Can any math-inclined people tell me if this works...?

I'm looking for a better way to seed mt_rand(), something that would yield more than the million seeds you get from (double)microtime()*1000003.

Some people have suggested something like:
PHP Code:
$seed=hexdec(substr(md5((double)microtime()*1000003),-8)) & 0x7fffffff; 
But I don't see how this is any more random, since there are only a million md5 hashes you can generate from microtime anyway; the pool is the same size.

In my application, a secondary (though less-reliable) source of randomness is that I don't know the exact last time the generator was seeded, because it hinges sometimes on user interaction. (Other times, reseeding happens in a loop). So my thought was that I could try this:
PHP Code:
$seed=hexdec(substr(md5((double)microtime()*1000003),-mt_rand(6,8))) & 0x7fffffff; 
which I think should leverage the (unknowable) last seed to choose the length of the md5 substring, which could now yield 6, 7 or 8 digit long hexidecimals (instead of 8 digits only).

My question is, does this increase the true number of possible random seeds to 3 million? Or is that an error in my reasoning?