As a note on Control structures, I have found that I actually have a harder time reading the control structure based code than the code with branches. It seems to me that the control strucutres for problems of this size don't lend anything to readability. Rather, the density suddenly changes between fairly low to failure high as I move from normal APL to the control structures, and that usually makes my code read like Swiss cheese. For problems of this size, I find that using labels with branches seems to keep the density fairly constant while lending me hints in the code that I don't get from the control structures. Admittednly, if I through in more comments, I could probably get the same, but then the density goes down even further.
⍝ Compute the sum of the decimal digits of 2*1000
R←1 ⋄ I←0
ADD:R←2×R ⋄ C←⌊R÷10
→(~1↑C)/CARRY ⋄ R←0,R ⋄ C←0,C
I←I+1 ⋄ →(I<POW)/ADD