The 8 test slices are the last thing the required all-checks-pass gate
waits for on every python PR: each slice carries ~968s of LPT-balanced
per-file work and runs 150-190s wall, while every other gated job is
done by ~90s. Per-slice fixed overhead is ~10s with the warm uv cache,
so slicing wider is nearly free: at 12 slices per-slice work drops to
646s and stays balanced (makespan == min within 1s in simulation
against the live durations cache).
Verified on this PR's own CI run: slices 141-187s -> 97-140s wall.
Peak run concurrency rises 23 -> 27 jobs; observed queue delay at 23
is 2-10s, well under the org limit.
71f1b371c6