zorch.sumcheck.jagged.fs

The per-round Fiat-Shamir hop + scalar reduce: observe the round poly, squeeze the challenge, and fold the claim/pad scalars -- traced into the whole-layer jit, one fused region per round. (The round's eval-point coordinate is sliced statically by the caller's loop index.)

The reduce (eval_coeffs(poly, r) + mass * eq_factor(r, z)) stays plain device ops that consume the zorch.duplex_fs composite's challenge output -- zorch stays optimization-agnostic, leaving any FS-kernel absorption to the consumer's xla layer rather than baking it in here.