Thought in jumping here, hope these suggestions help.
Anyway, at this stage in our field is pretty hard to know a priory the effect size of the different parts or studies and more meta-analyses, everything matters, from sample collection to wet lab and sequence processing. However, we are getting better at this, so before continuing I would suggest checking these 2 papers: Meta-analyses of studies of the human microbiota, and Tiny microbes, enormous impacts: what matters in gut microbiome studies?. At this stage, my suggestion for meta-analyses will be to not change anything in the sequencing processing (and more if they come from different primers); thus, suggest using deblur (and use fragment-insertion for the tree) or close reference with the same length. Note that this will not assure that you will not have a separation due to primer but you might not; really depends on the effect sizes of your datasets.
Now, answering your question and why IMOO using different lengths is complicated. Imagine that you have 5 denoised sequences:
Perhaps it will make sense to merge 2 and 3 but how will you merge 1 with 4 and 5? Furthermore, how can you assure there is no other possibilities in nature for those sequences? For example: for 2 and 3 that: AACTA, AACTC, will never occur and is fine to merge 2-3.
My 2 pesos.