So I see 2 steps here: dropping UTF-16 altogether (well, already, because there are plenty of extended codepoints above 0xFFFF), and when approaching the 31-bit limit - inventing something like "zero-width codepoint joiner" to compose codes of arbitrary length.

For example, in a hypothetical alien language, a hypothetical character "rjou" would have a code 0x2300740457 (all the previous codes are exhausted). We can't express this with a single code, so actually we split it into 2-byte parts and write "#" (0x0023), joiner, "t" (0x0074), joiner and "ї" (cyrillic letter yi, 0x0457). As we have a joiner between these codes, we know that we must interpret and display them not as a "#tї" sequence but as a single alien "rjou" character. I think you get the idea.