[akanyaani]

Fixes for the review findings agreed before open-sourcing. This branch includes #5 (citation and bidirectional results), so merging this PR covers both.

• Missing data fails first. train checks its data files before loading any model and prints the exact decisionformer data split … command when recipe data is missing. Previously it downloaded the model, then failed with a traceback and left an orphan output folder.
• One-line config errors. For example, error: unknown setting train.learning_rate (did you mean train.lr?). An override without = is rejected instead of silently ignored.
• Serving limit. Exports record the length they were trained on in decision_config.json (max_length). Loaded and served models refuse longer requests (HTTP 422, never truncated) unless max_length= / --max-length raises the limit. Previously a 5,088-token request to a model trained on 1,024 tokens was answered silently.
• Stricter requests. Every question needs a type (previously it silently defaulted to choice), and the server requires state.
• README:
  • Defines confidence (1 − normalized entropy, not the top probability).
  • Lists the GitHub install first; PyPI comes once v0.2.0 is published.
  • Says the backbone tests use tiny models, and lists the real checkpoints actually trained.
  • Notes CPU latency and that the Decision Index's 1,000 ms latency rule needs a GPU.
• deciformer15k_qwen3 saves resumable state every 125 steps; with 375 steps and the 500-step default, it never saved before.
• decisionformer --version.

Not included: the fp32 vs bf16 gap between the training metrics and eval on the export, which needs investigation.

Testing: the full suite passes on the GPU box with the GPU hidden. A smoke test of decisionformer train local_smollm2 in an empty folder fails instantly with the data split command and leaves nothing behind.