Extend LLM interpretation response to suggest sample_input and use it as interactive default

[AK11105]

What problem does this solve?

During interactive deploy, the user is prompted for sample_input with no guidance on what format the model expects. For numeric models this is especially confusing — the user doesn't know whether to enter [1.0, 0.0, ...], a dict, or a scalar. The inspector already extracts n_features_in_, tensor shapes, and input hints that are sufficient to suggest a concrete value.

Proposed solution

Extend the LLM interpretation response schema (#16 ) to include a suggested_sample_input field alongside framework, load_format, and input_hint. No separate rule-based logic is needed — the LLM infers the suggestion from the same raw_facts it uses for interpretation.

Updated interpretation response schema:

{
  "framework": "sklearn",
  "load_format": "joblib",
  "input_hint": "numpy array, shape (n_samples, 20)",
  "output_hint": "integer class label",
  "confidence": "high",
  "question": null,
  "question_field": null,
  "suggested_sample_input": "[1.0, 0.0, 1.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0]"
}

The suggestion is shown as the default value in the sample_input prompt during interactive deploy:

? Sample input for validation: [1.0, 0.0, 1.0, 0.0, ...]
  (suggested from n_features_in_=20)

The user can accept or override. In --yes mode (#24 ), the suggestion is used automatically if no --sample-input flag was provided.

suggested_sample_input is always a JSON-serialisable string so it passes through _parse_sample_input (B1) correctly.

Alternatives considered

Rule-based suggestion logic (e.g. generate a zero-vector of length n_features_in_). This requires separate implementation per framework and still produces less useful suggestions than the LLM, which can account for feature semantics visible in feature_names_in_ or config.json.

Area

CLI (deploy / fix)