SDK entrypoint import costs

Tracked by pm-38bskj and pm-cg1sjb.

This table measures fresh-process ESM import and module evaluation. The bare Node v26.5.0 process floor on linux/x64 was 44 ms p50 (56 ms p95) across 5 measured runs after one warm-up. Focused entrypoints are compared with the compatibility aggregate; negative reduction means the focused entrypoint was slower in this sample. The gate admits the median measured run against the unchanged upper-bound budget and 30 ms scheduler margin. A single cold or descheduled process therefore cannot fail the gate, while a majority of over-budget samples still does; p95 remains visible as diagnostic evidence.

Package export p50 p95 p50 above Node Reduction vs aggregate overhead
./sdk 320 ms 348 ms 276 ms 0%
./sdk/authoring 91 ms 99 ms 47 ms 83%
./sdk/contracts 152 ms 156 ms 108 ms 60.9%
./sdk/core 285 ms 291 ms 241 ms 12.7%
./sdk/governance 211 ms 220 ms 167 ms 39.5%
./sdk/graph 92 ms 105 ms 48 ms 82.6%
./sdk/merge 130 ms 145 ms 86 ms 68.8%
./sdk/query 132 ms 156 ms 88 ms 68.1%
./sdk/runtime 283 ms 312 ms 239 ms 13.4%
./sdk/testing 274 ms 319 ms 230 ms 16.7%

The aggregate @unbrained/pm-cli/sdk remains supported for compatibility. New packages should import the narrowest subpath that owns their capability. The committed budget file is an upper-bound ratchet and must not be weakened to hide a regression.


Performance - Sdk-entrypoint-import-costs remote
报告问题