fix(agent): empty assistant turn never completes the node

[wolfy-j]

Problem

tool_calling = "auto" completed the agent node on an empty assistant turn: agent_result.result ~= nil is true for "" (the empty string is truthy in Lua), so a model that had just created its answer through CreateArtifact and then returned {"result":"","tool_calls":[]} ended the node "successfully" with a zero-byte result. none rejected "" but accepted whitespace. Downstream consumers (kickside/research deliver) received an empty report and recorded a false success. Report: "Kickside Research false-success with empty report and orphaned artifact".

Two adjacent defects surfaced while proving the fix at loop level:

• The NO_TOOLS_CALLED feedback observation was queued after the turn's durable yield, so the next prompt (built from persisted history) did not contain it; the model only saw it one turn late.
• Control artifacts were persisted without node_id, so they could not be attributed to the creating agent node.

Change

• has_usable_result: text needs a non-whitespace character, a structured table needs at least one field, nil/false never complete. Shared by none and auto.
• An empty turn in auto records FEEDBACK.EMPTY_RESULT (key <iteration>_empty_result); a nil result keeps NO_TOOLS_CALLED.
• check_completion reports whether it recorded feedback; finalize_iteration yields immediately in that case so the feedback is applied before the next prompt (yield waits for the orchestrator to apply the commit; a bare submit only queues it).
• process_artifacts sets node_id on the artifact row.

Tests

• check_completion_test: 19 cases across none / auto / any (empty, whitespace, nil, false, empty table, structured, numeric, pending tool calls, min_iterations, exit-tool hint).
• agent_empty_result_test (loop level, deterministic model stub): the model creates an artifact through a control tool, returns an empty turn, then answers once the feedback is in its context. Asserts the artifact is retained with node_id, exactly one empty-result observation exists, and the workflow output is the real answer. A second case runs all-empty turns to the iteration limit and asserts COMPLETED_FAILURE with no output.
• control_handler_test: artifact rows carry node_id.
• Full wippy test --profile sqlite suite: 994 passed, 0 failed (two consecutive runs).

https://claude.ai/code/session_01TLD4uyqDLGD9JfgZESaaXH