Slate v2 post-closure architecture review

Decision

Do not pivot to Lexical, ProseMirror, or Tiptap.

The Slate v2 direction is still the best Slate-shaped architecture:

• Slate model and operations as collaboration truth
• Lexical-style editor.read / editor.update
• ProseMirror-style transaction and DOM-selection authority
• Tiptap-style extension ergonomics
• React 19.2 runtime using live reads, dirty commits, semantic islands, and projection sources

The honest post-closure status is stronger than "promising", but still not "impossible to regress".

What Is Strong

• editor.update is the right public write boundary.
• Primitive editor methods inside editor.update preserve Slate's flexible custom-node DX better than semantic method explosion.
• EditorCommit is the right local runtime fact for history, collaboration, React, DOM repair, and tests.
• Generated browser gauntlets now assert model state, DOM state, DOM selection, focus owner, commit metadata, kernel trace, and follow-up typing.
• Primary examples are guarded against Transforms.*, stale mutable editor fields, and direct method replacement patterns.

What Is Still Not Absolute

• Compatibility still exists. Direct primitive calls outside editor.update currently auto-transaction instead of failing everywhere.
• Transforms.* still appears in internal runtime, legacy transform tests, history tests, and compatibility lanes. That is acceptable only because the primary public surface is guarded.
• Public state mirrors are read-only, but still present. They are not the final cleanest possible API shape.
• Kernel authority is audited by inventories and contracts, but the source still contains explicit bridge/worker escape hatches. That is disciplined, not mathematically closed.
• The automated mobile claim is scoped. Semantic mobile handles and Playwright mobile viewport rows are not raw human-device keyboard, clipboard, glide, or voice proof.
• The React huge-document claim is reopened for typing: fresh 5000/10000-block comparisons show current shell islands winning ready/full-document lanes but losing steady typing/select lanes to legacy chunking-on. The kept direction is DOM-present auto by default and explicit shell mode behind proof gates.

Harsh Take

The architecture is the right one. The remaining risk is not "wrong engine"; it is proof maturity and escape-hatch pressure.

If future cursor regressions keep appearing, the answer is not another API pivot. The answer is to harden proof:

• real persistent-browser gauntlets, not only Playwright rows
• generated scenario expansion with shrinking and replay artifacts
• device-lab/Appium lanes before native mobile claims
• tighter allowlists for every Transforms.*, public mirror, repair bridge, and selection bridge use
• development/test assertions that fail illegal kernel ownership transitions at runtime, not only through inventory counts

Confidence

Architecture direction: high.

Implementation closure for the scoped release claim: high.

Regression-free browser editing across unknown user behavior: not high enough to claim perfection. Browser editing earns confidence through adversarial soak, real-device coverage, and dogfooding, not through a closed plan file.