After the ingest-bug streak, the obvious move was making sure those bugs stay dead. Mid-June we wired the format-matrix end-to-end into CI: every accepted input format — CSV, parquet, JSONL, the numpy family, the raw *vecs family, and the rest — gets uploaded to a live cluster, ingested, and queried for exact expected results. A core subset runs on every push; the whole matrix runs nightly. Ingest regressions now have a red light with their name on it.
Making that sustainable was its own engineering. The suite got per-run collection namespacing so runs can't collide, create-confirm and upload retries so infrastructure blips don't read as format bugs, and hard timeouts on upload and smoke steps so a hang fails fast instead of wedging the runner for an hour. Right-sizing the per-case fixtures made the suite about 8x faster without losing the exact-match property. A slow, flaky gate is a gate people learn to ignore; the whole value is in being trustworthy enough to block on.
Two side quests from the same week. Our docs-example tests were hitting real third-party APIs — which meant every CI run sent traffic to Wikipedia and a couple of museum archives from one shared egress IP, an accidental slow-motion load test of someone else's servers. Those dependencies are mocked with recorded fixtures now; the docs tests test our code. And we switched release builds from fat to thin LTO, which parallelizes the link-time tail that had been serializing every deploy — with the follow-through of actually bumping the deployed version so the change shipped to real machines instead of living only in CI. A build-speed win that never reaches production is a benchmark, not an improvement.
June ended with the gates green, which is the least glamorous sentence I'll write all year and the one that made July's benchmarking month possible.