Email from Noam Elkies:
{{{

About your concluding parenthetical comment of page 16: the snarky reply
would be "what part of lambda don't you understand?" :-) but seriously
is there anything it might help for me to explain?  The idea is:

 \lambda is a rational coordinate of X(2), which is a normal cover
 of X(1) with Galois group (P)SL_2(Z/2Z) =~= S_3.

 If E has good reduction outside a finite set S of primes of K,
 then the S_3 extension generated by \lambda generates an S_3
 (or smaller) extension K'/K unramified outside 2S.  Since S_3 is
 solvable, all such extensions can be listed effectively (and easily
 with existing tools, according to Cremona).

 Given K', let S' be the set of primes of K' above S.
 Since both \ lambda and \lambda-1 are modular units,
 it follows that both are 2S'-units of K'.  Thus
 \lambda + (1-\lambda) = 1  is a solution of the 2S'-unit
 equation in K'.  All such solutions can be obtained effectively
 (albeit not as easily as the previous step).  It should help
 in practice that the Galois conjugates of \lambda are contained
 in its images 1-\lambda, 1/\lambda, etc. under Gal(X_2/X_1) =~= S_3.
 This is what's needed for a solution of the 2S'-unit equation in K'
 to give rise to a value of j in K.

 For each of the resulting values of j, it remains to find all twists
 (usually quadratic, but also quartic or sextic if j=1728 or j=0)
 that yield a curve with good reduction at S.  This again is a finite
 and (again modulo existing number-field tools) easy computation.

}}}