Motir

Estimate: 25m · Depends on: 3.3.5, 3.3.7

A real crash in the shipped 3.3.5 swimlane board, found by the 3.3.7 closing E2E (the cross-lane reassign journey) and logged as PRODECT_FINDINGS #61. Dragging a card from one swimlane into another intermittently throws React "Maximum update depth exceeded" in LaneCell → the 3.2.5 useRowWindow windowing primitive, crashing the board mid-drag (so the reassign never lands). It is a timing race — the drag sometimes settles before the loop trips — which is exactly the kind of flaky crash a real product cannot ship.

Root cause. Unlike BoardColumn (the flat board), which hands useRowWindow an explicit getScrollElement (its own internally-scrolling body, a STABLE viewport), LaneCell calls useRowWindow({ count, estimateRowHeight, gap }) with no getScrollElement — so the hook falls back to findScrollParent(container) and observes whatever ancestor scrolls (the page / main scroller, or the swimlane board’s overflow-x-auto region) via a ResizeObserver. During a drag, relocateCardToCell mutates the cell’s cards on every onDragOver, re-rendering it; recompute’s "content fits the viewport → render-all, else window" decision then oscillates (windowing → absolute positioning collapses the cell height → ResizeObserver fires → render-all → height grows → ResizeObserver → window …), each flip a synchronous setRange/setMeasured that triggers the next, until React’s update-depth guard throws. The flat board does NOT loop because its viewport height is fixed. The unit tests miss it because useRowWindow degrades to render-all under happy-dom (no layout) — only a real browser surfaces it, which is why only the E2E caught it.

The fix (durable shape, decision-ladder rung 2 = match the working flat-board pattern). Give LaneCell a STABLE, explicit scroll viewport the way BoardColumn does — OR, since a single (column × lane) CELL holds only a small slice of a column (a cell rarely overflows its own slot), disable windowing in LaneCell by passing a getScrollElement that returns null (the hook’s sanctioned render-all degrade): per-cell virtualization buys little and is what introduces the oscillation. If windowing is kept, the viewport element must be stable AND recompute must not be able to flip windowing on/off in a feedback loop (guard setRange against no-op/oscillating transitions; confirm measureElement can’t feed an oscillating height back in). Either way the board’s per-column virtualization on the FLAT board (BoardColumn, 3.8.3) is unchanged — this is a LaneCell/useRowWindow-only fix. No new dependency (the finding-#57 "no second virtualization library" rule still holds).

Un-quarantine the regression test. The reproducing E2E already exists, shipped by 3.3.7 as test.fixme in tests/e2e/board-swimlanes.spec.ts ("a cross-lane drag reassigns the assignee with no status change; a drop into the catch-all unassigns"), body intact and verified to drive the path end-to-end. This subtask flips test.fixme → test and proves it green over the real stack — that is the acceptance proof the fix worked (hence the 3.3.7 dep: the test file must be on main first).

Acceptance criteria

• A cross-lane drag on the swimlane board (group-by Assignee/Epic/Priority) no longer throws "Maximum update depth exceeded"; the board never crashes mid-drag, repeated drags are stable.
• The fix is confined to LaneCell / useRowWindow (a stable explicit scroll viewport, or render-all in cells); the FLAT board’s BoardColumn virtualization (3.8.3) is unchanged; no new virtualization dependency is added.
• tests/e2e/board-swimlanes.spec.ts’s cross-lane reassign test is flipped from test.fixme back to test and passes green (reassign field changed, status unchanged; catch-all drop unassigns) over the real stack — alongside the existing group-by / collapse / WIP tests.
• pnpm test (the 3.3.5 component reducer + lane-render suites) still passes; pnpm typecheck / lint clean.

Context refs

• PRODECT_FINDINGS.md #61 — the full root-cause + suggested-fix write-up
• app/(authed)/boards/_components/LaneCell.tsx — the useRowWindow call with no getScrollElement (the bug); app/(authed)/boards/_components/BoardColumn.tsx — the working pattern (passes an explicit getScrollElement) to mirror
• components/ui/useRowWindow.ts — recompute / measureElement / the windowing decision + the render-all degrade path
• tests/e2e/board-swimlanes.spec.ts (3.3.7) — the quarantined test.fixme to un-quarantine; tests/components/board-swimlanes.test.ts — the reassign reducer coverage that already passes
• motir-core/CLAUDE.md — finding #57 (no second virtualization lib); the --el-* / shape-token rules