examples: full-SYM ECG reconstruction AE (Conv SYM + Quantization layers + MSE)

[LeoBuron]

Goal

First end-to-end full-SYM training proof. ECG-AE v2 (examples/ecg_anomaly_ae_v2: Conv1d + ConvT1d + ReLU + MSE, strides instead of pooling) is separated from full-SYM by exactly two dependencies: #45 (Conv/ConvT SYM) and #192 (Quantization layer).

Proposed work (milestone M3, one PR)

• SYM twin of the v2 example: explicit Quantization (requant) layers after every accumulator-range producer (Conv/ConvT emit raw accumulator mantissas, same contract as Linear/LayerNorm), MSE loss (already SYM-capable), SYM optimizer.
• Call the opt-in model validator from feat: Quantization layer — SYM→SYM requantization + full layer wiring #192 in the example setup — its first real consumer outside the chain test.
• FLOAT32-twin comparison + parity harness per the Stage-2 example precedent (PR feat(examples): ECG5000 reconstruction autoencoder (Stage 2/4) #163); tolerance derived analytically and documented in the example README.

Dependencies

Hard: #192 (M1) and #45 (M2) must ship first.

Relations

Enables #137 / #141 — a real fully-quantized-training topology Jan can evaluate against. Part of the SYM completeness program.

🤖 Generated with Claude Code