c10 · Review stop 1

Branch claude/guide-walk-fixes-f9dd48, on main's e4fc308 (this morning's New to Qpoint review). Not pushed, not merged. Each item's outcome is under it in the report.

What to look at

1. The agent view (dark, light). The tour's MCP step is driven by an agent. There's no panel, and the list's cut-out, the step's callout ("sentry: 8 tool calls…") and the agent's ring and callout on postgres-readonly are all drawn. The pill reads Driven by Claude Code · Show the tour. Is that pill the right way back for the person, and are its words right?
2. What the panel was hiding (Models step). With the panel gone, the Spend card and its callout show in full. At 1120 wide the Spend card used to sit under the panel.
3. The panel on the agent's word (capture). This is set-view { panel: true }. It also shows the new synthetic note at the top of the step: "Synthetic data: these numbers come from a made-up machine (world meridian-day90), not this one. Where the tour says "this machine", it means that one." Is the wording right, and should it sit above the lede?
4. A row's callout (Opus, last row, postgres-readonly). The pointer now sits on the ringed row's own edge.
   • Neither row has room below at this size, so both callouts go above, lined up with the row's end.
   • They cover the neighbouring rows' numbers (calls, spend), never their names.
   • Is that acceptable, or should the page scroll to leave room below the ringed row?
5. Free-form pointing now dims the page. Out of a tour, point and highlight draw the same dim as a step, with what's pointed at cut out. In dark mode that's the lighter scrim (the Opus capture).
6. Padding. The agent's callout padding is now 12 / 16 / 16 (top-bottom / left / right), up from 8 / 8 / 14. The step's callouts are 14 / 14 / 24, up from 12 / 12 / 20. Both use the radius scale, since there's no spacing token yet. The step callouts you reviewed this morning are otherwise placed as before.

Calls for Mark

• Item 5: Skills read 0 calls (design call, not built). No harness reports skill use at runtime yet, so the count is zero rather than unknown. There are three options:
  • (a) show — with a tooltip saying "not reported by any harness yet";
  • (b) show "not reported" in the cell;
  • (c) keep 0 and let the tour and explain say it.
  
  The guide's Last used concept already says "never until a harness reports skill use".
• Item 1e: does the app grow a judgement surface? The guide now answers "is it safe?" honestly. A concept says this app gives no verdict, and that judgement lives in policies and findings on the org's plane. Should the native app show anything of that (a finding count, a policy line on an MCP server's page), or does it stay on the plane?
• The driver pill's place. The pill sits bottom left over the page, and in the agent view it carries Show the tour, so it's wider. In the captures it covers the first letters of a row name (linear). Should callouts keep clear of it, or should it move (for example, into the footer)?
• The key column is a guess. A row's "name" is taken as the first 40 % of the row. That holds for every listing today. A table could instead record its first cell as a region, but the Info View would then answer for the cell rather than the row.

Not done

• Not driven live through MCP by an agent on the running app. The tests drive the same runtime calls headlessly.
• No strobe numbers (R7). Nothing here touches the budget beyond one more hole in the scrim.