grilling: question numbers restart at Q1 in a later round, colliding with answers already given

## What happens The `grilling` skill tells the agent to "number each question" and shows `❓ **Q1** - …` as the format. It never says whether numbering continues across rounds. In one `/wayfinder` session, the agent asked four questions labelled Q1–Q4 in the first round. The user answered tersely, as the format invites: ``` * q1 a * q2 a * q3 a * q4 as recommended ``` In the next round the agent asked four new, unrelated questions, again labelled **Q1–Q4**. After that, "Q2" pointed to two different questions in the same conversation. The user's earlier answers ("q2 a") and any new ones became ambiguous for both the user and the agent. ## Expected Numbers stay unique for the whole conversation. Round two continues from where round one ended (Q5, Q6, …), and a question that is asked again keeps its original number. Other sessions already do this on their own: they continue numbering (Q1–Q8, then Q9–Q15) and ask open questions again under their original number. So this is underspecified rather than consistently wrong. ## Likely trigger It happened during `/wayfinder` map creation, which runs grilling twice in a row: step 1 "Name the destination" and step 2 "Map the frontier. Grill again…". The second pass seems to have been treated as a fresh grilling session, so numbering restarted at Q1. #1093 describes a different side effect of wayfinder invoking grilling more than once. ## Suggested fix In `skills/productivity/grilling/SKILL.md`, make numbering explicit. The current text only says "number each question". For example: > Number questions sequentially across the whole conversation, never per round: if the last round ended at Q4, the next round starts at Q5. A question you re-ask keeps its original number. This holds even when grilling is invoked again in the same conversation (for example, by another skill). Optionally, change the template from `**Q1**` to `**Q **`, so that the literal `1` doesn't suggest every round starts at 1. ## Related - #789: inconsistent numbering *schemes* (Q-1 switching to S2.1b). Its suggested rule, "Use consistent numbering … Q1/Q2/etc", wouldn't prevent this, because restarting at Q1 is perfectly consistent in scheme. - #1115: later rounds refer back to questions by number only. It assumes numbers are unique across rounds, which this issue shows isn't guaranteed. When a number has been reused, even `Q3 (ranking model)` points at the wrong question. - #953: numbers skipped inside a batch (Q17 → Q19). Same family, different failure. - #45: "enforce numbered questions" (wontfix, models already number). Numbering within a round is fine; numbering across rounds is not. ## How it was found I scanned the local transcripts of 86 sessions that used `grilling` or `wayfinder`. I looked at the `❓ **Qn** -` headers and checked for numbers reused for different questions in later rounds. One session restarted numbering at Q1 for new questions. Every other repeated number was the same question asked again.