History entries with missing outputs: Imp invents a filler DSPy doesn't; native replay drops answers

[deepfates]

What's true now. When a History entry has no value for an output field, Imp renders that field as "Not supplied for this conversation history message." (lib/imp/adapter/chat.ex ~1170, introduced in d2e6881). The resulting assistant message puts invented words in the model's own voice.

DSPy 3.2.1 doesn't do this. Its ChatAdapter renders a missing history output as None (dspy/adapters/chat_adapter.py ~204, outputs.get(k, None)), which was confirmed by running ChatAdapter.format in tmp/dspy-parity-venv. DSPy's "Not supplied…" text is its filler for demos (dspy/adapters/base.py:451), not for history.

Separately, the native-tool replay path (render_native_tool_history_turn) shows only next_thought as the assistant's content. So a generic Imp.Predict that sets tool_calls_field and has no next_thought loses its recorded answer when history is replayed natively.

Question. What should a history entry with missing outputs render as? Three options:

• DSPy's None, for parity;
• the user message alone, with no assistant message, if the entry is not an exchange;
• the current filler.

Parity says follow DSPy unless there's a recorded reason not to, and there's no recorded reason for the filler.

Done when.

• The chosen rendering is tested and matches the parity differentials, or the difference is recorded with its reason.
• Native replay keeps a generic Predict's recorded outputs, with a test that fails on main.

Found while reviewing #260.