Expand description
The closure tier: every node runs its generated op over a flat u64
slot buffer, by-reference outputs as Ref2 pairs into step-owned
scratch (compiled_handles.md).
Four kernel types, each produced by a distinct compiler path. They
differ in what set_inputs marks and whether eval_for_slot
consults the cone guard; the shared step loop reads the mode’s
use_clean flag per step.
| Type | Push (per-node skip) | Pull (cone guard) |
|---|---|---|
CompiledKernelRaw | — | — |
CompiledKernelPush | yes | — |
CompiledKernelPull | — | yes |
CompiledKernelPushPull | yes | yes |
Structs§
- Compiled
Kernel Pull - The closure tier with the cone guard: an output whose cone no changed input reaches is not recomputed.
- Compiled
Kernel Push - The closure tier with per-step skipping: a changed input invalidates its dependents through the plan, and a current step is skipped.
- Compiled
Kernel Push Pull - The closure tier with per-step skipping and the cone guard.
- Compiled
Kernel Raw - The closure tier with no provenance: every evaluation runs every step.