Your retriever pulled 40 chunks instead of 8. Nobody noticed until the invoice.

Similarity thresholds drift, chunk overlap grows, top-k creeps up after a re-rank. The query still works, just five times more expensive. Calcis flags the cost spike on the next PR by computing per-route context-cost deltas and surfacing the regression with a link to the commit that caused it. Longer write-up plus a CI recipe landing here next.