Well, your "instantaneous bearing" thing matches my first feature
BTW I was still not clear enough for the second one : by "statistical computation", I meant it is not intended to be used during runtime but by the "convert" feature of gpsrecorder.

Your example does not seems to be a good one because if I understand it correctly, your pilot needs a vector from current position to a given known position (the field camp) to align heli's compass to the resulting vector.
So we definitely are talking about a third feature here !
Edit : Is it a feature request ?

Using accelerometer data will not make the bearing value really more accurate. I also have an example
A while ago, TomTom (a well-known GPS devices manufacturer for cars/motorbikes) used to integrate an accelerometer to make current position more accurate when driver entering into a tunnel for example.
After selling about 2/3 different models (can't remember exactly) with accelerometer, they removed it from next models partly because it was clearly not accurate at all and actually was even worse in a lot of situations.
My example tends to demonstrate that adding blind data to another blind data still gives blind data

But still, just to confirm : I will probably implement realtime bearing computation with some kind of settings to let user choose how much data he wants to use for computation. No reason not to do it. You can also propose a patch if you want to