fix(gf16): Fix F0.2 architecture description — Small CNN not MLP (#497)

- CIFAR-10 input is 32×32×3 = 3072, NOT 784×1
- Model must be a small CNN (2 conv + pooling + FC), NOT same MLP as F0.1
- Architecture: Conv(3→16)→ReLU→Pool → Conv(16→32)→ReLU→Pool → FC(1152→128) → FC(128→10)
- Expected baseline accuracy: ~70-75% (small CNN without augmentation)

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Author: Antigravity Agent

docs/research/IMPLEMENTATION_F0_2_CIFAR10.md

@@ -0,0 +1,61 @@
+# F0.2: CIFAR-10 Cross-Stack Validation Implementation
+
+**Status:** Ready to implement
+
+## Overview
+
+F0.2 validates GF16 format against competing formats (FP16, BF16) using CIFAR-10 dataset instead of MNIST. This provides scientific credibility beyond synthetic MNIST benchmarks.
+
+## Architecture
+
+```
+Dataset: CIFAR-10 (10k images, 32x32x3 RGB, 10 classes)
+Model: Small CNN — Conv(3→16)→ReLU→Pool → Conv(16→32)→ReLU→Pool → FC(1152→128) → FC(128→10)
+Formats compared: GF16, FP16, BF16
+Output: results/cifar10_metrics.json
+```
+
+### Architecture Details
+
+- **Input:** 32×32×3 = 3072 features (RGB images)
+- **Conv1:** 3 channels → 16 filters, 3×3 kernel, stride=1, padding=1
+- **Pool1:** 2×2 max pooling, stride=2 → output 16×16×16 = 4096
+- **Conv2:** 16 channels → 32 filters, 3×3 kernel, stride=1, padding=1
+- **Pool2:** 2×2 max pooling, stride=2 → output 8×8×32 = 2048
+- **FC1:** 2048 → 128 dense layer
+- **FC2:** 128 → 10 output layer (softmax)
+
+### Baseline Expected Accuracy
+
+Small CNN without data augmentation on CIFAR-10: **~70-75%**
+
+## Implementation Notes
+
+### 1. Build System Integration
+The Zig build system uses explicit `b.step()` calls. The `bench_cifar10` executable needs proper registration in build.zig.
+
+**Current issue:** build.zig expects steps like `const bench_XXX_step = b.step("...", "...")` pattern. We're using a simpler approach.
+
+### 2. Model Weights
+CIFAR-10 requires a trained small CNN for CIFAR-10. Options:
+- Train from scratch in PyTorch (small CNN architecture as above)
+- Export weights in binary format (compatible with F0.2 loader)
+- Use cross-entropy loss, Adam optimizer, 50-100 epochs
+
+### 3. CIFAR-10 Dataset
+The loader at `src/cifar10_loader.zig` downloads from:
+`https://www.cs.toronto.edu/~kriz/cifar-10-python.tar.gz`
+
+This is 10k images (32x32x3 RGB), requires downloading (~160MB).
+
+## Phase 2 Summary (Complete)
+
+- Competitive analysis documents copied to docs/research
+- arXiv + GitHub research completed (30+ papers, multiple repos analyzed)
+- Issue #497 updated with findings
+
+**Next Phase:** 3. Experimental Directions (academic paper, hardware bridge, optimizer suite)
+
+---
+
+**Document Status:** Created implementation plan for F0.2.