docs: OPERATORS.md — Session 5 summary (batched_matmul, groups conv, groupnorm, ~71 ops)

docs: 3-way simulator/PyTorch/Keras parity verified

docs: verified simulator<->PyTorch export parity

docs: one-click forward pass (run_forward + RunResultView)

docs: exec_coverage Tauri command + api.execCoverage() frontend surface

docs: nn.linear ONNX mapping = Gemm

docs: execution coverage audit API + io node passthrough

docs: embedding/dropout/batchnorm execute (~61 ops)

docs: exact multi-head MHA + reshape-0 lowering + post-lower inference

docs: layernorm/rmsnorm execute numerically (~58 ops)

docs: pooling kernels execute; full CNN forward pass runs in numeric runtime (~56 ops)

docs: nn.conv2d now executes numerically (direct NCHW conv, owned ParamSpec)

docs: nn.linear executes numerically via owned ParamSpec (deterministic seeded materialization)

docs: composite blocks now expose attrs in inspector

docs: exec now ~50 ops (reductions + flatten/squeeze/unsqueeze)

docs: attention 1/sqrt(d_k) scaling now numerically exact

docs: all movement ops now have CPU numeric kernels

docs: movement.transpose/concat/gather/onehot now have CPU numeric kernels

docs(operators): document math.scale

docs: document expanded nflow-exec CPU numeric kernel coverage (~37 ops) + exec step in 'adding an operator'

docs(operators): document movement.onehot

docs(operators): document movement.gather

docs(operators): document movement.slice

docs(operators): document movement.pad / tile / expand

docs: training-config / optimizer codegen (per-backend slot translation)

docs: add Loss operators section + framework-asymmetric loss export (83 prims)

docs: sync catalog (76 prims) with flow.where + math.clip

docs: sync operator catalog (74 prims) + param-ownership note with new ONNX/Keras-parity ops