docs

Background and decisions that are not derivable from the code. Nothing here should restate what CLAUDE.md already says, and nothing here should describe code that does not exist.

Removed: ai-cost-audit.md

It compared cloud stem-separation providers (Replicate, LALAL.AI) and a "copilot", and recommended keeping an execution order of Local → fake → Replicate → LALAL. It was also the one document CLAUDE.md's first working rule tells a reader to open before planning anything.

None of it was ever built. There has never been a Cloud-tier executor in this repository, "copilot" appears nowhere in the source, and ExecutionTier.Cloud — the enum value that whole ordering hung on — has now been deleted as well. A cost audit of spend that cannot occur is worse than no document, because it sends the next reader looking for a tier to extend rather than for the one that exists.

The costs that are real are all local and none of them bill: CPU time for HTDemucs (stem separation) and Basic Pitch (melody), and whatever model Ollama has installed for the written interpretation. If a paid tier is ever added, the audit belongs here again — written against code that is in the repo.