docs: e2e verification + Keras from_logits loss-instance fix

docs/ARCHITECTURE_PARAM_OWNERSHIP.md

@@ -174,6 +174,15 @@ no-selection state) lets the user enable/clear training, pick an optimizer + LR
 loss, and shows the live `train_summary` projected from Rust. It dispatches
 `set_train_config` through the zustand store. `npm run build` (tsc + vite) passes.
 
+**End-to-end verification:** `examples/e2e_intents.rs` builds an MLP classifier
+(Input→Linear64→ReLU→Linear10→Output) entirely through `Session` intents — the
+exact vocabulary the UI uses — attaches a TrainConfig via `SetTrainConfig`, and
+exports both backends. The exported code was run verbatim: PyTorch trains
+(loss 2.32→2.00), Keras `compile()/fit()` trains (2.51→2.28). The Keras emitter
+now emits the compile-loss as a proper **instance with `from_logits=True`** for
+logit-producing losses (the graph emits raw logits) — the string form would
+default to `from_logits=False` and be numerically wrong.
+
 ## Test coverage
 98 Rust tests pass (0 warnings) — added `loss_mean_is_scalar_none_is_elementwise`,
 `cross_entropy_none_drops_class_axis` (nflow-ops) and `pytorch_loss_in_forward`,