One language, deliberate execution choices

Aner parses, checks, and interprets source through a native executable. Compiling that interpreter is distinct from compiling an Aner program. Native compilation of user programs is a later milestone.

Rank two Tensor operations now support CPU Float64 and Float32, with optional native Metal/MPS and CUDA Float32 backends. Device selection, precision conversion, availability checks, memory transfers, reverse differentiation, and optimizer updates have explicit contracts. Distributed execution, multiple local GPUs, general rank autograd, and automatic memory scheduling remain future work.

Performance evaluation must report matched shapes and precision, native backend identity, synchronization, conversion and transfer costs, numerical agreement, and workload limits. The bounded benchmark runner supplies these measurements without implying that every GPU workload is faster or that results establish scaling, energy efficiency, or hardware parity.