Runtime-selected CPU specialization without target-cpu=native

[GeneralPawz]

Roadmap item, Parked: performance and execution providers.

Parked deliberately. See ADR 0013.

Why parked

Optimizing before the capability surface is real trades away the thing being
built. This is not declined — the ordering is intentional.

Unblocked when

A benchmark harness exists (#12) and the v0.3 intersection work is done, so a
specialization claim can be measured on the same harness rather than asserted.

Done when

• Specialization is selected at runtime, no target-cpu=native assumption
• Same-harness before/after evidence published
• Scalar path remains the differential oracle