Performance results

These numbers describe specific workloads on one machine. Lower is better within a row; cross-row comparisons are not meaningful.

Minimal Command-to-State delivery

Median nanoseconds per operation in the stable-instructions full run:

UseCase Forge is slower in this minimal row. It performs admission, Execution lifecycle, Snapshot publication, terminal processing, and bounded history that a direct observable or reducer update does not. That additional work explains a semantic difference, not a blanket performance exemption.

Queue-scaling work

The early core candidate exposed accidental queue scans. After optimization, a default 10,000-Command burst changed from 693.76 ms to 45.72 ms in JIT and from 1,469.83 ms to 33.92 ms in AOT. Later hardening changed a 5,000-head busy-lane scheduler case from 2,677.953 ms to 2.323 ms in JIT and from 2,205.400 ms to 1.289 ms in AOT.

Those are before/after results for internal candidates, not comparisons with other packages.

Regressions and variability retained

• the hardening candidate's 5,000-operation cancellation storm regressed by 12.35% in JIT while improving by 12.17% in AOT;
• the stable-instructions JIT 10,000-rejectNew row changed materially and had 13.31% coefficient of variation; its AOT counterpart also had 19.11% variation;
• the stable AOT 10,000 and 100,000 sequential rows measured 3,624.21 and 4,034.24 ns/op, with greater variability than the diagnostic hot-path rerun;
• no allocation or GC attribution is claimed.

Raw results

Read the methodology before using a number in a design decision.