All pseudo random generators have a finite cycle length.
Let's say, this deliberately short sequence.
1 5 9 3 2 8 4 7 6
So you call srand(0) and you get 1 5 9 3 2 8 4 7 6
If you did srand(4), you'd get 2 8 4 7 6 1 5 9 3
But if you keep changing the seed too often, what you do end up doing is risking landing on some sub-sequence you've already used. This most definitely isn't random.
Lets say 1 5 9 3 2, then you call srand() again, and get 9 3 2 8
Sure if you're only getting 10000 entries out of a potential sequence of 2^32, then the chance is minimal (but not zero).
Any decent run-length from a good random number generator (which rand() isn't) is going to be just as good as any other sequence. Calling srand() multiple times at BEST does nothing useful, and has a number of potential pit-falls which are not easy to predict (or detect).
> here's what I got my re-seeding srand() everytime with the output of the MT19937 prng:
Why are you messing with rand() when you've got MT to play with?
The standard library rand()/srand() functions are essentially useless for serious statistical or cryptographic work.