Current state: different GPUs produce different results so we have verification failures and an auditor can not repro
Desired state: bit-identical outputs across A40/T4/2080 via FP64 with controlled rounding
Actions: forward_f64() layers, RoundingLog struct, --deterministic flag, Ray worker integration
Success Criteria: Prover on A40, auditor on T4 then will get identical outputs, rounding log less than 2MB, and overhead less than 1.5x

Related to #19