df64 GPU conformance probe

One known-answer test per df64 operation, plus the composed loran chain, run on both backends — WebGPU (WGSL) and WebGL2 (GLSL) — through the same emit + _fp64 guard the examples use. Each result is read back as a real number and compared to the CPU f64 oracle. Every test is built so plain f32 gives a wrong answer, so PASS means df64 held on this GPU. The final Random composition family fuzzes seeded random df64 expression trees — where a driver's fast-math has a whole expression to reassociate, and so the most room to delete the error terms single ops survive. Run it, then hit Copy and paste the digest back (per device).

include stress battery (integer flavor + random trees — heavy compiles, may hang on D3D11 / Windows)

SwiftShader (CI / headless) passes everything — the diagnostic value is on real GPUs, where driver fast-math can fold the df64 error terms away.