I've written before about PCA returning the wrong nearest neighbors because a 32-dimension projection threw away most of the signal. The real lesson wasn't "back PCA out of the default." It was that an index can lie about how good it is -- score beautifully on the little sample the optimizer trains on, then fall apart on the real collection. If promotion trusts that self-reported number, a lying index goes live.
So promotion doesn't trust it. When the auto-optimizer promotes a new index, a probe re-measures it at serving scale: forced brute-force ground truth over real query traffic, not the sample it was tuned on. Under the recall floor, demoted. Immediately.
During the million-vector run I got to watch it happen. The optimizer built a PCA index, sample-scale recall looked great, promoted. 36 seconds later it was gone -- the probe measured it against forced brute force, saw it under the floor, pulled it. The old index kept serving the whole window -- recall never dropped for a single customer query. Sample-scale recall is a hopeful estimate. Serving-scale recall is the truth.
Watching a bad index promote and un-promote itself in 36 seconds is deeply satisfying. Like watching a mousetrap you set months ago finally go off.