Here's my 2 cents.
I have done both on-the-fly and post matrixes and use the latter when I really want to ensure I get it right.
Typically I'll run a JB3 from the board and a JB3 from a mic>UA-5 optical out.
The mic verson is always faster than the board, and without exception the difference is 0.0037 seconds per minute.
I have found that time stretching the mic recording creates minor artifacts and is not really desireable.
After synching the start of both recordings in Cool Edit Pro, I create a 0.0037 second cut and paste segment of silence and quickly run through pasting it in the mic recording at 1 minute, 2 minute, 3 minute, etc.
The mixdown sounds perfectly synched every time.
That method may have its critics but I swear by it.