For me, killer application is basic usable text input. I'd love to covertly send text messages, write down some thoughts and TODO items, while still maintaining eye contact and following discussion.
For it to be usable, it should not require sight and higher typing speeds should be achievable by training. This rules out virtual querty keyboards and predictive suggestions. I think glyph recognition would be usable, but also other gesture systems would work if they don't demand supervision by sight.
gesture input doesn't really need sight, especially if it is on a constrained surface (e.g. if the watch has a bezel that sticks out a bit, and a touch surface, you can easily "recalibrate" your absolute position on it every time you hit the borders). With adequate sensors it might be possible to do gestures in the air above it, or on the arm next to it.
Downside is that it requires training to learn the symbols, and of course without sight or other feedback you don't notice mistakes while doing it (same with keyboards though).
Not that it is perfect, but the Fleksy keyboard is pretty good about guessing what you meant to type, even if you miss several letters. I'm able to use it without looking, but I'd never trust it entirely.
edit: I forgot to mention Minuum. It's quite similar, but compresses all the keys into a single row. Consequently, it doesn't seem to be as accurate.
Maybe I'm weird but I don't look at my phone anymore when using a swype keyboard. I know where the keys are on the screen and can usually get about 90% word accuracy without looking. A quick glance at the final message when I'm done and off it goes.
I'd imagine the old Palm Graffiti (or similar) simplified glpyh recognition might be usable with training, but I'd still prefer some kind of physical references.
Maybe you could compensate a bit with some haptic feedback.
Yeah, seriously. Voice input is the primary way I use my watch. I use it for sending to FB & text messages, looking up the weather, setting reminders, etc.
Although I think a better input method would be voice (like the GoPro cameras) for most use cases. "Watch, send text to Mom...." Morse Code is just too slow and most humans can no longer send or copy it. It should be the input method of last resort.
For it to be usable, it should not require sight and higher typing speeds should be achievable by training. This rules out virtual querty keyboards and predictive suggestions. I think glyph recognition would be usable, but also other gesture systems would work if they don't demand supervision by sight.