── Stateful-variables variant of compile_generic_fused_step ────────

Same lifecycle and contract as compile_generic_fused_step — the trace covers forward + loss + (ghost-grad → opt update) — but a designated subset of the external feeds is promoted to MPSGraph stateful variables instead of placeholders, and the matching subset of the trace's output targets is bound to assignVariable operations instead of executable graph outputs.

Per-step semantics (when paired with run_executable_inplace):

• The transient feeds (model input x, loss target, per-step scalars) remain regular placeholders — same I/O as before.
  • Each variable feed initialises its variable from the current MTLBuffer contents of the matching Lucid Tensor at compile time. Later steps reuse the variable's persistent internal storage; the executable's input list excludes them entirely.
  • Each variable write target is emitted as an assignVariable op (added to targetOperations). Immediately after the assignment, a readVariable op is included as a regular targetTensor so the new value can be flushed back to the Lucid Tensor's MTLBuffer on each call — keeping model.fc1.weight-style reads up to date between steps.

Why this wins

The original run_executable_inplace path allocates one fresh MTLBuffer per output target (newBufferWithLength:, ~5-30μs × N), wraps it as MPSGraphTensorData, runs the executable, then swaps the Lucid Tensor's MLX array onto the fresh buffer. Variables eliminate the per-step buffer allocation: the readback writes directly into the existing Lucid Tensor buffer, and the assignment writes happen inside the variable's persistent storage with no host-visible buffer at all.