I also didn't notice when originally replying, but maybe it's worth tacking on:
> ... even if the NIST curves are not backdoored mathematically, they're designed in a way to strongly coerce implementors towards a route which leaves side-channels open ...
[emphasis mine]
The NIST curves are short Weierstrass curves with a=-3. I believe that before 2007 (Edwards), these were generally considered the best, fastest, easiest-to-implement elliptic curves for general operations. Yes, Montgomery curves are faster and simpler for ECDH, at the cost of having a cofactor, but not for signatures.
So it's worth noting that NIST/NSA/Certicom did not choose this family of curves as a trap for implementers. It was the best choice at the time, but Montgomery or Edwards curves are arguably a better choice now.
That said, I still think the finite-field inversion in ECDSA signatures (instead of doing vanilla Schnorr) is dumb and risky. (Bernstein's deterministic signature improvement is even better, and would probably have prevented the PS3 hack, but I can't blame NIST for this because it was not known at the time. I think you're allowed to do it with the NIST curves too nowadays.)
FIPS 182 gives a non-constant-time implementation of point multiplication with the note "The algorithm given below is for reference purposes. Other (constant time) algorithms that produce an equivalent result may be used." which should be a MUST not a MAY, and in my opinion the non-constant-time one should never have been in the standard in the first place because that (and the accompanying note's wording) encourages people to go with it.
Checking for invalid curve points is also a solved problem in theory, but as far as I know not always implemented correctly in practice.