Slate v2 Large-Document Shell Proof Batch

Goal

Run the first honest shell-only large-document proof:

• far islands stop mounting editable descendants
• top-of-document typing still works
• promotion is local
• default runtime stays unchanged when the layer is off

Kept Work

• added internal large-document planner and shell renderer:
  • create-island-plan.ts
  • classify-island-kind.ts
  • island-shell.tsx
• wired EditableBlocks to support opt-in shelling through editable-text-blocks.tsx
• added query-param control in huge-document.tsx
• added the dedicated proof lane: replacement-huge-document-islands-benchmark.mjs
• restored the kept breadth lane by fixing its JSX-runtime assumption in slate-react-rerender-breadth-benchmark.tsx

Critical Fix

The shell experiment would have silently corrupted document text if DOM reconciliation kept assuming every top-level subtree was mounted.

The kept fix was to make structured DOM reconciliation update only mounted top-level islands and preserve shelled top-level subtrees from the committed snapshot.

Without that, typing in one live island would blank out far shelled text.

Result

Default runtime stayed green

pnpm bench:replacement:huge-document:local at 1000 blocks:

• ready 527.95ms vs legacy 720.87ms
• type 12.48ms vs legacy 19.85ms
• select-all 3.59ms vs legacy 75ms
• paste 34.85ms vs legacy 107.75ms

Breadth lane stayed green

pnpm bench:react:rerender-breadth:local still reads:

• #5131: broad hook still broad by contract
• #3656: sibling leaves 0
• #4141: ancestors 0

Shell proof moved large-doc cost hard

pnpm bench:replacement:huge-document:islands:local at 10000 blocks:

• ready 527.86ms
• top typing 38.85ms
• shell promotion 54.10ms
• shell count 99

Compared with the last clean non-islands 10000 baseline:

• ready had been 1229.49ms
• top typing had been 97.47ms

So the first shell-only proof cut:

• ready by about 701.63ms
• top typing by about 58.62ms

That is real enough to keep.

Verdict

The next large-document strategy is finally better than chunking-shaped storytelling.

What this batch proved:

• shell-only far islands can remove real runtime work
• the win is large enough to justify continuing
• local promotion is cheap enough to be a viable path

What it did not prove:

• broad ops are solved
• shell-backed selection is solved
• far-shell paste is solved

So the next batch is obvious:

• Phase 2 and 3 from the proof-first plan
• promotion into editable DOM on interaction
• model-driven select-all and paste over shell-backed ranges