This one finishes the speed work we started in v1.52 and cleans up the untranslated leftovers you reported while testing it. Every item below came out of real screenshots and real logs. Thank you.
NEW: Text recognition now streams. Instead of recognizing the whole screen before anything else happens, the app recognizes text in reading order and starts translating and drawing while the rest is still being read. Combined with v1.52's incremental display, the first translation on a dense visual-novel screen now appears in about one second in our measurements — down from over five seconds two versions ago. The final result on screen is character-for-character identical to before; only the waiting changed.
FIX: Untranslated leftovers got a cleanup. When the offline engine cannot handle a word it used to leave broken fragments like "Un ed" on screen — those lines now show the readable original instead, while good partial translations are kept. And manga sound effects and short quips ("Ba-dump", "Just kidding") finally translate properly: we added a 146-term onomatopoeia glossary for English, Chinese, Korean and Russian, and fixed a bug where the Japanese-to-English direction had no common glossary at all.
FIX: The high-quality AI tier's local server now always shuts down with the app. If the app was force-closed or crashed, the server could keep running in the background and hold about 1 GB of memory until you noticed it. Its lifetime is now enforced by the operating system itself — when the app dies, for any reason, the server dies with it.