Fix CKEditor 4 typing after frame rewrites, a watchdog restart, and a lost arrow-key choice

[bartekplus]

Problem

The e2e test ckeditor4 typing accepts a prediction mid-line and keeps the caret after the word failed on master in Firefox (run 37345187610). The test is new in #476 and it is flaky.

Under CPU load I reproduced it locally: 7 of 120 CKEditor 4 typing runs in Firefox failed. In each failure, the prediction menu was visible, but Tab did not accept the prediction. An event log in the editing frame showed that nobody handled the Tab key, so Tab moved the focus out of the editor.

Root cause

CKEditor 4 writes its editing frame with document.open(). It does this at start and on each setData(). document.open() erases all event listeners of the document and its window, and it replaces the root element. The iframe, its window and the Document object stay.

• The MAIN-world bridge adds its own listeners again, but it sets itself to "disabled" and waits for the content script to send "enabled" again. The content script did not know about the rewrite, so it did not send it. The bridge then ignored the write request for the accepted word.
• In Chrome the same rewrite also kills the content script: all its listeners and its root observer are gone. After editor.setData(), no prediction showed at all.

In Firefox it is flaky because the order of "content script starts" and "CKEditor writes the frame" changes with CPU load.

Fix

The content script observes the Document node. A MutationObserver on the document stays after document.open() (I measured this in Chrome; the event listeners do not stay). When the root element changes, the content script destroys its instance and starts a new one on the new document. The new instance adds all its listeners again and sends "enabled" to the bridge again. A late config answer does not enable the destroyed instance.

Tests

• New e2e test: CKEditor 4 setData rewrites its frame with document.open, and typing still accepts a prediction. In Chrome it fails without the fix (the same "Tab not accepted" failure as the master flake) and passes 10 of 10 runs with it.
• New unit test in tests/content_script.behavior.test.ts. It fails without the fix.
• Coverage matrix: new behavior rewritten_document_restart.
• Local: bun run check, bun run test, bun run check:e2e:coverage, bun run test:e2e and bun run test:e2e:full --platform=chrome pass.
• Firefox: my local shell cannot start Firefox, so a temporary workflow ran the CKEditor 4 typing tests 20 times per job under CPU load (run 37352747215): 600 of 600 Firefox runs and 100 of 100 Chrome runs passed. Before the fix, about 6% of the local Firefox runs under CPU load failed. The last commit removes that workflow.

Two more fixes (found while I checked the CI of this PR)

2. Chrome: a watchdog restart closed the menu (ckeditor4 typing expands a snippet)

One CI run failed in Chrome: the menu showed "Best regards", but Tab did not expand ftsig.

• Cause: in Chrome, the content script in the CKEditor 4 frame can start while CKEditor still writes the frame (readyState: "loading", no <body>). Then the runtime observes <html>. Later HostChangeWatcher.watchDog() saw "observed node ≠ document.body" and restarted the runtime. The restart closed the menu. A restart between "menu shown" and Tab sent Tab to the browser.
• Fix: if the observed node is still in the document and contains the body, the watchdog observes the body and does not restart. A replaced root still causes a restart.
• Evidence: 4 parallel Chrome instances under full CPU load: before, 9 of 120 snippet runs failed; after, 120 of 120 passed, and all CKEditor 4 tests passed 288 of 288. New unit test.

3. A late answer moved the menu selection back to the first row

• Bug: the user pressed ArrowDown to a suggestion, then a late prediction answer came and highlighted the first row again. Tab then inserted a word that the user did not choose. Reproduction: type th, ArrowDown to that, a new answer comes, Tab inserts the .
• Fix: the keyboard handler records the chosen suggestion. A new answer keeps it highlighted when the answer contains it. An edit clears the choice.
• Tests: new e2e test A late prediction answer keeps the suggestion that the user chose with an arrow key (failed before the fix, passes 5 of 5 after) and two unit tests.

Not in this PR

• The MAIN-world start script (early Tab bridge) also loses its listener after document.open(). In a rewritten frame, the content script's own key handler accepts with Tab, so I did not change it.
• The CI log from master also showed a lost space ("We saww teh") once. I did not reproduce that. If it comes back with this fix, it needs its own look.

🤖 Generated with Claude Code

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.
To continue using code reviews, you can upgrade your account or add credits to your account and enable them for code reviews in your settings.