[Aschenbath]

Split out of #1131 at review.

the_embedding_gate_is_never_held_beyond_its_ceiling sends twelve batches through a fake embedding server that answers each one after a fixed delay. It makes three assertions: the peak number of batches in flight is at most EMBED_JOBS (4), at least two batches overlap, and total time is under three quarters of fully serial. With a 150 ms delay that bound is 1.35 s, and the fixed cost of writing 192 chunk vectors to PostgreSQL takes a large share of it. In #1131's local full run (Linux x86_64, 2 CPUs, PostgreSQL 16/pgvector under proot), the test took 1.51 s and failed even though peak and overlap were both correct. The gate was fine; the database was slow.

The delay is now 400 ms, so waiting on the fake network accounts for most of the elapsed time. The bound is still three quarters of serial (3.6 s against 4.8 s): an ideal run is three rounds of four, about 1.2 s, so a regression that serializes the gate still fails. The peak and overlap assertions, which are what actually check the ceiling, are unchanged. The test now takes about 0.75 s longer.

Not run locally on this branch: I don't build Rust on this machine, so CI's cargo test --workspace is the check.

[WaylandYang]

Thanks — the peak and overlap assertions are unchanged and a serialized gate still fails. Merging.