feat: a wheel pulls the wall too, and the bar stays put on a desktop

[antstanley]

The desktop half of the pull gesture, and the bar change that made the indicator need to think about where the wall's top is.

A wheel pulls the wall

Keep scrolling up when the wall is already at its top and it opens exactly as a finger opens it: same band, same 72px threshold, same 54px shelf, same indicator, same feed.refresh(). Two things are forced by what a wheel is, and they are the only differences:

• It reports movement, not position, so travel accumulates and each delta is scaled by 0.4. One notch of a mouse wheel is 100px in chrome; at 1:1 a stray notch at the top of the wall would arm a hundred-author fan-out. At 0.4 it takes two deliberate notches, while a trackpad's many small deltas give a pull that answers continuously.
• It never says it is done, so stopping is the release: 140ms of quiet past the threshold lays the wall. The pill says "stop to lay again" rather than "let go", because telling somebody holding no button to let go is an instruction they cannot follow.

A wheel pull starts from rest, and that is the whole defence against momentum. A trackpad flick that reaches the top keeps delivering wheel events after the fingers have left the glass: upward, at scrollY 0, exactly the shape of a pull. What a momentum tail structurally cannot have is a gap in front of it, so a gesture may only begin after 200ms of quiet. A reader who means to pull always has that gap.

The two inputs ask "where" differently, and that is not drift: a finger grabs what is under it, so a touch pull must start inside main#wall; a wheel only points, and the cursor sits wherever it was last left, very often over the bar. A wheel is refused only over a screen with its own scroll.

The bar stays put on a desktop

From md it sticks to the top of the scroll rather than scrolling away, so the layout picker, the switcher, refresh and settings are reachable from anywhere in an endless scroll. Unchanged on a phone, where it is fixed to the bottom under a thumb. z-20 so the switcher's scrim still dims the bar its panel hangs off; the page's own colour at 80% behind a blur, so it reads as nothing at the top of a wall and veils the bricks that pass under it; no border, because a rule across an unscrolled page is a line for its own sake.

The two are one change rather than neighbours: a sticky bar moves where the top of the wall is, so the indicator now measures main#wall's own top edge when a gesture starts. Zero on a phone, the height of the bar on a desktop, and no bar height written down anywhere it could go stale.

Verification

just check (254 vitest, 159 cargo nextest, two tsc projects, four guards, lint, fmt) and just test-e2e (85 browser tests, five of them new).

Thirteen new vitest cases cover the momentum tail and the rule keeping the two releases apart: a finger resting mid-drag must not lay the wall, and a touchend cannot commit a wheel pull. The browser cases dispatch wheel events 16ms apart rather than driving page.mouse.wheel, because the timing IS what is under test: every mouse.wheel call is a round trip that puts tens to hundreds of milliseconds between events, landing on the wrong side of both timers.

🤖 Generated with Claude Code

[github-actions]

🧹 Preview wall removed.