fix(tui): the sky dims with the room instead of going dark under the card [spec 10]

[wine-fall]

Implements specs/spec10/10-tui.md (§3.2-B spotlight amendment — the sky dims with the room).

What

Field report: whenever a consent/question card appears, the pixel figure and the ripple stardust vanish from the sky panel — the raster layer was deleted outright while hushed, because kitty images composite above text cells and could cover the card. A sky that goes dark under every ask reads as the interface breaking.

Now the rasters step down with the room instead of leaving:

• Figure: repaints faded toward the night by the same 0.55 step as hush() (HUSH_FADE, reusing encodeFigurePng's existing fade channel and cached per pose+state).
• Ripple: repaints in the hushed accent color, its raster height clipped so it ends above the card (waveRowsFor).
• Yield only on collision: the card's real footprint is computed by replaying the renderer's own width/wrap/chrome math (cardRows / cardTopRow in dock.ts, wrapping included); a raster is deleted only when the card's rows would actually reach it (stagePlan) — short windows keep the old behavior, everything else keeps the sky.

Verification

• pnpm test (root, includes the new cardRows/stagePlan/waveRowsFor unit coverage) and the TUI tree's own bun run typecheck (the CI tui job's command) — green; oxlint green.
• By-eye pass owed (sensory): on the next boot with a pending ask, the figure and ripple should stay visible but dimmed, and the card must stay fully readable — including on a small window, where the rasters yield.
• Peer review (codex gpt-5.5, xhigh): 1 finding, 1 applied (comment carried edit-history; reworded to current intent).

AI coding brief

• Original request: after live-testing the onboarding consent card, the user reported that the visuals behind the popup always disappear and asked for it to be handled.
• Manual interventions: one clarification round — "logo" was ambiguous between the MURMUR wordmark (actually still visible, just dimmed) and the sky rasters; the user confirmed it was the figure/ripple, which redirected the fix from the wordmark path to the raster paint loops.
• Retro: naming the on-screen regions in the clarifying question (with what the code says each does) resolved the ambiguity in one round; guessing from the screenshot alone would likely have fixed the wrong element.

🤖 Generated with Claude Code