Build 26296278

A detailed review made a fair point: on a CPU-only machine, one translation could take "a few seconds". So we profiled where every millisecond goes and shipped the two biggest wins - one automatic, one as an opt-in switch - plus clearer guidance for exclusive-fullscreen games.

NEW: Text recognition is faster on CPU - automatically. The recognizer paid a hidden cost every time it met a text shape it had not seen before (25-50% slower on first sight); it now reuses its memory layout and that cost is gone. On our CPU-only test rig a dialog box appears in 0.69s instead of 0.81s, and a very text-heavy full screen in 4.6s instead of 5.2s - with recognition results verified character-for-character identical on 3,000+ screenshots. Costs a bit more RAM (about 0.4 GB) on CPU; GPU setups are unchanged.
NEW: Optional speed mode for offline translation (Settings, off by default). It makes the offline English-to-Japanese engine translate 35-65% faster by considering fewer candidate phrasings - the meaning stays, but some sentences read differently. That wording trade-off is exactly why it ships off by default: flip it on if speed matters more to you than phrasing.
NEW: Clearer help for exclusive-fullscreen games. Windows does not let any app draw on top of true exclusive fullscreen - no overlay can. But the translation panel (a separate window you can put on a second monitor) works fine there, and the in-app notice and the first-run wizard now tell you so, instead of leaving you guessing.

Honest notes: the recognition speed-up shows mainly on CPU-only machines and on text the app is seeing for the first time that session - on GPU little changes. The speed mode currently covers the offline English-to-Japanese engine only, and we will not make it the default while it changes wording. If translation still feels slow on your machine, post your CPU model on the discussions board - exact reports like that are what made this update possible.