[BorisTyshkevich]

What & why

Closes #338.

On a long Dashboard, moving a tile from the end to the beginning previously required several separate drops-and-scrolls, because a move could only reach destinations visible when the drag began. This adds the standard board interaction: during an active tile move, holding the pointer near the top/bottom edge of the visible Dashboard viewport scrolls it continuously, and the destination follows the newly revealed tiles.

Extends the #332 grip / ⌘-Ctrl tile move — no native HTML drag, no change to ordering semantics or schemas.

How it works

• Pure, injected-seam module src/core/dashboard-autoscroll.ts (100/100/100/100):
  • edgeScrollVelocity(...) — signed px/frame from the pointer's distance to the visible edge; bounded, proportional acceleration; a lower bounded cap under prefers-reduced-motion; 0 in the center and on a degenerate viewport.
  • createDragAutoScroll(target, scheduler, opts) — a single-requestAnimationFrame-loop controller behind an injected DragAutoScrollTarget (visibleTop/visibleBottom/scrollBy/canScrollUp/canScrollDown) and FrameScheduler seam. At most one frame is ever pending; a stationary pointer keeps scrolling; leaving the edge zone or hitting a boundary idles the loop; stop() is idempotent and a frame already in flight can't mutate after it. A future nested scroll container can reuse the controller unchanged.
• DOM glue in src/ui/dashboard.ts wireTileDrag: builds the DragAutoScrollTarget over .dash-page (effective top = page top + sticky .dash-topbar height), injects real requestAnimationFrame, and stops it in the existing single cleanup() (so every exit path — pointerup / pointercancel / blur / Escape / rerender — tears it down).
• Destination stays correct while scrolling: the Use command-drag for Dashboard tile moves and add cell-detail preview to table/log tiles #332 gesture captures tile home rects once; those go stale as the page scrolls, so a currentRects() helper shifts them by -(scrollTop - scrollTop0) at all four resolution sites (onMove grid+flow, onUp grid+flow) and on the per-frame recompute. The floating dragged card is position:fixed (viewport-anchored) and is correctly left unadjusted. Persistence is unchanged: exactly one atomic move-tile on release, nothing during the gesture; a cancel changes neither order nor revision.

Both the Grid Tiles live-reflow path and the flow (Report / 2-column / 3-column) reorder path auto-scroll. No new bundled runtime dependency.

Tests

• Unit (tests/unit/dashboard-autoscroll.test.ts): edge-velocity zero/direction/accel/clamp/reduced-motion/degenerate/short-viewport; controller single-loop, stationary-keeps-scrolling, leave-zone idle, boundary idle, stop() cancel + idempotent, no post-teardown mutation.
• Integration (tests/unit/dashboard.test.ts): edge scroll down/up, threshold-gated non-start, destination recomputes from a scroll frame alone with no new pointermove, exactly-one move-tile after several frames (landing on the scroll-shifted slot), Escape cancels the loop + dispatches nothing, read-only never scrolls, sticky-topbar offset shifts the top edge zone, flow-engine scroll+reorder.

Reviewed by an adversarial pass (single-loop invariant, teardown on every exit, scroll-adjustment across all sites, non-vacuous tests) — no blockers/majors; the one minor it raised (untested nonzero sticky-topbar offset) is now covered by a dedicated test.

E2E (Playwright) not run locally (no engine in this sandbox) — relying on GitHub Actions.

Checklist

• npm test passes (per-file coverage gate: new core file 100/100/100/100; ui/dashboard.ts 100/92.48/98.82/100, pre-existing gaps unchanged)
• Tests added/updated in the same change as the code
• npm run build succeeds (single-file dist/sql.html)
• Layers kept honest: pure edge/loop logic in src/core/, DOM glue behind an injected seam in src/ui/
• No new runtime dependency
• CHANGELOG.md ([Unreleased]) updated
• Reconciled tracked work — Auto-scroll long Dashboards while Command/Ctrl-dragging tiles #338 is a discrete follow-on to the closed Use command-drag for Dashboard tile moves and add cell-detail preview to table/log tiles #332 and is not on the Roadmap to 1.0.0 #68 checklist, so no roadmap/ADR change was needed

🤖 Generated with Claude Code

---

Follow-up fixes from live verification on github.demo (same PR)

Live-testing the drag on a tall dashboard surfaced three real defects, now fixed and re-verified live:

1. Flow-layout tiles now float + follow the cursor (commit 1acf43b). They were position:static ("dim in place"), so the new auto-scroll scrolled the grabbed tile off-screen. Both engines now lift the tile to a floating follower (grid keeps its placeholder/FLIP reflow; flow has no slot grid).
2. Max-overlap tile-move commit (commit 1acf43b). The old rule ("cover ≥2/3 of the destination slot's area") was aspect-ratio-dependent and unreachable for short tiles — a 403×120 KPI can cover at most ~0.40 of a 403×296 slot, so it always snapped back and could never be moved. Replaced with max-overlap (commit to whichever slot the tile overlaps most; snap back only when it still overlaps its origin most). core/tile-reorder.ts, 100/100/100/100.
3. Fixed horizontal cursor/tile desync (commit eb6a3c3). The flow-float refactor reordered beginMove so the grid placeholder was inserted before the home rect was captured, displacing the card into the next CSS-grid cell → the floated tile sat ~one column right of the cursor (vertical fine). Now the rect is captured pre-placeholder; added a happy-dom regression test that models the grid displacement.

Verified live (github.demo, Google-auth, SHA-matched build): auto-scroll through an 11.6k-px dashboard; flow tile stays under the cursor while scrolling; grab point 0px offset on both axes; a KPI tile commits a move (index 55→54) where the old rule snapped back.

Known follow-up (separate, pre-existing): flow-layout KPI band tiles have no drag surface — filed as #340 (KPI tiles ARE movable in Grid Tiles).