[vicmaster]

Phase 30 slice G, the last slice. The phase began with a phone screen, "Ridgeline — Today", that had been hand-built in 93 nodes and scored 100/100 without a single mobile check running. This PR rebuilds that screen on the finished phase, through the real tools, and keeps the rebuild as a test. The re-run also surfaced two render defects that the gate could not see, and this PR fixes them.

The re-run

test-ridgeline.ts drives the MCP server over stdio and follows the spec's path:

1. canvas_create({ type: "mobile" }).
2. generate_design_system, with seed #277350 and the soft personality.
3. apply_structure("dashboard"), which stamps dashboard-mobile.
4. canvas_set_genre("data"). The honest-content tell's own suggestion names this genre for a fitness screen's own figures.
5. One batch_design batch.

The batch has no autofix, builds no device chrome by hand, and changes no rules.

	Result
Nodes	33 (the hand-built screen had 93, counted the same way)
canvas_evaluate, fast mode	100/100, READY TO PRESENT, three info findings (APCA advisories on secondary text in the dark theme)
canvas_evaluate, detailed mode	100/100, READY TO PRESENT, the same three info findings
Dark-theme contrast findings	none. A negative control proves the dark check is live: making dark secondary text match its surface produces dark errors.
canvas_stress, all five default perturbations	CLEAN, with 20 designed truncations (info)
mobile-handbuilt-target	0
surface.rulesOn	all six mobile rules plus mobile-artboard

Every tap target is a primitive: the bell is a nav-bar action, the four quick actions and "See all" are buttons, the runs are list rows, and the tabs and the start-run action are tab-bar items. The status bar, nav bar, tab bar and home indicator are the primitives.

The test also fails if the design system's fonts fall back, because a fallback face changes the glyph widths that every verdict above is measured on. It runs in the full tier.

The screen differs from the hand-built one in three places:

• The quick actions are a two-by-two grid of buttons rather than a row of four icon tiles. Four labelled 44px buttons don't fit one 358px row.
• The "FM" avatar is not rebuilt, because the nav bar has no avatar slot.
• The notification badge is not rebuilt, because nav-bar actions carry no badge.

Two render fixes the re-run needed (approved before building)

Large type on a phone rendered at 60%.

• The renderer scales any font size of 24px or more against a fixed 1440 design width. On a 390 artboard, $title (33px) rendered at 20px, smaller than the 21px card title, and $heading (27px) rendered at 16px, smaller than the 17px body. The rebuilt screen's headline figure read smaller than a card heading.
• Font sizes now scale against the root's own width, so a size renders as authored on its artboard and still shrinks on narrower viewports.
• A 1440 canvas emits exactly the CSS it did before, including in breakpoint renders.
• Padding keeps the 1440 reference, because the slice C safe-area rules model it.
• One visible change for existing work: a web canvas that isn't 1440 wide now renders its large type at the authored size. For example, a 32px heading on a 1280 page used to render at 28.4px. This belongs in the release notes.

Chart tick labels stretched with the plot.

• The chart's SVG stretches to fill its box, and its tick labels were SVG text inside it. On a width: "100%" chart in a phone card the labels came out squashed to about half their width.
• They are now HTML labels over the chart, at 11px at any width.
• Each x label has its own slot: its bar on a bar chart with one label per bar, otherwise the room halfway to the nearest labelled tick on each side. A label truncates with an ellipsis inside its slot, so dense labels never overlap or run past the plot.
• Empty labels skip their tick and leave the room to their neighbours. The docs now say to label every other tick on a narrow line chart.

A fix that followed from the font change

canvas_sync_from_url with a viewport override renders the imported page in a document as wide as the canvas root. Without this, the two sides scaled large type against different widths, and an in-sync page reported drift.

Docs

• README's screenshot_responsive paragraph and GUIDELINES' font-size and chart bullets describe the new behaviour.
• The batch_design docstring's chart line covers label slots.
• VISION ticks slice G and closes Phase 30.
• The spec gains slice G's build contract and an "As built" block, which records each departure, every verifier finding, and what changed because of it.

Testing

• New test-ridgeline.ts (full tier) and test-fluid-type.ts (no browser).
• New checks in test-chart.ts and test-sync-drift.ts.
• The full tier passes 116 of 117. The failure is test-text-props.ts, which hung in Chrome past the runner's timeout; it passes on its own in 3.5s. The same one-off hang happened in slice E.
• Verification ran in two rounds:
  • Round 1 used four agents: spec conformance, a render-fix adversary, a mutation pass and the docs steward. Their findings changed the test (the greeting and home indicator restored, and the assertions tightened), the label layout, and the sync render.
  • Round 2 used one agent on the fix diff. It found that half-step end slots truncated "Jan" and "Dec" on a fully labelled phone line chart, and that a 100% root synced above 1440 drifted. Both are fixed.
• The round-2 fixes were proven red on the previous commit and green after. No agent re-verified them. The sync change is a one-line pass-through in the tool handler. test-sync-drift.ts mirrors that handler rather than calling it, so the sync fix is verified by reading only.