If you're interested, here is a (still in-progress) simulator I wrote where you can try out Define-Combine on a simple grid. https://mpalmer.shinyapps.io/DefineCombine/
I gerrymandered during the define phase using classic packing / cracking strategies, such that I had 8 majority-B districts (2:3) and 2 majority-A districts (4:1 and 5:0), and unsurprisingly, the only districts I was able to combine that were majority-A were those that included the packed districts.
If the overall split was, say, 27:23 instead of 25:25 such that we could define 9 majority-B districts in the define phase, then I would only have been able to define a single majority-A district in the combine phase.
(And yes, all of these gerrymandered districts would be considered safe B seats, as one would expect with a 20% margin)
There are also potentially issues if the packed districts are geographically clustered -- we see this a lot in states with a single predominant urban center (e.g. Kansas, Minnesota, Kentucky). In those cases, you might be forced to combine multiple packed districts due to pathological maps. For instance, consider a map where a Democratic bastion is districted into concentric rings -- that satisfies the contiguity requirement, yet only the outermost district abuts any Democratic-minority districts.
If instead we had 9 population nodes we could assign to each Define district, then we might be able to draw nine 4:5 districts in favor of B, and a single 9:0 district in favor of A. In that case, A cannot recover any of the unfairness that B introduced.
It doesn't matter at the abstraction of proportional representation, but it potentially matters quite a bit when you get into the nitty-gritty of actual elections.
Please do not get discouraged by anyone or anything. We need more of this.