ADR: reverseMode result ABI

Context

Section 10 types handle f(args) with reverseMode as a nested pair (primal, (grads…)) and builds that pair in IR. The emitter then prints func.func @main() with no results, because infer_type only accepts tensors. Section 11 therefore takes g as an explicit argument.

MLIR tuple<> types do not bufferize on the existing one-shot-bufferize + LLVM C-wrapper path. The bootstrap surface also cannot write a one-element tuple type (T), so a nested (primal, (g)) cannot be used as a fn annotation.

Decision

Specify the ABI in 12-reversemode-result-abi.md:

• Language type is the flat tuple (primal, g₀, …, gₙ₋₁).
• Lowering is MLIR multi-result func.func, not tuple<>.
• Host .so loads use an existing llvm.emit_c_interface fn whose body is handle f(x) with reverseMode.
• Do not add tuple projection, @main arguments, or a new operator.

Consequences

• ad_neg / ad_nll / other reverse-mode goldens gain @main results.
• A wrapper example can return Kina-executed gradients from a .so.
• Loss → VJP → SGD in one @main still waits on naming those results inside .kina.