The #51 Bessel-series CDF landed with provisional dispatch thresholds on every tier ("initial value, pending benchmark tuning"). The v2.4.0 calibration legs measured them:
- kNeon (M1): sustained V→P crossover NEVER — BEST=VECTORIZED at 2M in all 3 runs (forced-parallel 0.20× at 2M). Encoded row dispatched parallel from 100000, a measured ~5× slowdown for every large batch. Bundle:
data/profiles/dispatcher/2026-09-04T04-22-28Z_darwin-arm64_dev-v2.4.0_sha-5f27ee1/.
- kAvx2 (Kaby Lake): sustained NEVER in all 3 runs (forced-parallel 0.27× at 2M). Encoded row dispatched parallel from 200000. Bundle:
data/profiles/dispatcher/2026-09-04T02-36-22Z_darwin-x86_64_dev-v2.4.0_sha-5f27ee1/.
Mechanism is #111: the batch CDF is O(count × j_max) full-array sweeps with heap allocations — memory-bound, the same class as Beta PDF/LogPDF, so a real fork adds contention without adding bandwidth.
Fixed on dev for kNeon and kAvx2 (CDF → NEVER, cross-referenced to #111 for re-profiling once that lands).
Remaining: the kAvx512 row (CDF 25000) is equally provisional but there is no checked-in Zen 4 sustained data to justify changing it — the 12-thread Zen 4 may legitimately differ. Needs one sustained re-measure on the Zen 4 machine; the row comment now says so. Close this issue when the kAvx512 cell is either confirmed or corrected from measurement.
The #51 Bessel-series CDF landed with provisional dispatch thresholds on every tier ("initial value, pending benchmark tuning"). The v2.4.0 calibration legs measured them:
data/profiles/dispatcher/2026-09-04T04-22-28Z_darwin-arm64_dev-v2.4.0_sha-5f27ee1/.data/profiles/dispatcher/2026-09-04T02-36-22Z_darwin-x86_64_dev-v2.4.0_sha-5f27ee1/.Mechanism is #111: the batch CDF is O(count × j_max) full-array sweeps with heap allocations — memory-bound, the same class as Beta PDF/LogPDF, so a real fork adds contention without adding bandwidth.
Fixed on dev for kNeon and kAvx2 (CDF → NEVER, cross-referenced to #111 for re-profiling once that lands).
Remaining: the kAvx512 row (CDF 25000) is equally provisional but there is no checked-in Zen 4 sustained data to justify changing it — the 12-thread Zen 4 may legitimately differ. Needs one sustained re-measure on the Zen 4 machine; the row comment now says so. Close this issue when the kAvx512 cell is either confirmed or corrected from measurement.