[vladwing]

Concurrent upserts can store one (symbol, filename) pair twice. Six identical cold runs of a build of main ended with symbol_filename row counts up to 99 apart.

insert_batch_and_mark() runs a batch's functions and types inserts with tokio::join!, and both end in SymbolFilenameStore::insert_batch(). Several inserters also run at once. When two merge_inserts of the same new pair open the table at the same version, both miss and both insert.

Serialize insert_batch() with a process-global async mutex, held across open_table and every chunk's merge_insert: the pattern content.rs uses for its shards. Every symbol_filename upsert in a process now waits for the one before it; the cost to a cold index is not measured. The mutex is a static rather than a field so that stores of different DatabaseManagers opened on one database also wait on each other.

The new test stores every pair twice without the lock (5,120 rows for 2,560 pairs) and once with it. It runs two insert_batch() calls concurrently with the same 64 new pairs, for 40 rounds with fresh pairs each round.