Frankly I'm no expert either.

Are you sure about that?
If I read [1] correctly then there are ARM devices out there that have no FPU at all and armel works on them. So this either means that armel doesn't use the FPU at all (and instead emulates FP operations using integer functions) or can detect if there is a FPU and make use of it. As I understand it the latter would require quite some overhead during runtime which makes me believe the former is true.
Under this assumption my understanding was that armhf requires an FPU which will then run FP operations much faster than armel using its integer emulation. This might also explain armhf's poor FP EMULATION test result since it simply wouldn't need to be optimized for that.
Anybody, please correct me if I'm wrong!

[1] http://wiki.debian.org/ArmHardFloatPort#Rationale

btw: As you can see I ran the armel test under Squeeze and the armhf test under Wheezy. I will repeat the armel test under Wheezy to see if the "illegal instruction" error is architecture or release related.

Edit:
I'd welcome any suggestion for a real world FP benchmark! I guess some of the (de)compression algorithms of different archive types make heavy use if it. But I don't have the knowledge which one would be suitable.