Would you mind elaborating? I.e., in what way does it not work well? Your upstream in this case is Debian, they still maintain i386 (dpkg arch, not “would run on an actual i386”) as a first class arch, and you don’t hear “Debian” and “not maintained correctly” in the same sentence very often. As long as the i386 packages are built from the same source as the amd64 ones, they stay current, most fixes trickle down to them as well. The exposure is limited to issues that only manifest in the i386 version of a package. That’s much better odds than amd64 universe and PPA packages, that may not receive any updates at all and be of questionable quality to begin with.
The fact is, current amd64 systems can run 32-bit code natively, removing support on the OS level is a regression in hardware support. I’d be against it in principle even if there weren’t any major use-cases – and there are plenty. (One could even make the case for re-introducing i386 as an installable architecture, if not for 32-bit/hybrid Atom boards and netbooks then for lighter [= more] VMs, but I digress.) The point is, 32-bit code is not legacy, hardware that cannot run 64-bit code is – that’s a big difference. Even if it were legacy, it looks like many people still want or even need to run 32-bit software, and why shouldn’t they? When choosing an OS, they ask “Does it run my stuff, and does it run it well?”, not “Is it pure 64-bit?” – and for once, they got it right.
IMHO, running both 32-bit and 64-bit code on amd64 at the same time should be fully supported and as seamlessly as possible – Windows does that very well – for the foreseeable future, and multiarch is, as of now, the best way to do that. Compiling the 32-bit packages to target amd64 CPUs is an option, as is selectively dropping non-essential packages that won’t compile and would be too much effort to fix.
Bundles (Flatpack/Snap/Appimage) currently negate all advantages of centralised package management, containers/VMs aren’t there yet, either. Too much overhead, too much complexity to manage, too fragile for the desktop, not at all well integrated, let alone seamless.
It’s highly likely that a better solution than multiarch will emerge, maybe in conjunction with a paradigm shift to containerise everything transparently (cf. Qubes), who knows, but it isn’t even on the horizon yet, so why the rush?