I started Aoede on Saturday. By Wednesday, instead of piling on more features, I stopped and ran a full quality and performance review of the app and its packages. About fifty findings came back, most of them real, and I spent the day clearing the board.
A few were the kind you only catch by looking hard. Pause then resume quietly re-synthesized the current sentence almost every time, so continuing meant a multi-second Kokoro spinner instead of picking up where I left off. The Kokoro warm-up ran on every launch regardless of voice, so anyone on a system voice or the silent pacer kicked off a ~330 MB download they would never use. "$1,000" was read aloud as "one dollar, zero." Cover colors reshuffled each launch because the hue came from String.hashValue. And the iOS string catalog had zero Japanese across all sixty keys.
The rest was unglamorous: deleting a dead M1-era speech path, folding duplicated download and tokenization code together, streaming a tar extraction that used to spike hundreds of megabytes on iPhone, and finally testing the playback controller, which ran all of audio with zero coverage. Two findings I left open on purpose: download checksums and resume (a correct version needs real resume-data plumbing and pinned hashes I do not have), and one per-word gesture cost I would rather measure in Instruments than guess at. The point of a review is not to close every line. It is to know which ones you are leaving open.