Fixes #31156

Two independent main-thread fixes found while investigating Sentry ui.long-animation-frame spans ("Main UI thread blocked") — 826 spans over two weeks, p99 1461ms, max 4304ms, 153.4 s of total blocking.

1. TipTap editors were never destroyed

useCustomEditor creates a ProseMirror Editor in an effect, but cleanup only flipped a local isMounted flag — instance.destroy() was never called (a grep across src/ matches only test files). The view, plugin stack, contenteditable DOM and MutationObserver outlived the component.

It compounds because the editor mounts once per table row — the shared description column (utils/TableColumn.util.tsx) renders RichTextEditorPreviewerNew, which mounts a BlockEditor.

The isDestroyed guard mirrors @tiptap/react's own useEditor cleanup (node_modules/@tiptap/react/dist/index.js:1013, :1039) — <StrictMode> double-invokes effects, so cleanup can run twice.

Verified at runtime

Production build served against a live backend, driving SPA navigation (pushState + popstate) so components unmount without a page reload — a full reload resets the count regardless and proves nothing:

round 0  tags: 16
round 1  my-data: 9    tags: 16
round 2  my-data: 9    tags: 16
round 3  my-data: 9    tags: 16
round 4  my-data: 9    tags: 16

.ProseMirror count returns to exactly the row count every round and drops 16 → 9 on leaving. No compounding.

2. ELK layout engine loaded on every authenticated page

EntityLineageLayoutUtils imported ELKUtil at module scope. That module also exports five pure geometry helpers, so importing any one of them dragged elkjs — 1.37 MB built.

LineageControlButtons imports only centerNodePosition, and is reached statically from Lineage.component.tsx. That was enough to pull elkjs into the shared chunk that 476 route chunks statically import. Every authenticated page downloaded and parsed the lineage layout engine — including glossary, which uses G6/antv-dagre and never calls ELK (OntologyExplorer chunk: 0 occurrences of org.eclipse.elk).

ELKLayout is referenced in exactly one place, inside getELKLayoutedElements, which is already async and already awaited at every call site. Moving the import inside it is behaviour-preserving; the ELK engine itself is untouched.

Measured on a production build

−1.37 MB on every authenticated route.

Verified against a live backend

• Glossary terms tab: 0 elk-related requests, page still renders (7 rows, 7 editors)
• /lineage: fetches ELKUtil + vendor-elk on demand, renders 83 nodes, 0 console errors

3. Destroying the editor exposed two unguarded deferred writes

Fix 1 above turned a latent bug into a live one. BlockEditor defers two pieces of
work onto the editor with setTimeout (to dodge a known TipTap flushSync
warning), with no clearTimeout and no re-check inside the callback:

• the content sync (setEditorContent)
• the editable toggle (setEditable)

Both effects check editor.isDestroyed at the top, but the timer fires later.
Before this PR the editor was never destroyed, so a stale timer harmlessly wrote
into an orphaned-but-alive editor. Once the instance is destroyed on unmount, the
same timer reaches a torn-down ProseMirror view:

TypeError: Cannot read properties of null (reading 'updateOuterDeco')
  at EditorView.updateStateInner (prosemirror-view)
  at updateState (src/utils/BlockEditorUtils.ts:195)
  at src/components/BlockEditor/BlockEditor.tsx:235

This is reachable in the app, not just in tests: any description cell whose
content or editable prop changes and then unmounts within the same tick —
pagination click, tab switch, search keystroke, autosave round-trip — throws from
a timer callback, outside React, where no error boundary catches it. That is
exactly the high-churn table this PR targets.

Both sites now clear the timer on cleanup and re-check isDestroyed inside the
callback. Clearing also drops a stale content write that could otherwise land
after a newer one.

Caught by CI, confirmed by A/B

ui-coverage-tests went red on EntitySummaryPanel and TestDefinitionList,
both with that identical stack, while the same job is green on every recent main
run. A/B on one stack, changing only the destroy() hunk:

A sweep of the UI for other post-destroy access paths found none: only three files
combine a TipTap editor with deferred work — BlockEditor (fixed),
useCustomEditor (its rAF is already isMounted-guarded), and FeedEditor
(Quill, not TipTap). No extension NodeView defers editor access.

Type of change

• Bug fix (non-breaking)
• Performance improvement (no functional change)

Tests

128 tests pass across the BlockEditor tree and the two suites that caught the
regression (TestDefinitionList, BlockEditor/**), plus the lineage areas
(EntityLineageLayoutUtils, Lineage, EntityLineage, LineageProvider).

useCustomEditor's own suite now covers the unmount path. Its MockEditor had
neither destroy() nor isDestroyed, so the new branch threw
TypeError: instance.destroy is not a function on every unmount — React swallowed
it and the suite still reported green, which is why the first pass of this PR
missed the regression above. The mock now implements both, and a new test asserts
the instance is destroyed on unmount; it is non-vacuous (it fails against main's
hook and passes against this branch's).

Suites that fail to run locally with "Jest encountered an unexpected token" are a
pre-existing environment issue with the ui-core-components prebuilt dist; they
run fine on CI. I A/B'd them with the change removed and they fail identically.

tsc --noEmit: 0 errors in all changed files. prettier --check: clean.

Not verified: ESLint could not run locally — this branch's eslint.config.mjs requires eslint-plugin-sonarjs, absent from the checkouts available here. Relying on CI.

Manual test steps

Editor leak: open a table of descriptions (/tags/General, a glossary terms tab), run document.querySelectorAll('.ProseMirror').length, navigate away and back within the app (don't reload), re-run. Before: grows. After: returns to the row count.

ELK: open a glossary page with DevTools Network filtered to elk — no requests. Then open /lineage — ELKUtil and vendor-elk load, and the graph renders.

Checklist

• Self-reviewed
• Verified at runtime against a live backend, not just unit tests
• Confirmed no test regressions via control run and per-change A/B
• UI screen recording — N/A, no visual change

🤖 Generated with Claude Code

Summary by Gitar

• Performance:
  • Dynamically import ELKUtil in EntityLineageLayoutUtils to prevent bundling elkjs (~1.37MB) across all authenticated pages

_{This will update automatically on new commits.}