Parent: #478
Current canonical evidence is environment-keyed, but optimization claims need to survive both Tier-1 Flow platforms. BLAS/LAPACK implementation, vector width, compiler and allocator differences can reverse small wins.
Run canonical operation-level and selected scaled benchmarks on Linux x86-64/OpenBLAS and macOS arm64/Accelerate with pinned threads. Maintain independent baselines per environment; never compare absolute timings across machines.
Acceptance: every canonical loss/win has per-platform status; regressions are evaluated only within matching environment IDs; platform-specific holes become explicit child issues; Pages can filter by platform; target #478 eventually has zero material losses on both supported performance platforms.
Parent: #478
Current canonical evidence is environment-keyed, but optimization claims need to survive both Tier-1 Flow platforms. BLAS/LAPACK implementation, vector width, compiler and allocator differences can reverse small wins.
Run canonical operation-level and selected scaled benchmarks on Linux x86-64/OpenBLAS and macOS arm64/Accelerate with pinned threads. Maintain independent baselines per environment; never compare absolute timings across machines.
Acceptance: every canonical loss/win has per-platform status; regressions are evaluated only within matching environment IDs; platform-specific holes become explicit child issues; Pages can filter by platform; target #478 eventually has zero material losses on both supported performance platforms.