Hi @deClare125,
Thank you for trying the extract-seq-segments pipeline and providing your outputs!
I think at the time I responded to that post about length variation, we did not implement both of the --p-min-seq-len and --p-max-seq-len options. When these options are used, there will be a noticeable drop in sequence length creep that we initially observed, and should not be as large of a problem as it used to be. 
Figuring out which sequence lengths are the appropriate bounds can vary depending on the amplicon segment you are after. For example, this can be problematic for highly variable length genes, as I discuss in some of our other forum tutorials. Anecdotally, I've found that using more iterations at high identity (after an initial 70-80% identity extraction) works better for high-variable length genes. But your mileage may vary.
Like you've done, I base my selection of possible gene segment lengths on the literature. Then I set the min & max values to that. Actually, I often add another 10-20% to the max length, as there is always length variation.
I personally always like to start with 70% identity for my first step, as you grab a far wider pool of reference sequences. You can see this in the first iteration of your "70>80>90" processing. You grab more references in your first iteration there, than all three iterations from your "90>94>97" processing.
This is kind of the conundrum, starting off with higher identities means you 'll need to perform more iterations, as your pool will grow more slowly, compared to lower identities. That is, at high identities you may need to run 6, 7, or 8 iterations, where as with lower identities you many only need ~3 or so. The only risk with applying too many low identity iterations is the potential to extract spurious segments. Especially, from the few larger length sequences that might be in your reference pool, e.g. chromosome fragments, etc....
Think of it this way, if I extract a sequence segment that has ~70% identity to my query (~30% divergent). Then that newly extracted segment becomes part of the new query pool. Then in my next iteration, that new query sequence might extract a fragment with a ~70% identity to it. Which means that, the newly extracted sequence from the 2nd iteration is now ~ 60% divergent from the initial query.
^^This is how the new query segment pool is intentionally grown. But you can also see how this can generate problems if too many low identity iterations occur. It appears that you observe some length variation, to a minor degree, in your "70>80>90" processing given the observed increase in 50 bp length. Which may not be a problem(?).
Historically, I've liked to start with 70% identity. For some reference data sets, I think simply performing one iteration at 70% identity followed by 90% identity worked well enough for my needs at the time. Some other gene segments I prepared required 2-3 initial iterations at 70%, because my initial segment pool was not very diverse to begin with.
I guess that is the key point here: knowing how broad, or not, your initial reference segment pool is. If the starting segment pool is already quite diverse, then you might be fine starting with ~80% identity or higher, followed by other more strict identities.
In a nut shell, my "go to" approach is to do the following (assuming I have nothing else to inform my decisions on identity or number of iteration cycles): 70 > ~85 > ~95, while using the --p-max-seq-len option.
To wrap up, given the information you provided, I'd trust your "70>80>90" processing. Especially, if you leverage the --p-max-seq-len option. Alternatively, you can combine that option with extracting segments at 80% > 90%.
Also, do not forget that you can assess the quality of your reference databases following the approach outlined here.
-I hope this helps!
-Cheers!
