Agent eval scenarios for Creating a Task and Answering a Question, on a fixture current with the entity tier #3763

Overview

The agent eval fixture (packages/agent/goldens) predates the entity tier and the workflow procedures, so it exercises none of the skills shipped by #3722 and #3719 (Creating a Task, Answering a Question, Asking the User, Splitting a Plan into Tasks). #3722 closed with its last acceptance criterion unmet: "the agent evals on the scenarios the change touches".

Problem Statement

• The confusion matrix and the live retrieval guards measure which skill is found, not whether an agent that has it follows it: checks before creating, a bug reproduced before it is shaped, a question put to the user rather than answered by the agent, a declined request left as an accepted decision.
• Without scenarios, a change to either procedure's wording has no behavioural check.

Proposed Solution

Add eval scenarios for the new procedures and bring the fixture up to the current model, then record a baseline:

• Creating a Task: a bug report is searched for (built and declined), reproduced or reported as not reproducible, then created with a brief, a checklist and its project, and is in Ready tasks.
• A declined request: an existing accepted decision that declines it is found, and no task is created.
• A question that stops work: a question task is created assigned to the person, blocking the stopped task, which keeps its status and assignee.
• Answering a Question: the agent puts the question to the user and records the answer as theirs; it never answers it itself.

Acceptance Criteria

• The fixture seeds the entity tier and the shipped workflow queues and skills
• One scenario per bullet above, with a pass condition that does not depend on wording
• A baseline run recorded (3 reps on these scenarios, then one full smoke run), with the contention check from the eval runbook done first
• Any scenario that fails is either fixed in the procedure text or recorded as a known gap with the reason

Related Issues

• Shape a standalone task when it is created, and let a task wait on a person (waiting status and waiting_on) #3722 (shipped the procedures; this is its unmet evals criterion)
• Adapt the community spec, plan-review and ticket-splitting workflows into NodeSpace planning skills #3719 (Asking the User, which both procedures name)