When your test suite needs a GPU, a model loaded into VRAM, and three minutes to run end-to-end, standard testing advice stops applying. Here is what I learned from building a 685-test Rust suite for a production TTS pipeline — what to mock, where to draw the integration boundary, and which four timeouts you can safely ignore.
When I added resume logic to a TTS pipeline processing hundreds of audio segments, I thought I needed four states. I needed two. Here is why minimal state machines produce more reliable recovery than exhaustive ones.
Before a TTS pipeline can give each character a distinct voice, it has to know who is speaking. That means parsing manuscript dialogue — curly quotes, unmarked speakers, continuation lines, omniscient narration. Here is why this is harder than it looks.