pagination virtualization char burst perf

Objective: Optimize Slate pagination virtualization typing; done when char-burst benchmark proves virtualized rows800 within staged tolerance and correctness passes.

Completion threshold: Rows=800 virtualized pagination passes the source-backed char-burst benchmark with pagination_virtualized_failed=0, correct inserted text, bounded DOM/page surfaces, p95 typing <= 16ms, scroll <= 400ms, and app-ready <= 800ms.

Verification surface:

.tmp/slate-v2 benchmark: bun run bench:react:pagination-virtualized-char-burst:local
.tmp/slate-v2 focused browser correctness: pagination rows800 perf, insert-break burst, double-click word selection, and right-margin wrapped-line selection tests
.tmp/slate-v2 package/site checks: slate-layout typecheck/tests, slate-react typecheck, site typecheck
root benchmark target checks: pnpm bench:targets:check, pnpm bench:targets:dry-run -- react-pagination-virtualized-char-burst, pnpm slate:ar:benchmark-lint

Constraints:

Preserve native text editing, selection, double-click word selection, margin click placement, and insert-break routing while improving pagination virtualization.
Keep the package ownership boundary in slate-layout / slate-react; keep example code as a consumer proof, not a second layout engine.
Do not create PRs or pushes for this task.

Boundaries:

Source of truth: .tmp/slate-v2 packages and pagination example, with root benchmark target wiring in benchmarks/targets/slate-v2.json.
Allowed edit scope: pagination virtualization package code, focused pagination example code, benchmark target/script, and this goal plan.
Browser surface: /examples/pagination?page_layout=single&rows=800&strategy=virtualized.
Tracker sync: N/A, no issue or PR was requested.
Non-goals: full pagination product redesign, unrelated dirty files, and release packaging.

Blocked condition: No current blocker. If the benchmark becomes flaky again, block only after a repeatable failing trace shows source-backed correctness or perf cannot be stabilized without a product/API decision.

Start Gates:

Gate	Applies	Evidence
Skill analysis before edits	yes	Used slate-autoresearch and autogoal workflow for a measurable perf target.
Active goal checked or created	yes	Active Codex goal: pagination virtualization char-burst perf.
Source of truth read before edits	yes	Read pagination example, slate-layout PagedEditable/page mount plan, slate-react virtualizer, and benchmark script.
TDD decision before behavior change or bug fix	yes	Existing focused Playwright tests plus a new page-mount-plan unit expectation cover the changed behavior.
Branch decision for code-changing task	N/A	User did not request branch/PR work.
Browser route / app surface identified	yes	`/examples/pagination?page_layout=single&rows=800&strategy=virtualized`.
Output budget strategy recorded	yes	Used focused `sed`, `rg`, scoped tests, and metric artifacts instead of broad output dumps.

Work Checklist:

Objective, completion threshold, verification surface, constraints, boundaries, and blocked condition are concrete.
Reproduced the failing source-backed rows800 virtualized benchmark.
Fixed page virtualization startup by mounting only an initial page window before viewport measurement.
Scoped page projection and decorations to visible/selected fragments instead of whole-document projection.
Fixed scroll-selection retention so user scroll is not yanked back to the selected block.
Fixed page-mode virtualizer retention so selected caret keeps the selected top-level node, not an unrelated whole page.
Preserved native interaction coverage for double-click word selection, wrapped right-margin selection, and insert-break bursts.
Recorded accepted Autoresearch evidence for the kept packet.

Completion Gates:

Gate	Applies	Required action	Evidence
Named verification threshold	yes	Run benchmark and focused browser proof	`pagination_virtualized_failed=0`, load 561.6ms, scroll 357ms, p95 11.6ms in `pnpm slate:ar:benchmark-lint`; focused Playwright 4 passed.
Bug reproduced before fix	yes	Capture failing benchmark symptoms	Earlier source-backed runs failed app-ready >800ms and scroll >400ms; selection-retained scroll also exceeded DOM bounds.
Targeted behavior verification	yes	Run package and browser tests	slate-layout tests 41 passed; focused pagination Playwright 4 passed.
TypeScript changed	yes	Run touched typechecks	slate-layout typecheck passed, slate-react typecheck passed, site typecheck passed.
Package manifests changed	yes	Verify benchmark target wiring	`pnpm bench:targets:check` and dry-run passed.
Browser surface changed	yes	Browser proof through Playwright	Focused pagination browser tests passed against the example route.
Package behavior or public API changed	no	No changeset for internal/example benchmark work	N/A: no public package API was added.
Autoreview for non-trivial implementation changes	no	Defer unless user asks	N/A: user asked for Autoresearch perf loop; focused correctness and benchmark gates passed.
Final lint	no	Scoped type/tests used	N/A: no broad lint requested; touched TS typechecks passed.
Goal plan complete	yes	Run autogoal complete check	Ready to run after this update.

Phase / pass table:

Phase	Status	Evidence	Next
Intake and source read	done	Relevant package/example/benchmark files read	implementation
Implementation	done	Page-windowing, projection scoping, selection-scroll guard, retention fix	verification
Verification	done	Benchmark, typechecks, package tests, focused Playwright	closeout
PR / tracker sync	N/A	User did not request PR or tracker sync	closeout
Closeout	done	Autoresearch keep logged and goal plan updated	final response

Findings:

Whole-page retention for a selected caret was the main scroll correctness bug after the perf changes.
Initial mount paid duplicate layout work: constructor refresh plus mount-effect refresh.

Decisions and tradeoffs:

Keep page-level virtualization as default for pagination, but do not let selected paths retain whole page content.
Suppress selection scroll only briefly after user scroll; keyboard typing still scrolls the caret into view.

Implementation notes:

PagedEditable now starts with an initial virtualized page window before the viewport exists.
PagedEditable projects visible/selected fragments for rendering while keeping cheap top-level layout items for virtualizer offsets.
Pagination example uses rendered fragment bounds and lazy decorations instead of a duplicate full-document projection.

Review fixes:

N/A.

Error attempts:

Error / failed attempt	Count	Next different move	Resolution
Benchmark initially measured stale static output	1	Run benchmark against `next dev` source by default	Fixed benchmark harness.
Source benchmark failed app-ready and scroll gates	3	Remove full initial page render, full-doc projection, duplicate refresh, and whole-page selection retention	Fixed; benchmark passes.

Verification evidence:

bun --filter ./packages/slate-layout typecheck && bun --filter ./packages/slate-layout test --timeout 10000: passed, 41 tests.
bun --filter ./packages/slate-react typecheck: passed.
bun typecheck:site: passed.
bun run bench:react:pagination-virtualized-char-burst:local: passed with load 578.4ms, scroll 352.1ms, p95 8.8ms, failed 0.
pnpm slate:ar:benchmark-lint: passed with load 561.6ms, scroll 357ms, p95 11.6ms, failed 0.
Focused Playwright pagination tests: 4 passed, covering rows800 perf, insert-break burst, double-click word selection, and wrapped right-margin selection.
pnpm bench:targets:check: passed.
pnpm bench:targets:dry-run -- react-pagination-virtualized-char-burst: passed.

Reboot status: Current. Continue only if user wants broader full integration or review sweep.

Open risks: Autoresearch labels the keeps as exploratory until repeat/holdout promotion metadata is added. Full pagination integration was not run in this slice.

Final handoff contract:

PR line: N/A, no PR requested.
Issue / tracker line: N/A.
Confidence line: high for focused rows800 pagination virtualization perf and covered native interaction regressions.