Thanks! I got a step further.
Collada indeed gives the bind-pose matrices in world-space coordinates. But, I found out the actual per-frame animation matrices are in joint-space coordinates! So, I exported one with the matrix of the first frame, and it looks like this:
pro-rsoft.com/screens/goblin-frame0.png
Looks like the first frame is set to the bind pose this time or so?
However, the rest of the matrices still don’t look right yet:
pro-rsoft.com/screens/goblin-weird.avi
But at least they are not as much out of proportion as they were before.