BatchEmbedWithPool fans a batch out over the ants pool, so with
BATCH_EMBED_SIZE=1 every text becomes its own sub-batch and several workers run
at once. Each worker starts by testing firstErr to see whether an earlier
sub-batch already failed, but that read was unsynchronised while the writes are
under mu. The race detector reports it between the guard and the failure branch.
The guard exists to stop billing a provider for work whose result will be
discarded. A worker that misses the write issues the request anyway, so the race
defeats the protection it was added for.