Slate v2 React Runtime Narrowing Batch

Goal

Run Phase 3 honestly:

• rebaseline huge-document browser lanes after the core fast path
• cut obvious slate-react subscription fanout before touching islands

Kept Changes

• editable-text-blocks.tsx
  • descendant binding collapsed to one selector
  • top-level runtime ids use structured comparison instead of string churn
  • text descendants pass resolved marks directly into EditableText
• editable-text.tsx
  • bound text state collapsed to one selector
  • resolved runtime-bound text can skip redundant selector work when parent already owns text and marks
• small presentational slate-react components switched to explicit createElement(...) runtime ownership so browser benchmarks can run against source reliably under Next dev

Verification

• pnpm install
• pnpm turbo build --filter=./packages/slate-react
• pnpm turbo typecheck --filter=./packages/slate-react
• pnpm --filter slate-react test
• pnpm lint:fix

Results

Huge-document gate lane (1000 blocks)

• current v2:
  • ready 470.20ms
  • type 13.78ms
  • select-all 2.74ms
  • paste 37.41ms
• delta vs legacy:
  • ready -160.17ms
  • type -6.90ms
  • select-all -71.49ms
  • paste -69.31ms

Chunking compare (5000 blocks)

• before:
  • ready 881.50ms
  • type 71.18ms
  • select-all 19.36ms
  • paste 96.68ms
• after:
  • ready 824.89ms
  • type 65.45ms
  • select-all 17.63ms
  • paste 93.44ms
• delta vs legacy chunking after the cut:
  • ready +2.75ms
  • type +47.63ms
  • select-all -23.33ms
  • paste -220.03ms

Chunking compare (10000 blocks)

• current v2:
  • ready 1458.56ms
  • type 166.76ms
  • select-all 43.59ms
  • paste 196.13ms
• delta vs legacy chunking:
  • ready +107.97ms
  • type +136.78ms
  • select-all -37.70ms
  • paste -430.56ms

Verdict

This batch was worth keeping.

It improved the large-document React lane without touching islands or chunking architecture. It did not erase the typing gap at 10000 blocks.

That means the next honest work is still Phase 3 follow-up or early Phase 4:

• narrower runtime invalidation again if there is still obvious fanout left
• otherwise semantic islands / active corridor / adaptive occlusion

What is not left:

• pretending the remaining gap is still mostly core
• pretending chunking is already obsolete at 10000