Slate v2 React-Perfect Runtime Optimization Plan

Goal

Make the slate-v2 rewrite earn the “React-perfect runtime” claim honestly.

That means:

• the React runtime is selector-first, predictable, and local by default
• huge documents do not need legacy chunking as the main rescue story
• the core does not sabotage the runtime with absurd transaction and read costs
• perf claims are backed by stable lanes, not heroic one-off reruns

Harsh Current Read

The current situation is blunt:

1. richtext blocker is fixed enough for the current curated RC gate
2. explicit normalization is no longer the big cliff
3. core huge-document typing is still bad
4. core read-after-write observation is still bad
5. legacy chunking still beats current v2 badly on huge-document typing at 5000 and 10000 blocks

So the next work is not:

• more normalization tinkering
• more richtext polish
• more drift census work
• or pretending chunking is irrelevant just because the docs say it should be

The next work is:

• fix the core/runtime architecture enough that the React runtime can actually become the right story

Non-Goals

Do not do these in this plan:

• import slate-batch-engine
• make child-count chunking foundational again
• widen the public contract just to hit a benchmark
• chase low-value microbench wins that do not move real user-facing lanes
• add another giant hook surface to paper over runtime leaks

Current Evidence

Runtime/blocker surface

From api-drift-perf-scoreboard.md:

• richtext blocker is green again
• placeholder is green
• markdown/table are green
• huge-document remains mixed because typing is still slower

Core-only evidence

From 2026-04-10-slate-v2-core-perf-batch.md:

• explicit normalization:
  • adjacent-text 134.20ms -> 3.08ms
  • inline flatten 1779.73ms -> 10.36ms
• core huge-document typing at 5000 blocks is still bad:
  • start-block 370.97ms -> 348.03ms
  • middle-block 368.58ms -> 297.34ms
• core observation at 5000 blocks is still bad:
  • editor.children.length 940.18ms -> 784.01ms
  • Editor.nodes(...) 1275.18ms -> 796.95ms
  • Editor.positions(...) 1013.11ms -> 693.64ms

Chunking comparison

Measured separately:

• at 5000 and 10000 blocks, legacy chunking still wins hard on typing
• current v2 wins hard on select-all and paste

That means:

• the runtime story is not coherent enough yet
• chunking is still outperforming us exactly where the current docs say it should become unnecessary

Root-Cause Ranking

Tier 1: Core transaction model

Most likely remaining wall:

• full draft bootstrap cost per outer op
• expensive snapshot publication cost
• too much per-op structural work for simple text edits

If this stays bad, React cannot save it.

Tier 2: Core read/observation paths

Still ugly:

• editor.children
• Editor.nodes(...)
• Editor.positions(...)
• runtime-id / index access patterns

This matters because the React-perfect runtime depends on cheap committed reads and narrow subscriptions. If core reads are gross, selectors become expensive too.

Tier 3: React invalidation and render breadth

This is the layer the rewrite is supposed to win on:

• selector-first subscriptions
• local rerender boundaries
• semantic islands
• active editing corridor
• adaptive occlusion

But this layer only matters after Tier 1 and Tier 2 stop poisoning it.

Program Decision

The optimization program should run in this order:

1. stabilize and keep the current benchmark truth
2. attack core transaction cost
3. attack core observation/read cost
4. then attack React/render invalidation
5. only then build the huge-doc escalation layer

Anything else is backwards.

Execution Plan

Phase 0: Freeze Truth

Goal:

• stop the benchmark surface from drifting while we optimize

Required artifacts:

• api-drift-perf-scoreboard.md
• a kept huge-doc vs chunking benchmark lane:
  • pnpm bench:replacement:huge-document:chunking:compare:local
• kept core-only huge-doc typing lane
• kept core-only observation lane

Work:

• promote the current ad hoc huge-doc chunking comparison into a repo-owned script with stable outputs
• keep the core-only compare scripts in slate-v2/scripts/
• ensure every lane writes JSON to .tmp/

Acceptance:

• no more temp harness dependence for roadmap truth

Phase 1: Kill Per-Op Core Bootstrap Cost

Goal:

• make plain collapsed text inserts stop paying whole-world transaction setup

Why first:

• huge-doc typing is still catastrophic at core level
• this is the strongest sign that every edit is still bootstrapping too much

Likely files:

• with-transaction.ts
• apply.ts
• apply-operation.ts
• draft-helpers.ts
• replace-snapshot.ts

Required ideas to evaluate:

• a fast outer-op path for simple collapsed insert_text / remove_text
• partial draft reuse instead of full createDraftTree(...) on every outer op
• smaller publication work for unchanged subtrees

Guardrails:

• range refs stay correct
• selection stays correct
• snapshot immutability stays real
• no hidden mutation of committed snapshots

Acceptance:

• core huge-document typing lane shows a large step down
• no contract regressions

Phase 2: Kill Core Observation Cost

Goal:

• make observation reads cheap enough that selector-first runtime is plausible

Likely files:

• get-current-children.ts
• get-current-index.ts
• get-current-node.ts
• editor.ts

Work:

• cache transaction reads only where the cache is mutation-version-safe
• remove accidental full-tree cloning on read paths where possible
• make Editor.nodes(...) and Editor.positions(...) honestly lazy only after measuring a design that preserves current behavior

Guardrails:

• do not leak draft metadata into public reads
• do not regress snapshot-contract rows
• do not accept clever fast paths that win one microbench and corrupt helper semantics

Acceptance:

• core observation lane steps down materially
• package tests stay green

Phase 3: React Runtime Narrowing

Goal:

• make the runtime behave like the architecture contract says it should

Reference:

• architecture-contract.md

Work:

• keep useSyncExternalStore selector-first
• remove broad rerender defaults
• make leaf and element subscriptions local by default
• keep derived UI off broad editor subscriptions

Likely lanes:

• #5131-style selection subscription breadth
• #3656-style leaf rerender breadth
• #4141-style ancestor rerender breadth

Acceptance:

• React lanes improve without chunking becoming the main story

Phase 4: Huge-Doc Escalation Layer

Goal:

• beat legacy chunking where it still wins, without making chunking the foundational v2 story

Reference:

• chunking-review.md

Work:

• semantic islands, not numeric chunks
• active editing corridor
• adaptive occlusion outside the corridor
• optional <Activity> / hidden-tree posture where it helps
• optional planning layer later, but DOM truth stays authoritative in the active corridor

Acceptance:

• huge-doc typing at 5000 and 10000 is no longer badly behind legacy chunking

Verification Stack

For every phase:

1. targeted contract tests
2. full slate package tests
3. build
4. typecheck
5. core-only compare lanes
6. if React/runtime changed, user-facing replacement lanes

Do not claim a win without:

• before numbers
• after numbers
• proof that behavior stayed intact

Regression Policy

Hard no:

• draft metadata leaks into public reads
• weakened range-ref behavior
• weakened clipboard behavior
• weakened snapshot immutability
• broad live normalization reopening by accident

When an optimization regresses contract behavior:

• revert it immediately
• keep the idea in notes as rejected
• do not “fix later” unless the regression is already understood and bounded

Harsh Feedback

The rewrite says it wants a React-perfect runtime.

Right now the harsh truth is:

• the React story is better than legacy on some user-facing flows
• the core still does too much work per edit
• huge-doc typing is not remotely good enough
• chunking still embarrasses us on typing at scale

So stop pretending the next move is another cute hook or another normalization rule.

The next move is boring, deep engine work:

• transaction bootstrap
• snapshot publish
• transaction reads
• then runtime invalidation
• then semantic islands / active corridor

If the core keeps rebuilding too much state per keystroke, the React-perfect runtime is just branding.

Success Condition

This program is done when all of these are true:

1. core huge-doc typing no longer looks ridiculous versus legacy
2. core observation reads are no longer absurdly expensive
3. React rerender breadth lanes are locally scoped by default
4. huge-doc typing no longer needs legacy chunking as the main rescue story
5. the remaining perf debt is small enough to be honest follow-up work, not architectural failure