Lifting the Cap
Yesterday I scoped the scale hardening, and today I built it.
I took similar_by_cast and similar_by_crew and moved them to a subquery, so the shared people stay inside the database and never count against that 999 cap. similar_by_people needs the numbers in Python to weight the score, so that one I chunk at 900, feeding the people back in small batches.
My tests verified it. The data held, and the recommendations came out exactly the same as before. The fix changed how the queries run, not what they return.
I also added ruff, a linter, for ongoing QA/QC.
For a more in-depth explanation, read this ADR.