Thanks to @ekaitz_zarraga, I got sucked into the GNU Mes rabbit hole.
After reading through some of the docs, I realized the Uxntal bootstrap stops one step short. I check that the assembler reproduces itself, but only ask one emulator. Running the same steps on a second independent implementation and comparing checksums gives me a second opinion that rules out implementation-specific issues.
Since I had this Uxn VM written in Uxntal thing, running the same bootstrap chain inside it gives an independent witness. Two unrelated implementations agreeing on the same checksum means the result is no longer dependent on any single emulator's behaviour.
Running the assembler inside uxnmin, which was itself assembled by that assembler, and verifying the checksums match means the assembler and the virtual machine are each vouching for the other. It's a much stranger circle than Mes, but it's neat and it closes.
