I would have thought DNA composition would have played some role (for example, if 10% of the genome came from different organisms, with a different codon bias, that might result in a non-50:50 ratio even if that 10% had a random insertion direction).

I like this idea, because it would seem to follow that if codon bias affected the 50:50 ratio, then it would explain why organisms that do horizontal gene transfer break Chargraff's second parity rule, and why organelles - genomes highly dependant and interacting with an "external" genome - are the worst offenders. In short, my theory would be that the more your genome interacts with others, the less optimised your codon/nucleoside ratios are, the further you deviate from 50:50 even in the presence of random insertion direction. So is matted's random-insertion-direction sufficient to give a 50:50 ratio?

The code below tests random 50:50 reverse-complimenting of DNA, of user-editable codon usage:

import string
import random
def rc(DNA):
    return DNA.translate(string.maketrans('ACGT','TGCA'))[::-1]

genome = [
    'AAA' for _ in xrange(1000)      ]+[
    'AAT' for _ in xrange(10000)     ]+[
    'ATT' for _ in xrange(100000)    ]+[
    'TTT' for _ in xrange(1000000)   ]+[
    'CCC' for _ in xrange(1000000)   ]+[
    'CCG' for _ in xrange(100000)    ]+[
    'CGG' for _ in xrange(10000)     ]+[
    'GGG' for _ in xrange(1000)      ]

before = ''.join(genome)
A,C,G,T = before.count('A'),before.count('C'),before.count('G'),before.count('T')
print 'Before random insertion:'
print float(A)/(A+T), float(C)/(C+G)

for position,codon in enumerate(genome):
    if random.randint(1,100) > 50:   # Flip codon 50%
        genome[position] = rc(codon) #  of the time

after = ''.join(genome)
A,C,G,T = after.count('A'),after.count('C'),after.count('G'),after.count('T')
print 'After random insertion:'
print float(A)/(A+T), float(C)/(C+G)

Sample output:

Before random insertion:
0.036903690369 0.963096309631

After random insertion:
0.500247890819 0.499687968797

So as you can see, even if you start with a highly 'deviant' genome, you still end up with near 50:50 ratios if every codon has a 50% chance of being reverse-complimented. Therefore, I would say my idea of codon usage is false, and matted's answer is correct and sufficient to explain Chargraff's second rule :)

It also stops working when the codon number get small - obviously, because if you only had a single, highly biased gene, it doesnt matter how you orientate it, it's going to result in a highly-biased nucleoside composition. In this test i'm flipping every codon, but in nature you can only flip whole genes. So this suggests that the genomes that break the second rule are probably due to the number and size of the genes, or non-random insertion direction, and bad luck.