Module runtime_fused_arm 

Runtime-fused-op arm emission — the gated Op::Branch that makes an adopted (JIT-synthesized / imported) fused op reachable, mirroring crate::decode_flash.

Same constitutional posture as the flash arm: the optimizer emits/prunes arms; backends never decide. Arm 0 is the region’s existing primitive subgraph (the decompose / correctness oracle the graph already holds); arm 1 is the synthesized Op::Fused(runtime_id, Runtime) kernel, offered only when a kernel is bound for the target backend (fused_kernel_available_in) and pinned to that backend. finalize_branches guarantees the merge reads arm 0, so a finalized-but-unpicked graph — and every backend with no synthesized arm — realizes on the primitive route, byte-identical. The route picker chooses arm 1 when it wins; the kernel resolves from the FusedOpId-keyed runtime sidecar.

Unlike the flash arm (emitted at graph-build by the decode builder), a runtime fused op is synthesized after load, so this arm is emitted during (re-)optimization — a pathfinder scans the base map for regions matching an adopted op and calls this per match. This module is the emitter; the pathfinder that finds matches + splices reconverge points registers it into optimize_graph.

Structs

RuntimeFusedSpec

A matched runtime-fused region the optimizer offers as an arm candidate. Arm 0 (primitive_sink, the region’s existing output) is already in the graph; this describes how to build arm 1 (the fused kernel) and where to splice the branch.

Functions

offer_runtime_fused_arm

Offer a runtime-fused Op::Branch arm for an adopted op, gated on kernel availability. Returns Ok(Some(branch)) when the arm is emitted, Ok(None) when the gate declines (no kernel for backend → no arm; the region stays on arm 0’s primitives). Never a realize-time decision — the capability gate and the backend pin happen here, in the optimizer.