The SpectreRF noise algorithms are a good deal more sophisticated than you give them credit for, and are perfectly capable of handling the delay. If you are interested in learning more about them, try reading the following paper:
Ricardo Telichevesky, Kenneth S. Kundert, Jacob K. White.
Efficient AC and noise analysis of two-tone RF circuits.
Proceedings of the 33rd Design Automation Conference, June 1996.
I believe by reading it you will find that SpectreRF works much differently than you think.
So then the question becomes, if SpectreRF noise analysis is accurate on dividers with delay, why did you get results that were far from measurement? I believe it is because you used SpectreRF to measure the time-average noise rather than the phase noise at the threshold crossings. This is actually the point of slides 36 and 36 in
http://www.designers-guide.com/Theory/cyclo-preso.pdf. If you use the time-average noise, which SpectreRF produces by default, you include the noise produced by the output stage even when the output is not switching. This is why the noise from the output stage appears to dominate over the noise from other stages.
I recommend that you follow the procedure described in
http://www.designers-guide.com/Analysis/PLLjitter.pdf.
-Ken