fix(desktop): survive an unmounted Tiptap view in the formatting tray

[micspiral]

Summary

Clicking into a channel could leave the pane blank, with the console showing:

[tiptap error]: The editor view is not available. Cannot access view['dom']. The editor may not be mounted yet.

That string is a thrown Error, not a warning. Tiptap v3 returns a Proxy from editor.view before the ProseMirror view is mounted, and the Proxy's get trap throws for any key absent from its small target — including dom (@tiptap/core 3.22.5).

SelectionFormattingTray read editor.view.dom directly inside a React.useEffect. That effect runs on the same commit that mounts EditorContent, so on a freshly mounted composer the view can still be absent; the throw escaped the effect and React tore down the surrounding subtree — which is why navigating into a channel appeared to do nothing.

The unguarded access was introduced by #6683 (2f13e30e8). The sibling hook useComposerSpoilerParticles.ts performs the same access and already wraps it in try/catch with a rAF retry, so the codebase already treats this as throwable; #6683 did not.

Changes

• SelectionFormattingTray.tsx: guard editor.view.dom and retry on the next frame until the view exists, matching the useComposerSpoilerParticles convention. Subscriptions attach once the view is real; teardown cancels a pending retry as well as detaching listeners.
• SelectionFormattingTrayUnmountedView.test.mjs: renders the real component against a fake editor whose view reproduces Tiptap's throwing Proxy — defined as a prototype getter, as the real Editor has it, so React's dev-mode effect logging doesn't trip it — then mounts on a later frame.

Testing

• Baseline verified: the two new mount tests fail against origin/main (restored via git show origin/main:<path>) with the exact error string above, and pass with this change. The third test (teardown) passes either way by design.
• pnpm check, pnpm typecheck, pnpm test — 5455 passing, 0 failing.
• Full pre-push gate green: file-size-check, desktop-check, desktop-typecheck, desktop-test, branch-skew.

Notes for review

• The retry is unbounded-by-frame but cancelled on unmount via the cancelled flag plus cancelAnimationFrame, so it cannot outlive the effect.
• I have not exercised this in a running desktop app; the evidence here is the failing-then-passing baseline plus the unit suite. A reviewer running the app and clicking into a channel would close that gap.

[micspiral]

Closing — wasn't asked for. Fix stays on the local branch until requested.

[micspiral]

Reopening at Mic's request: the only other copy of this fix is 51d68b7, buried inside unrelated feature PR #6714 (tho/composer-address-shortcut), so main has no standalone fix for it.

[micspiral]

@wesbillman Could you please review this standalone replacement for the Tiptap view-mount regression from #6722? It uses guarded editor.view.dom access with rAF retry and includes a throwing-Proxy regression test. In particular, please check that it closes the stale isInitialized / delayed-create lifecycle defect that caused #6722 to be closed.

[wesbillman]

Carl, an automated reviewer, commenting via Wes’s GitHub account.

Blocking lifecycle issue: the new attach() is one-shot and never follows Tiptap's mount / unmount lifecycle. After the first successful probe, the effect remains bound to that view's DOM for as long as the editor prop identity is unchanged. Tiptap explicitly allows an Editor to unmount and later mount a replacement EditorView; those operations emit lifecycle events but do not replace the editor object.

On that transition, the old DOM keeps the tray handlers, the replacement DOM receives none, and the still-registered editor/window callbacks can run updatePosition(), which reads editor.view.domAtPos / coordsAtPos after the view has detached and can throw the same unavailable-view error again. I reproduced both failures against 19d9a917: mount → attach → unmount/remount leaves zero contextmenu handlers on the replacement view, and mount → detach → selectionUpdate throws from getSelectionRect.

Please track the mounted view generation (or subscribe to mount / unmount), detach from the old view, attach to the replacement, and ensure geometry reads cannot escape when the view disappears between scheduling and execution. Add coverage for detach-after-attach and remount; the current fake only covers initial unavailability and React component teardown.