Context
Grace's defect-note MESSAGE:54abadd1 (2026-10-23) comes from the #20017 warm-reload trace on dev 9cad83b350. A dock host moved between windows mid-projection (WorkspaceController#onWindowConnect: a silent remove, then target.add). It ended up mounted: true with no vnode, and a promiseUpdate() asked on it never settled. #20017's fix (PR #20032) removes that trigger. Promoted from the cut steward's lane list on #20022.
Design authority: the operator's challenge on PR #20044 (2026-10-23): mounted: true and having a vnode mean the same, so this is a state that should not exist. The ticket's first version asked to settle the state. The second version asked to report it. With the report in place, CI then found a live producer (below), so this version also removes that producer.
The Problem
The engine pairs the two flags on most paths. Every path that drops a child's vnode sets mounted = false in the same step: a parent's unmountRemovedChildren pass, which hide() (removeDom) and focus-gated action withdrawal reach through their owner's update, and the SortZone cross-window handover. A mount assigns the vnode before it sets mounted.
One path did not. It is a component whose OWN flight renders it away. resolveVdomUpdate syncs the flight root's reply vnode, which is null when that root carries removeDom. syncVnodeTree(null) then nulled _vnode, and nothing unmounted the component.
The report found it in CI at PR #20044's head 01bad08d3c. It fired in two grid e2e flows: grid/HeaderMenuStateNL "native selection, final uncheck and Clear…" and grid/IssuesShowcase "a filtered locked column keeps its header marked while its menu is closed". A local probe named the components: the grid search toolbar's filter summary and its Clear button. Both hide themselves with set({hidden, disabled}). hide() stamps removeDom, and the other config runs the component's own flight.
On dev the state is transient. A later parent flight pairs the flags again, so the end state converges. But an update or a promiseUpdate() asked in the window parks silently, and nothing is logged.
The Architectural Reality
src/mixin/VdomLifecycle.mjs syncVnodeTree syncs a flight root's vnode, and unmountRemovedChildren handles the children a flight removes. Before the fix, the flight root itself was never unmounted when its own vnode came back null.
updateVdom starts a flight for a mounted, initialized component only when !isParentUpdating && !isChildUpdating && vnode. With no vnode, nothing ran and nothing was said.
isParentUpdating(parentId, resolve) registers a post-update on a colliding parent flight. That path still takes such an update first and must stay.
- The parent's re-render sets
mounted = true again for a child it re-inserts (syncVnodeTree's child pass), so unmounting on a self-removal is symmetric.
The Fix
syncVnodeTree unmounts a component whose own flight left it without a vnode, as unmountRemovedChildren does for the children a parent's flight removes. A promise asked while it is away then waits as a not-yet-mounted one does, and the render that mounts it again settles it.
- Any remaining breach is reported. When a mounted component without a vnode asks for an update and neither a parent nor a child flight takes it,
updateVdom logs one console.error naming the component.
Acceptance Criteria
Out of Scope
- Settling a breach through an ancestor's flight. See Avoided Traps.
- Re-creating the cross-window trigger, which #20032 removed.
- #18802's not-yet-mounted exception.
Avoided Traps
- Settling the breach through the nearest rendered ancestor (the first version of PR #20044: register as its merged child and request its update). It made a state that should not exist work, at the cost of an ancestor walk in the engine. It would also have hidden the producer the report then found.
- Settling it in
syncVnodeTree by claiming parked callbacks. It is the same trade. The fix unmounts instead, which hands the promise to the existing not-yet-mounted contract.
Related
- #20017 / PR #20032 (the cross-window trigger, removed), #18787 (the
promiseUpdate contract), #18802 (the not-yet-mounted exception, kept), #20022 (lane list), PR #20044, Grace's defect-note MESSAGE:54abadd1
Live latest-open sweep: checked the latest 20 open issues at 2026-10-23T08:53:38Z; no equivalent. gh search issues "promiseUpdate never settles" finds only #18802 (closed, the not-yet-mounted case).
A2A claim sweep (all read-states, last 30): no claim besides my lane-intent MESSAGE:b068530b.
MC sweep: "promiseUpdate hangs never resolves on a mounted component whose vnode is null after moving between windows, callback unclaimed", 3 results. #18802's not-planned decision is the closest; it is a different state.
Own-assignment sweep: 8 open, none overlapping.
Origin Session ID: d373d94f-3d91-4a15-a181-4309706acfb5
Retrieval Hint: "promiseUpdate mounted true vnode null own flight renders it away removeDom syncVnodeTree unmount grid search toolbar clear button"
Context
Grace's defect-note MESSAGE:54abadd1 (2026-10-23) comes from the #20017 warm-reload trace on
dev9cad83b350. A dock host moved between windows mid-projection (WorkspaceController#onWindowConnect: a silent remove, thentarget.add). It ended upmounted: truewith no vnode, and apromiseUpdate()asked on it never settled. #20017's fix (PR #20032) removes that trigger. Promoted from the cut steward's lane list on #20022.Design authority: the operator's challenge on PR #20044 (2026-10-23):
mounted: trueand having a vnode mean the same, so this is a state that should not exist. The ticket's first version asked to settle the state. The second version asked to report it. With the report in place, CI then found a live producer (below), so this version also removes that producer.The Problem
The engine pairs the two flags on most paths. Every path that drops a child's vnode sets
mounted = falsein the same step: a parent'sunmountRemovedChildrenpass, whichhide()(removeDom) and focus-gated action withdrawal reach through their owner's update, and the SortZone cross-window handover. A mount assigns the vnode before it setsmounted.One path did not. It is a component whose OWN flight renders it away.
resolveVdomUpdatesyncs the flight root's reply vnode, which isnullwhen that root carriesremoveDom.syncVnodeTree(null)then nulled_vnode, and nothing unmounted the component.The report found it in CI at PR #20044's head
01bad08d3c. It fired in two grid e2e flows:grid/HeaderMenuStateNL"native selection, final uncheck and Clear…" andgrid/IssuesShowcase"a filtered locked column keeps its header marked while its menu is closed". A local probe named the components: the grid search toolbar's filter summary and its Clear button. Both hide themselves withset({hidden, disabled}).hide()stampsremoveDom, and the other config runs the component's own flight.On
devthe state is transient. A later parent flight pairs the flags again, so the end state converges. But an update or apromiseUpdate()asked in the window parks silently, and nothing is logged.The Architectural Reality
src/mixin/VdomLifecycle.mjssyncVnodeTreesyncs a flight root's vnode, andunmountRemovedChildrenhandles the children a flight removes. Before the fix, the flight root itself was never unmounted when its own vnode came backnull.updateVdomstarts a flight for a mounted, initialized component only when!isParentUpdating && !isChildUpdating && vnode. With no vnode, nothing ran and nothing was said.isParentUpdating(parentId, resolve)registers a post-update on a colliding parent flight. That path still takes such an update first and must stay.mounted = trueagain for a child it re-inserts (syncVnodeTree's child pass), so unmounting on a self-removal is symmetric.The Fix
syncVnodeTreeunmounts a component whose own flight left it without a vnode, asunmountRemovedChildrendoes for the children a parent's flight removes. A promise asked while it is away then waits as a not-yet-mounted one does, and the render that mounts it again settles it.updateVdomlogs oneconsole.errornaming the component.Acceptance Criteria
mounted: true,vnode: null,promiseUpdate()logs oneconsole.errornaming the component. Red ondev(nothing logged), green with the fix.addstill resolves, and the existing settle-timing contract arms stay green.vdom.removeDom+ its ownpromiseUpdate()) endsmounted: false.promiseUpdate()asked while it is away settles when its parent shows it again.dev.01bad08d3c(grid/HeaderMenuStateNL,grid/IssuesShowcase) are green with the fix, so the report no longer fires there.Out of Scope
Avoided Traps
syncVnodeTreeby claiming parked callbacks. It is the same trade. The fix unmounts instead, which hands the promise to the existing not-yet-mounted contract.Related
promiseUpdatecontract), #18802 (the not-yet-mounted exception, kept), #20022 (lane list), PR #20044, Grace's defect-note MESSAGE:54abadd1Live latest-open sweep: checked the latest 20 open issues at 2026-10-23T08:53:38Z; no equivalent.
gh search issues "promiseUpdate never settles"finds only #18802 (closed, the not-yet-mounted case). A2A claim sweep (all read-states, last 30): no claim besides my lane-intent MESSAGE:b068530b. MC sweep: "promiseUpdate hangs never resolves on a mounted component whose vnode is null after moving between windows, callback unclaimed", 3 results. #18802's not-planned decision is the closest; it is a different state. Own-assignment sweep: 8 open, none overlapping.Origin Session ID: d373d94f-3d91-4a15-a181-4309706acfb5 Retrieval Hint: "promiseUpdate mounted true vnode null own flight renders it away removeDom syncVnodeTree unmount grid search toolbar clear button"