Are you sure? In my tests moiree looked more or less identical between 3x5 readout (1736*6?? (16:9) horizontally stretched by 1.67) in MLV and standard H264(1080p). For e.g. the 5D2 I agree with what you wrote. Standard H264(1080p) on EOSM looks way worse than e.g. on 7D, but sensor has the same resolution.
Yes, I'm sure. EOS M uses 3x5 readout in standby, at 30 fps, regardless of settings in Canon menu (this is unusual about it). However, when recording H.264 at 1080p, it switches to 3x3, where it behaves like all other ML-enabled Canons other than 5D3 (i.e. column binning and line skipping).
Now, how bad the aliasing actually looks like, depends on both how strong the high-frequency details are (how sharp the focus actually is), and - to a lesser extent - on the image processing pipeline (which may be different on different camera models; currently it's a black box to me).
Feel free to do a controlled test and prove me wrong. You will need a static scene with lots of aliasing (e.g. a resolution chart or any kind of fine print), same lens, same aperture, same focus point (triple-check this one!), same exposure, same picture style settings... you know the story. You could also capture a LiveView frame in x5 and in 1080p without moving the camera (e.g. by tweaking
this script); from that pair of images, I can identify the pixel binning pattern (but I don't expect any surprises here).