fix(agent, llm): improve message handling to support image, document in extract_recent_actions

[msmakouz]

What was changed

• agent: part_text helper renders a message's content part — its .text when present, [image] / [document] for media, [non-text content] otherwise. Used in all four branches of extract_recent_actions
• llm: the mapper tracks whether the merge succeeded and appends a text part when no text part exists, instead of dropping the instruction

Why?

Any tool returning an image to the calling agent through the _images tool-result channel killed the agent process. wippy.llm:prompt.get_messages() strips _images out of the function result and re-injects the pictures as {role = USER, content = collected_images} — a message whose parts are all images, with no .text anywhere. Two separate defects fire on that message.

1. extract_recent_actions crashes the agent process "user: " .. message.content[1].text concatenated nil, killing the process mid-turn. The dataflow node was left running and the failure surfaced as COMPLETE_WORKFLOW must be the first command in a command batch — a downstream symptom, not the cause. Nothing reached the logger because the process was gone. Only agents with a memory_contract were affected: extract_recent_actions is reached only via perform_memory_recall, which returns early without one. The FUNCTION_RESULT branch already guarded this with or "nil"; USER, ASSISTANT and DEVELOPER did not

2. The Claude mapper silently drops the developer instruction agent:step appends the memory recall as a DEVELOPER message last — directly after that image-only message. should_create_new_message only tests for tool_result, so an image-only message takes the merge path; the merge loop scans for a text part, finds none, and falls through. The recalled memory was discarded with no error, on every turn following an attached image