JSONL duplicate keys under parallel load. Hot-append path was creating duplicate entries because content-hash dedup wasn't wired up for appends -- only for initial ingestion. So the first file gets deduped correctly and every subsequent append just blindly adds rows, duplicates and all.
Looked like one bug from the outside. Turned out to be two. The dedup bypass was one. The other was a race in the append path where two concurrent appends could both read the existing key set, both decide their keys are new, and both write. Classic read-check-write without a lock. Fixed both but I only found the second one because the first fix didn't fully resolve it.