Module compile 

Kernel compilation: assembled DAG → fast executable kernel.

Everything in this module is on the path between assembly::PolydatAssembler and an executable kernel. The pipeline:

PolydatAssembler ──resolve──▶ ResolvedDag ──compile_with(Engine)──┐
  (fusion, adapters,                                              │
   round-trip lint,          ┌────────────────────────────────────┤
   topo sort)                ▼                  ▼                 ▼
                 Interpreter(JitMode)  Closures(Provenance)  Native(Provenance)
                 cone::extract_jit_cones  closures::           hybrid::
                 → PolydatKernel          CompiledKernel*      HybridKernel*

The host names the engine (select::Engine); under Provenance::Auto the selector picks the provenance mode from the graph’s shape. Pure native code (jit::JitKernel*) is the differential tier behind the hybrid kernel.

• assembly: the public construction surface (assembly::PolydatAssembler + assembly::WireRef) and the per-engine compile paths.
• fusion: graph-level subgraph fusion pass; runs during assembly after wiring resolution.
• roundtrip_lint: the structural type-round-trip lint run at resolution.
• cone: cone-level JIT inside the interpreter kernel (SRD-105), under a cone::JitMode.
• lattice: the engine-mix report of a compiled program.
• select: the engine and provenance enums, and the heuristic that picks a provenance mode under Provenance::Auto.
• closures: the closure tier, one generated op per node over a flat u64 slot buffer, by-reference outputs as Ref2 pairs.
• hybrid: the native engine (native segments + closure steps sharing a flat u64 buffer).
• marshal: the boundary marshalling between slots and Values.
• externs: extern inputs and shared cells on the compiled engines.
• simd_plan, simd_tier1: scalar-flow SIMD promotion.
• jit: Cranelift lowering and the pure native kernels (feature-gated on jit).

Modules

assembly

Programmatic assembly API for building Polydat Kernels.

closures

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).

cone

SRD-105 — cone-level JIT inside the interpreter kernel.

fusion

Graph-level node fusion optimization pass.

hybrid

Hybrid kernel: per-node optimal compilation level.

jit

Phase 3: Cranelift JIT compilation of Polydat Kernels.

lattice

SRD-105 lattice report — what extraction actually did.

marshal

The boundary marshalling a compiled node kit reads its arguments and writes its outputs through (compiled_handles.md §4): a node crate’s kits use it as the core’s own do. Marshalling between typed Values and compiled-tier slots. One place owns the rule for every port type, so embedded cones and whole compiled kernels decode alike:

roundtrip_lint

Structural type-round-trip lint (C6b; governing principle: feedback_no_auto_string_conversion).

select

The engine a host chooses, the provenance mode a compiled engine is built with, and the selector that picks a mode from a graph’s shape when the host leaves it to Provenance::Auto.

simd_plan

Static type and node-variant qualification for scalar-flow SIMD promotion.

simd_tier1

End-to-end Tier-1 scalar-flow SIMD promotion.