You have commit messages with errors ⧗ input: feat: PyTorch-style GradientTape with op-level backward recording (#1057) Core GradientTape<T> rewrite for automatic differentiation: - AsyncLocal<T> tape stack for async pipeline safety (replaces ThreadStatic) - RecordOp(name, inputs, output, backward) — Engine ops register backward functions directly, no separate ExportComputationGraph needed - Gradient(Tensor<T> loss) → Dictionary<Tensor<T>, Tensor<T>> — reverse-mode AD via reverse topological traversal of recorded ops - NoGradScope<T> — disables recording for inference (like torch.no_grad()) - Gradient accumulation for tensors used in multiple ops - Persistent tapes for multiple gradient computations - Backward-compatible: Watch(ComputationNode), Gradient(ComputationNode), RecordOperation(ComputationNode), StopRecording/ResumeRecording still work for existing callers during migration 11/11 tests pass: - Add, Multiply, Chain rule, Gradient accumulation - Persistent/non-persistent, NoGradScope, disposed safety - AsyncLocal flows across await, unwatched tensor exclusion Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> ✖ subject must not be sentence-case, start-case, pascal-case, upper-case [subject-case] ✖ found 1 problems, 0 warnings ⓘ Get help: https://github.com/conventional-changelog/commitlint/#what-is-commitlint ⧗ input: feat: Conv2D, BatchNorm, LayerNorm, Softmax backward functions (#1057 Step 2 cont.) New DifferentiableOps: - Conv2D: backward via Engine.Conv2DBackwardInput/BackwardKernel - LayerNorm (Ba et al. 2016): full backward through mean/variance/gamma/beta - BatchNorm (Ioffe & Szegedy 2015): full backward through batch statistics - Softmax: Jacobian-vector product dx_i = s_i*(dout_i - dot(dout,s)) All verified against finite-difference gradient checking (rel tol < 1e-4): - Conv2D, LayerNorm (weighted), BatchNorm (weighted), Softmax (weighted) - Chain: Conv→ReLU, LayerNorm→Tanh, tanh(x@w+b), MSE loss 35/35 tests pass (11 tape + 24 gradient checks). Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> ✖ subject must not be sentence-case, start-case, pascal-case, upper-case [subject-case] ✖ found 1 problems, 0 warnings ⓘ Get help: https://github.com/conventional-changelog/commitlint/#what-is-commitlint ⧗ input: feat: GPU path for named-port multi-input layers (#1058 Step 4) LayerBase<T>: - ForwardGpu(IReadOnlyDictionary<string, IGpuTensor<T>>) — default delegates to ForwardGpu(input) using "input" key - BackwardGpuMulti(IReadOnlyDictionary<string, IGpuTensor<T>>) — default delegates to BackwardGpu(gradient) NeuralNetworkBase<T>: - ForwardGpu(input, auxiliaryInputs) — routes named GPU tensors to multi-port layers during GPU-resident forward pass, mirroring the CPU ForwardWithMemory(input, auxiliaryInputs) 58/58 tests pass. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> ✖ subject must not be sentence-case, start-case, pascal-case, upper-case [subject-case] ✖ found 1 problems, 0 warnings ⓘ Get help: https://github.com/conventional-changelog/commitlint/#what-is-commitlint ⧗ input: feat: Concatenate/Split ops + perf optimizations from audit New ops: - Concatenate(tensors, axis): forward concat, backward splits gradient - Split(x, splitSizes, axis): forward splits, backward concatenates Both essential for U-Net skip connections and multi-head attention Performance optimizations: - ReLU mask: byte[] (1B/element) instead of Tensor<T> (4-8B/element) - InputPorts/OutputPorts: cached with ??= to avoid allocation per access All 66 gradient check tests pass. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> ✖ subject must not be sentence-case, start-case, pascal-case, upper-case [subject-case] ✖ found 1 problems, 0 warnings ⓘ Get help: https://github.com/conventional-changelog/commitlint/#what-is-commitlint ⧗ input: refactor: migrate DeepBeliefNetwork, FastText, GloVe, Word2Vec, NeuralNetwork, ⚠ body must have leading b