0064 — Throughput is min(supply(batch), ack_capacity(workers))

Context. Measuring the pipeline’s ceiling required a model of what bounds it: the single prefetch loop supplying messages, or the workers committing per-message results.

Decision. The mental model is throughput = min(supply(batch), ack_capacity(workers)). Claim cost is sublinear in batch size (~3.4µs/row asymptote, so a single prefetch loop can supply ~290k rows/s); at batch=100 supply is round-trip-capped at ~17–18k/s, right at the ack ceiling. Raising batch exposes the ack wall (~20–22k msgs/s at 64 workers on this box), which scales with worker count.

Consequences. The prefetcher is not a hard ceiling until far beyond current throughput; the bottleneck is the per-message commit on the success/failure path. How it was diagnosed: pg_stat_activity wait-event sampling, plateau when raising batch (not supply-bound), and scaling with workers at fixed large batch (commit-path-bound). Rejected: “the single prefetch loop is the ceiling” — an earlier wrong conclusion drawn from measuring only batch=100.