[kelkalot]

Fixes #124.

run_scenario passes the whole conversation as history, current turn included. On the conversation's message dicts, the engine keeps markers beside role and content:

• documents with their marks;
• file_uri;
• decision.

HTTPAppTarget._build_body forwarded this list unchanged. As a result:

• Duplicate turn. The current user turn reached the app twice. In messages mode it was appended again; in other modes it went out as message and again at the end of history.
• Marker leak. The markers reached the app, including the evaluation-only document marks.

examples/audit_openwebui_rag.py uses messages mode, so it was affected.

Changes

• History. History goes out without the engine's markers and without a trailing entry that is the current user turn. Other fields (tool_calls, tool_call_id, name, …) stay, so direct callers that pass earlier turns get the body they built. In message mode, history is a new list of the earlier turns.
• Documents.
  • A multi-turn scenario's documents are taken from the history marker and sent through documents_field as text, as in context-grounding runs.
  • They're parsed with parse_documents, like the model and decision targets: a dict without a string text fails, and empty documents keep their position, matching the judge's numbering.
  • An explicit documents=[] is respected.
• The history contract. Target.send now documents it, and history_documents() in targets/base.py replaces the copies in http.py and decision.py.

Behaviour changes:

• Documents without a destination. A multi-turn scenario with documents and no documents_field now ends in an ERROR that names documents_field, the same rule context-grounding runs follow.
• Attached files. A scenario with file_uri now ends in an ERROR; an HTTP app can't receive the file, and it used to be dropped silently.

Tests

Eight new tests in tests/test_targets.py. Through real audits against a recording HTTP client, they cover:

• each turn sent once in both body modes;
• documents sent as text with no marks;
• the missing documents_field and file_uri errors.

Direct _build_body calls cover:

• marker removal;
• OpenAI tool messages kept as sent;
• document parsing and positions.

Live check against a real OpenAI-compatible chat endpoint (Anthropic's, with claude-haiku-5-5), recording every request body:

• In a three-turn audit, each request carried every turn once and no SimpleAudit keys. On dev, the same run sent the current user turn twice in every request.
• Scenarios with documents and no documents_field, or with a file, stopped before any request was sent.

Not covered here

CallableTarget also passes history to the wrapped function unchanged, markers included.