To test this, we implemented fallback-to-HTTPS behavior in a very widely used previously non-HTTP client. We then observed the number of clients that failed to connect via our custom protocol, but succeeded in falling back to HTTPS.
The numbers were negligible.
It's ridiculous that we'd seriously believe that we can't trust that TCP works on the internet. We joke about it being the "interweb", but I see no reason to sow fear, uncertainty, and doubt, and thus and actually turn the interweb into reality.
I don't see that we should model the internet architecture on bad technical choices made on a limited number of open wifi networks.
Or, we just frame our standard protocol over websockets as an (unfortunate) fallback, if it ever is revealed to be a real problem.
Given how many other things are broken by networks that foolishly only open port 80 and 443, and their (in my experience) relative rarity, I'd suggest that it's not worth bothering with, except possibly as a fall-back to measure the actual number of people trying to use your service behind such a network.