binary-trees at one thread on a Xeon E3-1231 v3: 1110 -> 870 ms, now ahead
of Node (950 ms) there; at 8 threads 461 ms against Rayon's 601.

release:check on WSL (5541 s) for b2a867757: 257 of 258 commands passed. The
perf gate found 0 GeaStack regressions and every claim held (binary-trees at
20 threads +41.2% over Rayon; parallel geometric mean at 20 threads +21.1%
over Rayon, +15.3% over the C++ pool). The one failure is 2 control drifts in
the comparison programs (idiomatic strings, Rust -5.4% and C++ -13.3% at one
thread against their baselines), not GeaStack code.