Orchestrator marks soft-failed worker tasks as success=true #70

Problem

In execute_task, the Ok(result_str) arm after rig's multi_turn loop marks TaskStatus::Complete, success=true whenever the loop terminates without a rig-level error. The result text is never inspected. A worker that enters a duplicate call loop and writes "I was unable to complete the task" is indistinguishable from genuine success.

This means:

• build_task_context embeds soft-failure text as PRIOR_WORK for dependents
• ready_tasks() unblocks dependents that should be blocked
• The coordinator has no structural signal forcing a replan

Evidence

Live SSE captures from SRE integration testing show "success": true on tasks where the worker explicitly reported inability to complete, counted as task_failed_count: 0 in analysis.

Solution

After stream_and_forward returns, check the escalation flag set by the duplicate-call guard (#69). If set, run the worker response through maybe_create_artifact (inline if short, artifact pointer if long) and return Err so the task is marked Failed and dependents are blocked in the DAG. The coordinator sees the worker's findings via the existing reflection prompt and can read_artifact for long outputs.

Related code

• crates/aura/src/orchestration/orchestrator.rs — execute_task flag check after stream_and_forward
• crates/aura/src/orchestration/types.rs — ready_tasks() blocks dependents on TaskStatus::Failed