3,000+ screenshots), just faster: roughly 50 ms less per scan on a 1080p screen, with the biggest gains on CPU setups and small capture regions.

NEW: A safety net for local AI models (Ollama / LM Studio). Very small models (around 0.5B-2B) sometimes return broken output - JSON-like fragments, echoed instructions, or a long ramble for a short label. Such lines are now caught and instantly re-translated with the built-in offline engine, with a one-time notice. Normal models are unaffected (calibrated to zero false alarms on 13,000+ samples), and list-shaped answers from mid-size models are now unwrapped so you keep the model's own translation instead of seeing brackets on screen.

NEW: A smoother first run. The setup wizard now ends with a "your first minute" card (the start hotkey, where the control bar appears, and how long the first translation takes), and on your very first launch the control bar's location is pointed out once - never again after that.

Under the hood: all three changes sit behind safety valves and the full 11,800+ test gate passed before shipping. If anything behaves oddly, please report it on the community hub - recent threads directly shaped this update.