went through everything running on the server and asked which services pay rent in ram and cpu for work a compiled language would do for almost nothing. some of the answers were fine as they were, a few were genuinely worth a rewrite, and the tagger got the first slot — it was the biggest memory hog for the least complexity.
the method is boring on purpose: run the rust version and the original side by side against the same inputs and diff the outputs until they match. nobody should trust a rewrite that changes behavior while it changes languages, so parity is the gate, not a hope.
the diffs were the interesting part, because they exposed how much behavior hides in a language's details. first diff: a hashmap loses insertion order, python dicts keep it, so palette ties sorted differently depending on which language ran. second diff: a float total summed in iteration order, so the same numbers produced a slightly different result depending on where they happened to live in memory. both were fixed the same way — make the order explicit instead of inherited, and suddenly the outputs agree.
next up is the taste engine port, with a backup kept safe so the old version can come back while the new one proves itself. the plan stays the same the whole way down: one service at a time, parity first, swap only when the diff is empty.