Three of you built this update: one player asked for per-game settings, one filmed his chat with a phone camera until we could see exactly where names got mangled, and one attached the error log that explained his aborted translations. This is what that kind of feedback turns into.
NEW: Per-game profiles. Under Options you can now save the current setup — capture regions, source and target language, engine and model — for the game you are playing. Next time you start translating with that game in front, its profile applies automatically; switch games mid-session and the regions follow immediately, while language and engine changes politely wait for your next start. This also makes multiple capture regions survive a restart for the first time. Nothing changes until you save your first profile.
FIX: Chat lines stay separate, and player names stay out of the translation. A player filmed exactly what went wrong: after a long message, the next chat lines glued themselves to it, and once glued, the name-exclusion could not see where a new message began — so names ended up translated. A line that starts like a chat message (tag, name, colon) now always begins its own block. Calibrated against eleven thousand real text lines with zero wrong splits.
NEW: Ollama got tougher in three ways. If the model spirals into repetition and the server aborts, the affected batch is now translated by the built-in offline engine instead of showing the original — clearly logged, never silent. A tuned anti-repetition setting makes those spirals rarer in the first place. And a new guard catches a small-model quirk where translations came back wrapped in braces and quotes, and unwraps them before they reach your screen.
