I see what you are saying and I'm all for wrapping existing CGIs in adaptors, when possible, but I disagree on principle. This is like insisting that mixing hash access with Moousueusue-based code isn't a problem. Yes, it works, yes it's probably future proof in many cases and somewhat transparent. No, it's still not correct and mixing the two approaches increases the cognitive load/overhead and potential bugs. You insist your future devs be adept at a dead-end and the idiosyncrasies that made it such instead of embracing ideas that span several high-level languages and foster tremendous re-use and cross-pollination. Probably the single best app engine for PSGI stuff is Python/C: uWSGI. It supports PSGI because it was easy.
I feel pretty credible here. I still write CGI.pm based code, I even think, to the appalled reactions of many monks, that it's a perfectly cromulent library for HTML formatting. I'm not the modern for modern's sake monk. I'm the modern for the love of pity why am I still working 90% of the time on code this 1995-awful monk. I would never foist any of this on someone at $work and the times when I end up revisiting this kind of stuff in my own projects I only have regret that I didn't see it might grow/outlast the shoddy/quick path I gambled.
You also gloss the fact that once one knows enough to be able to wrap CGIs to PSGI, which isn't always easy or possible, one knows better than to write the stuff that needs to be wrapped in the first place.