feat(examples): mnist_mlp — PyTorch + C MNIST dense-MLP parity demo

[LeoBuron]

What

Adds examples/mnist_mlp/ — a self-contained MNIST dense-MLP classifier demo with exact PyTorch ↔ C bit-parity, plus a shared examples/_shared/mnist_data.py loader. First of two MNIST examples (the conv twin mnist_cnn follows in a separate PR), replacing the deleted legacy example/MnistExperiment with the factory layer API.

Mirrors the canonical har_classifier/ pattern: prepare_data.py → train_pytorch.py (reference + per-layer weight export) → train_c.c (factory-API trainer, two modes) → compare.py (informational) → README.md + CMakeLists.txt, wired into the c-bit-parity CI job.

Model: Flatten[1,28,28]→Linear(784→64)→ReLU→Linear(64→10)→Softmax, CrossEntropy (~51 K params). The framework is 1D-only (no Conv2d); the MLP's flatten consumes [1,28,28] directly (the channel-1 acts as batch), so no loader reshape is needed — matching the legacy MLP.

Verification

• Bit-parity gate (the CI gate): C predictions int32 bit-identical to PyTorch, 10000/10000. BIT_PARITY=1 loads PyTorch weights via StateDictApi (npyLoadFlat, per v2 examples (HAR/ECG) forwards not parity-faithful — bit-parity diff fails after the crash fix #177) and runs inference only.
• C unit tests 62/62, pytest 22 passed, all examples build (rot guard), train_c.c clang-format-21 clean.
• Final whole-branch review (Opus): READY TO MERGE, no Critical/Important findings.

Notes

• The train-from-scratch demo (compare.py) is informational, not a gate — independent random init, and the known C-vs-PyTorch training divergence is deferred. On full MNIST the C demo runs ~75 min (framework trains one sample at a time over 54k samples); the README documents this and points at the fast bit-parity check instead.
• Spec: docs/superpowers/specs/2026-06-26-mnist-examples-design.md; plan: docs/superpowers/plans/2026-06-26-mnist-examples.md (both gitignored scratch).

🤖 Generated with Claude Code