Built hosted E2E on .150. One collection per format, fixed query vector, exact top-k validation against the real cluster. Not mocks. Not local. Actual uploads, actual queries, actual result comparison against known-good answers.
The test repo is telling me what correctness means better than the code comments do. When the safetensors ID bug was live, the E2E caught it immediately -- expected ID "abc123", got "7". When the bvecs seek bug was live, the dimension mismatch blew up the query before results even came back. Every format bug from the last 2 weeks would've been caught on first run.
Been meaning to add historical run summaries so I can compare before/after when I change things. That's for tomorrow-me though. Today-me is going to bed.