[diffs/edit] Programmatic edits join the undo timeline like local edits

[fat]

What

Fixes the largest bug cluster from the editor-behavior audit (7 pinned bugs, one root cause): history entries were frozen while applyEdits(..., updateHistory=false) bypassed the EditStack, so any programmatic edit left the stack holding stale offsets — undo corrupted text, redo replayed at wrong positions, and undo-to-exhaustion could never restore the original document.

Design decision

Programmatic edits are not a separate semantic class: a mixed programmatic/local sequence must behave exactly like the same sequence applied all-locally. Concretely, after any sequence, undo-to-exhaustion restores the original text and redo-to-exhaustion the final text, with per-step states matching the all-local run.

// before: undo corrupts ('ilot?' …); after:
doc.applyEdits(insert('!'), true);   // local  -> "base!"
doc.applyEdits(insert('>> '));       // programmatic -> ">> base!"
undoAll(doc);                        // ">> base!" -> "base!" -> "base"
redoAll(doc);                        // -> ">> base!?" — byte-exact both ways

How

• TextDocument.applyResolvedEdits always creates/coalesces/pushes an EditStack entry; updateHistory now only gates caller-supplied selection + undo-boundary metadata on the entry. Redo clears on every recorded edit, tracked or not.
• Editor.applyEdits always tracks; its updateHistory param is kept for API compatibility but deprecated.
• No recursion / no bypass leak: undo()/redo() always wrote through the private #applyResolvedEditsToBuffer and are untouched; that private method is the only non-tracking door. History reset remains document rebuild with a fresh stack.
• Composition unaffected: previews never touch the document (browser composes in DOM; compositionend commits once through the tracked path) — verified end-to-end in review.

Tests

• The 7 equivalence pins flip to permanent regression tests with assertions unchanged — they encode all-tracked reference runs, independently re-derived by a snapshot-based simulator during review (25/25 values matched).
• The 2 old-model companion pins are rewritten to the new contract; exactly one other test in the suite pinned the bypass (applyEdits default does not record undo) and is flipped with a cross-reference.
• Docs updated (applyEdits example + History section) — they described the old flag.
• Full diffs suite: 1278 pass / 0 fail (run twice, identical); typecheck + oxlint clean. README known-bug inventory: 32 → 25.

Review notes

The riskiest surfaces were checked specifically: recursive recording via undo/redo replay, bypass leaking publicly, composition previews entering history, redo-clear semantics on the coalesce path, metadata-less entries in selection restore, and version-counter drift — all clean. Second PR of the failing-test burn-down (after #986).

🤖 Generated with Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
pierre-docs-diffs	Ready	Preview	Jul 15, 2026 5:07pm
pierre-docs-diffshub	Ready	Preview	Jul 15, 2026 5:07pm
pierre-docs-trees	Ready	Preview	Jul 15, 2026 5:07pm
pierrejs-diff-demo	Ready	Preview	Jul 15, 2026 5:07pm

[ije]

@codex review

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 97812b4aed

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[chatgpt-codex-connector]

Keep editor selections coherent for false-history undo

When Editor.applyEdits(..., false) is used while the editor has a caret, this now creates an undoable entry but strips both selection snapshots here. Undo/redo then returns undefined selections, so Editor.#applyChange leaves #selections in the coordinates from the other document version; for example inserting a line above the caret with updateHistory=false remaps the caret down, but undoing the now-undoable edit leaves it on the non-existent lower line and the next input/focus can act on stale coordinates. Since this commit makes these edits user-undoable, the editor still needs some selection remap/restoration path even when caller metadata is omitted.

Useful? React with 👍 / 👎.

[ije]

LGTM