If you try "raspivid -w 1280 -h 720 -o foo.h264 -md 4" and "raspivid -w 1280 -h 720 -o foo.h264 -md 5" they should have produced the same result (one cropping in the ISP, and the other on the sensor), but they don't.
I'm not intending to fix this as it would mean that anyone using that mode would suddenly get a different field of view and recording things they weren't expecting.
However it prompted me to look closer at the registers, and it's shown up that if I correct another register so we actually read out a 16:9 image, then I can get 57fps out of it instead of the old 49fps
The changes are pending code review, but should be out in the next few days for those who are prepared to use rpi-update and accept the risks.
