Reference model and training loop for the elicitation path

[ajtritt]

Part of #43. Depends on #45 (sample contract) and #47 (package scaffolder).

When the user has no model code (tier 4 of the evidence hierarchy, pure elicitation), the builder has nothing to verify the consumer contract against, and the user receives a Dataset with no demonstration that it feeds anything. Both problems are solved by generating a minimal reference model alongside the training loop, so the deliverable is a working example rather than an untested class.

The reference model is a contract witness, not an architecture recommendation. Its purpose is to prove tensors flow, to make the model-forward check runnable in every case, and to give the user a working starting point to replace. It must be labeled that way in the generated code and the README, or a domain scientist will reasonably read it as DSAgt endorsing an architecture for their science.

Scope

• Emit examples/train_minimal.py in every case. When the user has model code, it imports their model; otherwise it imports the generated reference model.
• Emit examples/model.py only when the user has no model code. Smallest architecture that consumes the declared contract, chosen from the contract's shapes and semantic roles.
• Sample kinds covered in v1: array, tabular, and graph. Array and tabular samples get an MLP head over flattened inputs. Graph samples get a two-layer message-passing network.
• The graph reference model is implemented in plain torch using Tensor.index_add_ for neighborhood aggregation, not torch_geometric. PyG is a heavy dependency with compiled extensions version-matched to the torch build, and a reference model that fails to import defeats its own purpose: the acceptance criterion is that the example runs on CPU in under a minute, which it cannot do if the user first has to solve a PyG installation.
• The exception is a contract that already commits to PyG (as the XGC case does, where samples are torch_geometric.data.Data objects). There the collation is PyG's, so the reference model uses PyG too and the dependency is already the user's, not one the builder introduced.
• The training loop runs a handful of steps over a small subset, prints the loss, and exits. It is a smoke harness, not a training script: no checkpoints, no schedulers, no logging integrations.
• Both files carry a header comment stating the model is a placeholder for contract verification.

Acceptance criteria

• python examples/train_minimal.py completes on CPU in under a minute against the project's data and shows a loss value.
• The model-forward check in Built-in check-dataset code #46 passes using the generated model when no user model exists.
• The generated model is derived from the contract, so a contract change that alters input shape produces a correspondingly changed reference model.