Personally I'm not too fond of parenting poles directly to the feet. I used that setup a long time ago, and found myself counteranimating too much. Perhaps that was due to lack of experience; I see now in retrospect, that if I'd kept my poles down on the ground plane (instead of at knee height), they would have been a lot easier to manage in a foot-parented situation. Nonetheless, the impression I'm left with presently is that foot-based poles often cause more trouble than they solve.
It's even worse when the poles are on the pelvis.
To tell the truth, I've done a fair amount of animation with the poles in world space. It's not terribly exciting, but at least it offers predictable results and full control. It might even be the best way to go, if the character is staying in a small area.
But if the character covers much distance, then keeping his world-poles in reasonable proximity becomes a chore. In those cases, some automation is in order. So here is a setup I have found comfortable:
One locator (A) is constrained half way between the hip and foot. It is aim constrained towards the hip. Its up vector is determined by another locator, which is constrained between two more locators, which are placed ahead of the foot and hip (on pelvis), and parented to each, respectively.
(edit: that is, the "ahead" locator for the hip is parented to the pelvis. Parenting it to the hip joint would cause a dependency loop.)
So, (A) sits precisely between the hip and foot, aligned along that line, and in plane with a point half way between each of their "aheads", and with the same pitch as the pelvis.
The pole vector control object is parented to (A).
The nice thing about this setup, is that it is really, really dull. It is about as "neutral" a position as one can get; you can do almost anything with the foot and pelvis, and the leg stays in a very boring inbetween position. This is nice because it offers a clean starting point for animation, and is stable enough that it rarely requires counteranimation. The IK goal doesn't pitch with the foot, so it will never "swim" up through the leg, no matter what you do with the foot. And because the leg's rigging is diluted between input from the foot and input from the pelvis, wiggling the foot won't cause excessive motion in the leg.
In fact, the solution won't fail until you twist the foot a full 180 degrees out of alignment; that is, turn it all the way backwards - at which point the leg will finally flip. Such a situation is rare enough to be dismissable; but should it occur, you can always switch back to world space tangents for the duration of the event.
Hope that gives everybody something fun to chew on.
