@a3s-lab/office 0.12.0 → 0.13.0

package/dist/work-spreadsheet-package-scan.worker.js.LICENSE.txt

@@ -0,0 +1,11 @@
+/*!
+
+JSZip v3.10.1 - A JavaScript class for generating and reading zip files
+<http://stuartk.com/jszip>
+
+(c) 2009-2016 Stuart Knightley <stuart [at] stuartk.com>
+Dual licenced under the MIT license or GPLv3. See https://raw.github.com/Stuk/jszip/main/LICENSE.markdown.
+
+JSZip uses the library pako released under the MIT license :
+https://github.com/nodeca/pako/blob/main/LICENSE
+*/

package/docs/latest/en/browser-editor-architecture.md

@@ -9,7 +9,7 @@ grid, slide canvas, or PDF page.
 
 | Product | Interaction surface | TipTap responsibility | Kernel responsibility |
 | --- | --- | --- | --- |
-| Document | TipTap and ProseMirror | Complete logical document, selection, commands, history, comments, and collaboration boundary | Page layout, style and font resolution, OOXML semantics, and serialization |
+| Document | TipTap and ProseMirror | Canonical structured model plus an equal-position lazy live tree for eligible giant DOCX files; selection, commands, history, comments, and collaboration boundary | Page layout, style and font resolution, OOXML semantics, and serialization |
 | Spreadsheet | Virtualized grid and canvas | Rich text inside a cell or floating text object | Formula calculation, workbook semantics, print layout, OOXML, and serialization |
 | Presentation | Scene graph and slide canvas | Text inside an individual text box | Masters, layouts, themes, object geometry, OOXML, and serialization |
 | PDF | PDFium page surface | None | PDF parsing, rendering, annotation serialization, and document save |
@@ -17,6 +17,15 @@ grid, slide canvas, or PDF page.
 This separation prevents an editor framework from becoming a false abstraction
 over products with different selection, layout, and performance requirements.
 
+TipTap and ProseMirror are not competing choices in this stack. TipTap owns the
+extension, command, schema, React integration, and public customization layer;
+ProseMirror is the model and view engine underneath it. Ordinary document
+features stay at the TipTap layer. Position-sensitive infrastructure such as
+pagination decorations, transaction mapping, collaboration bindings, and
+large-document NodeViews uses ProseMirror primitives through TipTap extensions
+because those primitives expose the exact model positions and view lifecycle
+that the higher-level API intentionally abstracts.
+
 ## Shared platform layers
 
 The products share infrastructure without sharing a false universal document
@@ -280,7 +289,7 @@ focus restoration.
 | --- | --- | --- | --- |
 | Document | One TipTap/ProseMirror body tree retained across editing and read-only preview, controlled TipTap header/footer surfaces with direct paper-margin editing and a contextual ribbon, one typography and page-chrome baseline across editing, preview, and PDF surfaces, a persistent typed heading-navigation pane with active-selection tracking, filtering and collapsible keyboard traversal, host-owned selected-text menus with full document context and range-mapped async commands, responsive and keyboard-operated paragraph-style and list galleries, typed bullet and numbering commands with restart/continue/start controls, typed physical-page and section-page descriptors, repeated first/default/even page chrome, a versioned structured model with an HTML compatibility representation, prefix-reused visual-line measurement and pages, page decorations, page-aware horizontal and vertical rulers for page margins, paragraph indents and typed tab stops, structured list-item pagination, explicit paragraph and list-item direction, compact spacing and pagination controls, typed inline/square/tight/through/top-and-bottom/no-wrap image layout and floating-image drawing layers, imported style-inherited paragraph properties, structured inline tabs, and theme-aware run font/size/color/background import | Worker plus resumable Rust/WASM flow pagination and Rustybuzz shaping across CSS-matched registered text runs; the same live result remains mounted in read-only preview and is consumed by bounded browser PDF capture, including eligible list paragraphs, Unicode bidi level segmentation, ordered per-grapheme font fallback, packaged Latin/CJK/Arabic/Hebrew faces, and structured left-to-right tabs, with explicit DOM and JavaScript fallbacks for text affected by supported floats | Language-complete font substitution, complete Word style and numbering coverage, locale-complete and bidirectional tabs, arbitrary floating-object geometry and non-image drawing layers, complex table flow, searchable/vector PDF output, and loss-preserving OOXML package state |
 | Markdown | TipTap visual editing with a resizable source-and-preview split view by default, source-aware ribbon formatting and shortcuts, bounded source-native history with coalesced typing and selection restoration, host-defined selection menus across both editing surfaces, GFM tables, strikethrough, autolinks and nested task lists, controlled source state, coalesced preview rebuilds, proportional pane scrolling, keyboard-adjustable pane sizing, a flat measure-bounded reading surface, optional visual or source-only views, and a full-workspace Source/Preview switch on phones | No kernel required for normal editing | CommonMark differential fixtures, multi-megabyte profiling, and an off-main-thread parser boundary when measurements justify it |
-| Spreadsheet | Fortune Sheet grid integrated with the shared Office shell, a WPS-oriented command catalog and tab order, quick-access Undo and Redo, a priority-aware adaptive and collapsible ribbon with temporary tab expansion, Home and Styles Conditional Formatting, Home and Cells row/column insertion and deletion through the shared typed structure commands, Home and Alignment Merge and Center plus Merge Cells, Merge Across, Unmerge Cells, and Unmerge and Fill through one native-model-backed controlled batch per intent and the WPS `Ctrl+M` shortcut, Home and Editing Clear All, Formats, Contents, Comments, and Hyperlinks with Delete/Backspace mapped to content-only clearing and merge geometry retained, Data sorting and AutoFilter with bounded current-region discovery, controlled filter state, WPS toggle and header-menu shortcuts, and accessible vendor menu semantics, View and Window Freeze Panes with WPS current-cell, top-row, first-column, and unfreeze patterns backed by controlled state and native XLSX frozen panes, Formulas and F9 workbook recalculation, an executable Home clipboard group with permission-resilient Paste, Cut, and Copy plus a WPS-style one-shot and locked Format Painter for bounded cross-sheet native-style patterns, shared grid/formula/footer visual tokens and one spreadsheet accent across selection, sheet tabs, menus, and zoom, an A3S-owned accessible single-row workbook footer for creation, activation, rename, duplicate, color, hide, reorder, delete, selection status, and zoom, a resize-observed worksheet viewport that keeps the active tab and its touch action visible across desktop, compact, and phone-width changes, one cell/worksheet context-menu surface whose displayed accelerators remain executable while the menu owns focus, Arrow/Home/End tab navigation plus Shift+F10 and native context-menu access, typed editing and calculation command ports, A3S-owned deterministic Arrow, Enter, Tab, Home, PageUp/PageDown, extended-selection, row, column, and all-cells keyboard commands, selection-preserving focus restoration across controlled workbook remounts and F2/Escape editing transitions, Cmd/Ctrl formatting, undo/redo, clear, a permission-resilient editor-local clipboard fallback, Shift+F11 sheet creation, and Ctrl/Cmd+PageUp/PageDown sheet-navigation shortcuts, direct font-family, horizontal/vertical-alignment, text-wrap, general/number/percent, and decimal-place controls backed by native cell-style keys, live count/sum/average selection summaries, operation-driven sparse-workbook projection, guarded controlled-value remounts that reject stale engine callbacks, and no-history result patches with cell-scoped Fortune fallback | Versioned, cancellable Worker/Rust-WASM calculation sessions using the shared bounded Rust formula parser, retained formula ASTs, incremental forward/reverse dependency graphs, dirty-subgraph recalculation, cross-sheet references, and a dynamically loaded JavaScript fallback | A3S-owned virtual grid, moving replacement projection off the main thread, broader Excel formula semantics, A3S-owned custom number-format evaluation, and print layout |
+| Spreadsheet | Fortune Sheet grid integrated with the shared Office shell, a WPS-oriented command catalog and tab order, quick-access Undo and Redo, a priority-aware adaptive and collapsible ribbon with temporary tab expansion, Home and Styles Conditional Formatting, Home and Cells row/column insertion and deletion through the shared typed structure commands, Home and Alignment Merge and Center plus Merge Cells, Merge Across, Unmerge Cells, and Unmerge and Fill through one native-model-backed controlled batch per intent and the WPS `Ctrl+M` shortcut, Home and Font native strikethrough through the same typed formatting command and WPS `Cmd/Ctrl+5`, Home and Editing Clear All, Formats, Contents, Comments, and Hyperlinks with Delete/Backspace mapped to content-only clearing and merge geometry retained, Data sorting and AutoFilter with bounded current-region discovery, controlled filter state, WPS toggle and header-menu shortcuts, and accessible vendor menu semantics, View and Window Freeze Panes with WPS current-cell, top-row, first-column, and unfreeze patterns backed by controlled state and native XLSX frozen panes, Formulas and F9 workbook recalculation, an executable Home clipboard group with permission-resilient Paste, Cut, and Copy plus a WPS-style one-shot and locked Format Painter for bounded cross-sheet native-style patterns, shared grid/formula/footer visual tokens and one spreadsheet accent across selection, sheet tabs, menus, and zoom, an A3S-owned accessible single-row workbook footer for creation, activation, rename, duplicate, color, hide, reorder, delete, selection status, and zoom, a resize-observed worksheet viewport that keeps the active tab and its touch action visible across desktop, compact, and phone-width changes, one cell/worksheet context-menu surface whose displayed accelerators remain executable while the menu owns focus, Arrow/Home/End tab navigation plus Shift+F10 and native context-menu access, typed editing and calculation command ports, A3S-owned deterministic Arrow, Enter, Tab, Home, PageUp/PageDown, extended-selection, row, column, and all-cells keyboard commands, selection-preserving focus restoration across controlled workbook remounts and F2/Escape editing transitions, Cmd/Ctrl formatting, undo/redo, clear, a permission-resilient editor-local clipboard fallback, Shift+F11 sheet creation, and Ctrl/Cmd+PageUp/PageDown sheet-navigation shortcuts, direct font-family, horizontal/vertical-alignment, text-wrap, general/number/percent, and decimal-place controls backed by native cell-style keys, live count/sum/average selection summaries, operation-driven sparse-workbook projection, guarded controlled-value remounts that reject stale engine callbacks, and no-history result patches with cell-scoped Fortune fallback | Versioned, cancellable Worker/Rust-WASM calculation sessions using the shared bounded Rust formula parser, retained formula ASTs, incremental forward/reverse dependency graphs, dirty-subgraph recalculation, cross-sheet references, and a dynamically loaded JavaScript fallback | A3S-owned virtual grid, moving replacement projection off the main thread, broader Excel formula semantics, A3S-owned custom number-format evaluation, and print layout |
 | Presentation | Scene canvas with ordered typed multi-selection, persistent nested browser groups, native PPTX group-node export, exact keyboard-accessible table-dimension insertion, native slide/object context actions with optional AI actions, a separate object/content editing state, one on-demand TipTap instance, collective move/scale/nudge/clipboard/delete/layer commands, selection-bound alignment and distribution, typed group/ungroup commands, one typed dispatcher for ribbon commands, editor-scoped shortcuts with post-command selection focus, direct beginning/current-slide playback with fullscreen fallback, frame-coalesced transactional move/resize previews that commit once on pointer release, two-level thumbnail node and scene windowing, and a phone stage that top-aligns the primary canvas while retaining one readable slide indicator beside view and zoom controls | Revisioned, cancellable Worker/Rust-WASM slide-relative alignment and object-set snapping with typed visual guides and a JavaScript fallback | Arbitrary rotated or reflected PPTX group transforms, connectors, theme resolution, text fitting, kernel-owned thumbnail layout, and slide serialization |
 | PDF | PDFium-backed page rendering with an A3S-owned responsive toolbar, a scrollable page-thumbnail rail with active-page synchronization and bounded bitmap/DOM windowing, a focus-contained phone page drawer, and typed capability controllers for navigation, zoom, search, basic annotations, annotation color, opacity, compatible stroke-width defaults and selection updates, history, and save; page and search drafts cancel without accidental commands, shortcuts remain scoped to the PDF root, and compact widths retain page status while exposing search-result traversal, navigation, zoom, and history through the keyboard-operated overflow menu | PDFium WebAssembly | Forms, redaction review, page organization, and reopen fixtures |
 
@@ -321,6 +330,96 @@ batches when their geometry is uniform or from exact per-page viewports when
 it differs. Explicit descriptor pages remain a fallback for surfaces without a
 registered live document.
 
+Eligible structurally plain large DOCX files take a narrow import path in a
+dedicated Worker. The main thread inflates `word/document.xml` once and
+transfers its exact `ArrayBuffer`; the Worker decodes it, performs the strict
+WordprocessingML tag and envelope checks, and streams paragraphs or table rows
+in batches of 2,048. Table messages carry one text vector plus transferable
+`Uint32Array` bands for row-to-cell and cell-to-paragraph counts, so neither
+side clones a 100,000-row nested object graph. The main thread constructs the
+complete, schema-validated `WorkDocumentModel` and canonical HTML from those
+bounded messages, but it does not create the complete live ProseMirror tree.
+The TipTap source initially materializes
+only the first two leaf chunks. Every later leaf contains an internal
+`documentLazyBlock` whose position tape has exactly the same `nodeSize`, text
+offsets, and block boundaries as the canonical payload. Ordinary text uses
+128-block leaves above 2,048 blocks, while tables use 16-row leaves above 512
+rows. Runs longer than 32 table leaves are nested under geometry containers, so
+a 100,000-row table exposes 6,250 logical leaves through only 196 top-level
+containers without changing any canonical position.
+
+The Worker boundary is deliberately fail-closed. An explicit `ineligible`
+result continues to the complete Mammoth and OOXML-marker importer. Worker
+creation, execution, message-validation, or timeout failure terminates that
+exact Worker and retries the same narrow parser synchronously; cancellation
+terminates it and propagates `AbortError`. The production import Worker is
+23.9 kB uncompressed, about 97.4 percent smaller than the earlier
+dependency-heavy prototype, and contains neither JSZip nor the rich
+section-normalization pipeline. Envelope counting uses one scan, and the row
+parser emits the same columnar shape that the Worker protocol consumes.
+
+Hosts may reserve an artifact ID before starting import. The Playground mounts
+a blank editor shell under that ID while the Worker parses, then applies the
+final controlled model to the already mounted editor. Parsing and editor
+initialization are therefore parallel critical paths rather than serial work;
+source-backed export is rekeyed atomically to the reserved identity. A failed
+or cancelled import removes only its own placeholder and restores the previous
+workspace.
+
+The chunk NodeView omits descendant DOM for an off-screen leaf. A visible lazy
+leaf receives a read-only semantic preview from a structure-keyed DOM pool;
+paragraph and table nodes are updated in place and returned to the pool as the
+window moves. Pointer selection, search, model-boundary navigation, or a real
+edit hydrates only the selected chunk from its canonical payload before the
+selection is applied. Hydration is history-free, does not publish a controlled
+edit, and transfers cached document statistics because its logical content is
+unchanged. The placeholder retains estimated content height plus indexed
+pagination spacer height, and nested leaf heights aggregate into the collapsed
+container. Expanding, collapsing, and pooling therefore do not change the
+scroll range. A cumulative geometry index and binary search map the viewport
+to its exact leaves. `Ctrl+Home`, `Ctrl+End`, and the macOS equivalents operate
+on model positions first and never depend on mounted DOM.
+
+Pagination reads the canonical JSON payloads of lazy chunks without hydrating
+them into ProseMirror. A text edit reuses the measured block and page prefix
+before the first changed position, locates the current page with a binary
+search, and shapes only real paragraph or heading nodes. This prevents a local
+tail edit from filtering, comparing, or recreating all 100,000 prior blocks.
+Selection, IME, undo, comments, revisions, and exported model positions remain
+model-authoritative throughout this process.
+
+Controlled ownership remains intact. Text-only transactions in an eligible
+lazy document are coalesced to one animation-frame publication. The editor
+finds the changed persistent chunk, updates document statistics by subtree
+delta, materializes the complete JSON model from the payload registry, serializes
+only that chunk, and patches its indexed range in the canonical HTML. The host
+still receives complete `html` and `model` values. A composable polynomial
+fingerprint hashes the changed chunk and combines its cached prefix, chunk, and
+suffix segments, so consecutive publications do not rescan the complete HTML
+string. Process-local schema trust admits that precomputed fingerprint only for
+the exact live model snapshot; cloned or persisted models still rescan and
+verify the complete HTML, including fingerprints written by earlier releases.
+Formatting, structure, or unsupported edits deliberately fall back to complete
+schema serialization. The lazy projection is process-local and
+parser-authenticated; rich compatibility imports, cloned untrusted models, and
+collaboration bindings retain the complete TipTap path.
+
+Physical page chrome has a separate React window above 24 pages and mounts only
+the nearby sheets. A WeakMap-backed page-surface registry retains all page
+frames for PDF geometry without recreating thousands of page-sheet elements.
+Pagination widgets are indexed by leaf chunk and appear only with a mounted
+leaf; logical spacer height remains in the chunk geometry when the widget DOM
+is absent. Single-column sections deliberately use ordinary block flow rather
+than CSS Multi-column, while true multi-column sections retain the column
+algorithm. Word content-autofit tables ignore TipTap's inline grid widths at
+render time so the browser sizes them from content; fixed and percentage table
+geometry remains unchanged.
+
+This path is not a claim that every rich 100,000-block DOCX is already bounded.
+Unsupported tags, relationships, tracked structures, drawings, or other rich
+features route to the complete compatibility importer. Extending the same
+model-preserving window to those paths remains a performance gate.
+
 The Word navigation pane uses that same capture boundary for real physical-page
 thumbnails instead of reconstructing an approximate text card. The current and
 adjacent pages are admitted first; the remaining captures use an
@@ -852,6 +951,48 @@ use Fortune coercion and eager-branch evaluation for formulas beyond the
 shared parity fixtures; this is an explicit first-slice compatibility limit,
 not a cross-engine equivalence claim.
 
+XLSX file import starts two bounded, one-shot Workers before the calculation
+session. The general Worker runs SheetJS off the main thread and streams dense
+worksheets in 512-row chunks or sparse worksheets in 8,192-cell chunks. In
+parallel, the package Worker decompresses each worksheet XML part once and
+builds the feature and style gates used by the rest of the importer.
+
+The package Worker also owns a deliberately narrow, fail-closed path for plain
+OOXML workbooks. It first authenticates an exact package envelope: workbook,
+content types, relationships, and worksheet ownership must form a closed set
+with no styles, shared strings, external targets, or unknown parts. It then
+parses only ordered primitive numeric, Boolean, error, and simple inline-string
+cells. A worksheet-start message authenticates its declared range before any
+cell is accepted. Each later transferable columnar chunk covers at most 256
+rows and 4,096 cells and carries packed coordinates, value kinds, numeric
+values, and the remaining text vector instead of a million nested cell
+objects. The hot cursor validates row and cell coordinates in place with
+bounded character-code arithmetic; it creates neither an address substring nor
+a regular-expression match per cell. The main thread treats those chunks as
+provisional until the complete worksheet XML has been consumed.
+A formula, style index, rich/shared string, unknown element or attribute,
+malformed entity, invalid address, relationship, or trailing payload rejects
+the candidate instead of guessing. Package authentication cancels the
+speculative SheetJS Worker early; a later cell-level rejection discards every
+provisional row and restarts SheetJS with the original private bytes. Worker
+creation, protocol, timeout, and cancellation failures retain the synchronous
+parser as the final compatibility fallback. This removes the second complete
+worksheet decompression and SheetJS object-materialization pass for eligible
+million-cell files without weakening the full XLSX path.
+
+Primitive imported cells and their sparse rows are frozen once, and an
+identity-keyed profile authenticates the resulting matrix as directly
+consumable by Fortune. A host may reserve worksheet IDs while the Playground
+mounts a blank workbook shell. If the final workbook is editable, preserves
+the same sheet count and identities, and contains no stateful structure such
+as charts, protection, merged geometry, filters, validation, images, or custom
+view state, the editor replaces that certified matrix through the existing
+Fortune instance. Fortune adopts the frozen matrix and rebuilds only its
+formula index; it does not clone every populated cell or remount the workbook.
+Any failed predicate or update exception returns to the complete controlled
+remount path. This narrow eligibility gate preserves correctness for rich and
+multi-sheet imports while removing the simple million-cell first-mount task.
+
 The browser pagination implementation is split by capability instead of
 accumulating in one editor module. `work-document-pagination.ts` owns only the
 TipTap extension and public facade; dedicated modules own block measurement,
@@ -1145,6 +1286,26 @@ change. Environments without `IntersectionObserver` render complete scenes for
 the bounded mounted window. Kernel-owned thumbnail layout remains a separate
 performance gate.
 
+Presentation design metadata is memoized per controlled content identity.
+Thumbnail scenes consume that prepared master/layout catalog directly, and the
+editor does not copy the complete slide array merely to resolve one visible
+scene. The deterministic performance fixture contains 1,000 slides and 9,000
+scene elements. In five fresh headless Chrome 149 processes on the recorded
+10-logical-core host, normal mode mounted 18 thumbnail buttons, rendered 13
+full thumbnail scenes, retained 1,006 DOM nodes and a median 10.9 MiB
+JavaScript heap, reached the final slide in 6.1–13.4 ms, and recorded zero Long
+Tasks during import, final-slide navigation, and ten object nudges. The median
+keydown-to-object-commit time was 4.5 ms.
+
+`content-visibility: auto` remains a benchmark mode rather than a production
+rail rule. The interleaved five-process comparison reduced median editor-ready
+time only from 1,637.6 ms to 1,600.3 ms, while increasing median object-commit
+time from 4.5 ms to 5.9 ms and median nudge script time from 8.0 ms to 11.5 ms,
+with no retained-heap improvement. The rail therefore uses actual DOM and
+scene virtualization; CSS skipped rendering cannot replace the JavaScript and
+data-processing bounds and can add invalidation cost to the already bounded
+window.
+
 Exit criteria: object drag and resize stay interactive on complex slides;
 partial rich-text formatting survives PPTX round trips; masters, layouts,
 themes, tables, charts, links, and notes have compatibility fixtures for
@@ -1169,8 +1330,27 @@ their direct toolbar controls are hidden. The product-owned thumbnail rail uses
 PDFium thumbnail capabilities without enabling the embedded sidebar, keeps page
 numbers and active-page state synchronized with the main scroll controller,
 automatically reveals the current page, and mounts only a bounded page window
-for long files. At phone widths the same navigation becomes a focus-contained,
-dismissible drawer rather than reducing the document canvas.
+for long files. Files above 48 pages retain virtual spacers and mount only the
+viewport plus five items of overscan on either side, with a hard ceiling of 32
+thumbnail buttons. Leaving the window aborts the exact pending PDFium thumbnail
+task and revokes its object URL. Long-distance Home and End navigation uses an
+instant rail jump so intermediate windows do not start disposable bitmap work
+or unmount the keyboard destination after it receives focus. At phone widths
+the same navigation becomes a focus-contained, dismissible drawer rather than
+reducing the document canvas.
+
+`preloadOfficeEditor('pdf', { preloadRuntimeAssets: true })` can move the PDF
+editor chunk and PDFium response body ahead of a high-confidence navigation.
+Runtime-asset loading is opt-in because the unpacked binary is about 4.4 MiB,
+and the helper does not initialize the PDFium Worker. Runtime fetching is best
+effort: a failure leaves editor opening unaffected and clears the request cache
+so a later intent can retry. The reference benchmark therefore treats
+module/asset availability and Worker-side initialization as separate
+boundaries. Host CSS `content-visibility` is not applied to the internal PDF
+pages: EmbedPDF already mounts only six to seven pages behind its Shadow DOM
+boundary, while CSS containment cannot cancel PDFium tasks or release bitmap
+memory.
+
 Existing PDF password handling remains in the PDF document lifecycle rather
 than the removed toolbar. Form-authoring controls, redaction review, page
 organization, and compatibility fixtures remain part of this stage.
@@ -1212,9 +1392,145 @@ repeatable fixtures. The targets below are release gates, not current claims:
 - Every editor has fixture-based memory, interaction-latency, load-time, and
   output-size budgets in CI before it can be called production-ready.
 
+### Current 100,000-unit evidence
+
+The following cold local Playground measurements were captured on 2026-08-19
+on an Apple M2 Pro with 16 GB of memory, Playwright 1.61.1, headless Chrome 149,
+and a 1,440 × 1,000 viewport. Document values are the median and range of three
+fresh browser processes; Spreadsheet values use five. The benchmark drains
+diagnostic memory-counter work before starting the scroll profile so
+DevTools-triggered V8 cleanup is not charged to the product. These measurements
+are evidence of current limits, not universal release targets. The spreadsheet
+readiness milestone is observed inside the page on animation frames; automation
+runner polling backoff is therefore excluded from the load time.
+
+| Fixture | Editor visible | Pagination ready | Longest task before visible |
+| --- | ---: | ---: | ---: |
+| DOCX, 100,000 text paragraphs, 4.50 million characters | 0.397 s (0.396–0.480) | 1.460 s (1.381–1.525), 3,125 pages | 131 ms (128–140) |
+| DOCX table, 100,000 rows × 3 columns | 0.813 s (0.810–0.814) | 1.747 s (1.739–1.749), 2,381 pages | 96 ms (88–144) |
+| XLSX table, 100,000 rows × 10 columns, 1 million populated cells | 0.480 s (0.474–0.494) | n/a | No task ≥50 ms in any of five runs |
+
+The median text phases were 121.6 ms for the complete import, 76.1 ms inside
+the Worker, and 56.6 ms for Worker parsing. The table phases were 462.6 ms,
+377.6 ms, and 311.6 ms respectively; the already mounted editor applied that
+100,000-row controlled model in 32.3 ms. Editor-shell initialization overlaps
+these import phases instead of extending the critical path. Against the prior
+Worker/windowing baseline, median editor visibility improved from 0.815 to
+0.397 seconds for text and from 1.316 to 0.813 seconds for the table;
+pagination readiness improved from 1.749 to 1.460 seconds and from 2.366 to
+1.747 seconds respectively.
+
+For the XLSX fixture, the complete import measured 375.8 ms (372.3–386.4).
+The authenticated package Worker fast path owned 345.5 ms (343.6–355.8), while
+main-thread canonical conversion took 13.8 ms (12.7–15.4). The hot worksheet
+cursor now authenticates row and cell coordinates directly in the 35.7 MB XML
+string instead of allocating and matching one million address substrings. Its
+isolated median fell from 178.3 to 124.3 ms (30.3%). Against the immediately
+preceding corrected end-to-end baseline, editor visibility improved from 536.7
+to 480.2 ms (10.5%), complete import from 432.0 to 375.8 ms (13.0%), and the
+Worker fast path from 403.5 to 345.5 ms (14.4%). No import Long Task, browser
+error, page error, or console error occurred in the retained five-run result.
+
+These values supersede the earlier 830.5 ms visibility figure. Page-local
+observation showed that Fortune had become visible before Playwright's next
+progressively backed-off locator poll; the old figure therefore included test
+runner latency rather than product work.
+
+| Fixture | Continuous scroll | p95 frame interval | Scroll long tasks | Retained browser state | Bounded render state |
+| --- | ---: | ---: | ---: | ---: | --- |
+| 100,000 text paragraphs | 120.0 FPS in every run | 10.7 ms (10.2–10.8) | 0 in every run | 70.4 MiB; 2,867 median DOM nodes | 1 of 782 content chunks at readiness, 2 after end navigation, and 4 of 3,125 page sheets mounted |
+| 100,000 table rows | 120.6 FPS (120.5–120.9) | 10.1 ms (9.9–10.4) | 0 in every run | 145.8 MiB; 3,106 DOM nodes | 4 of 6,250 leaves at readiness, 1 after end navigation; 196 outer containers; 4 of 2,381 page sheets mounted |
+| 100,000 × 10 spreadsheet cells | 120.1 FPS (120.0–120.2) | 12.9 ms (12.9–13.0) | 0 in every run | 38.41 MiB; 903 DOM nodes | Fortune Canvas paints only the visible row and column range; Ctrl+End reached `J100000` in 77.3 ms (76.0–79.7) |
+
+The controlled-edit benchmark sends one browser text-insertion event at the
+exact final paragraph or cell, waits for the complete host publication, then
+does it again against the new controlled revision. A page-local `beforeinput`
+listener starts `Wall latency`, and a `MutationObserver` stops it when the
+publication counter changes. This excludes Playwright input transport and
+polling backoff while retaining the input transaction, React and pagination
+scheduling, and publication. `Publish CPU` measures only construction of the
+complete `html` and `model` values. Document values below are the median and
+range of five fresh browser processes.
+
+| Fixture | First edit wall / publish CPU | Second edit wall / publish CPU | Publication path |
+| --- | ---: | ---: | --- |
+| 100,000 text paragraphs | 70.3 ms (69.0–115.4) / 21.8 ms (20.0–54.7) | 53.9 ms (22.5–56.4) / 10.0 ms (9.4–16.8) | `lazy-chunk` in all ten publications |
+| 100,000 table rows × 3 columns | 102.3 ms (93.8–121.0) / 57.8 ms (53.5–70.1) | 85.1 ms (78.2–93.3) / 41.2 ms (36.9–45.7) | `lazy-chunk` in all ten publications |
+| 100,000 × 10 spreadsheet cells | 52.5 ms (49.6–56.4) / 2.5 ms (2.5–2.6) | 45.9 ms (45.9–47.0) / 2.3 ms (2.2–2.3) | Incremental 3-operation projection of one changed cell; 0 Long Tasks |
+
+Every run reached the exact final paragraph or row, both consecutive edit
+markers survived in the final controlled value, and no browser, page, or
+console error was recorded. Pooled previews reduced detached scroll DOM by
+about 92 percent during profiling. Cached statistics removed the previous
+233–246 ms scroll task, while early non-paragraph rejection, binary page
+lookup, and measured-prefix reuse removed document-sized work from tail edits.
+
+The earlier document wall figures used a Node-side clock around Playwright
+input and polling. They therefore measured runner scheduling as product work
+and are superseded by the page-local figures above. Against the same page-local
+five-run baseline, cached segment fingerprints reduced the median second text
+edit from 101.4 to 53.9 ms and the second table edit from 170.8 to 85.1 ms;
+their publication CPU fell from 56.0 to 10.0 ms and from 123.2 to 41.2 ms.
+
+The remaining cold-open boundary is the 100,000-row Worker's approximately
+222 ms content scan plus main-thread construction and retention of the
+canonical model. The live ProseMirror tree and strict OOXML inspection are no
+longer document-sized main-thread constructors. A WASM parser is justified
+only if it consumes the transferred bytes directly and emits the same columnar
+protocol without adding a second complete copy or JSON serialization step.
+During editing, controlled ownership intentionally requires a complete HTML
+string; the indexed chunk patch and segment fingerprint avoid complete
+semantic serialization and repeat hashing, but allocating that host value
+remains a measurable lower bound. A future optional patch callback may remove
+that copy only if the complete `onChange` contract remains available.
+
+The million-cell XLSX path now retains about 38.41 MiB after collection and has
+neither import nor continuous-scroll Long Tasks in any run. The remaining
+cold-open boundary is the package Worker's approximately 344 ms XML scan and
+columnar cell construction; canonical main-thread conversion is already about
+13.8 ms. A WASM parser is justified only if it consumes the transferred bytes
+directly, emits the same bounded columnar protocol, and improves that Worker
+phase without introducing another complete copy. Fortune Canvas still owns
+painting, but imported matrix adoption is now outside the rich-workbook remount
+path for the strictly authenticated simple case.
+
+### Current 1,000-page PDF evidence
+
+The deterministic PDF fixture is 412,852 bytes with SHA-256
+`eb86216416a633ce86dd4b31c1c0d9e42792e7bb802946cc0bb6e2b395ae79cb`.
+It contains 1,000 A4 pages with stable page markers. The measurements below
+combine two counterbalanced three-run batches on the same 2026-08-19 reference
+machine and browser described above. Every value is the median and complete
+range of six fresh browser processes. `Runtime preloaded` waits for the editor
+module and PDFium response body before starting the file-open clock.
+
+| Mode | Editor shell | Viewer ready | First page bitmap | End to page 1,000 | Retained heap |
+| --- | ---: | ---: | ---: | ---: | ---: |
+| Import-time warm-up | 410.7 ms (399.4–433.1) | 7.430 s (4.403–13.742) | 7.626 s (4.611–13.944) | 21.7 ms (19.6–28.7) | 10.04 MiB (9.48–10.52) |
+| Runtime preloaded | 369.1 ms (322.8–428.6) | 9.464 s (5.286–12.390) | 9.655 s (5.478–12.580) | 23.5 ms (18.0–33.9) | 9.72 MiB (9.61–10.47) |
+
+Runtime preloading moved the shell median forward by 41.6 ms (10.1%) but did
+not improve viewer-ready or first-bitmap medians on the local server. The
+dominant and highly variable boundary is PDFium Worker initialization, not
+main-thread layout or local transfer. Runtime preloading remains an explicit
+network-latency tool rather than a claimed local-startup acceleration.
+
+All twelve runs mounted exactly seven main-view pages and 15 thumbnail buttons
+at readiness. The final window contained pages 986–1,000, page-local End
+navigation produced no Long Task, and no browser, page, or console error was
+recorded. Cold import produced no task at or above 50 ms. Three runtime-preload
+runs recorded one import task each, between 59 and 89 ms; this remains a gate
+for later Worker-initialization work rather than being hidden by the preload
+API. The deterministic A3S Test suite independently verifies the 15-item first
+and final windows, retained keyboard focus on page 1,000, the page counter,
+accessibility output, and empty console/page-error evidence.
+
 ## Performance and safety rules
 
-- Editing and selection stay on the main thread; parsing and layout do not.
+- Editing and selection stay on the main thread. XLSX parsing, eligible
+  large-DOCX inspection/parsing, and document layout use dedicated Workers;
+  canonical large-document model assembly remains a measured main-thread
+  boundary.
 - The public React entry loads editor engines as independent asynchronous
   chunks. `preloadOfficeEditor` may warm one engine from hover or keyboard
   focus without mounting it.
@@ -1227,7 +1543,8 @@ repeatable fixtures. The targets below are release gates, not current claims:
   CI against the scripts referenced by the generated `index.html`.
 - One active layout request exists per editor. A newer revision cancels the
   previous request.
-- Requests reject invalid dimensions and more than 10,000 layout blocks.
+- Requests reject invalid dimensions and more than 200,000 layout blocks.
+  Page-style tables remain bounded to 10,000 entries.
 - Text requests are bounded to 1,024 paragraphs and 1 MiB of UTF-8 text.
   They accept at most 16,384 contiguous runs, with no more than 4,096 in one
   paragraph and no more than eight ordered faces per run. Font registration is
@@ -1255,18 +1572,20 @@ The browser kernel is covered at four boundaries:
 1. Rust unit tests for deterministic pagination, Unicode line breaking,
    grapheme-safe emergency wrapping, whitespace modes, shared formula parsing,
    scalar dependency calculation, cycles, targets, and validation.
-2. JavaScript fallback tests for protocol parity, safe page-prefix reuse,
-   no-Worker operation, cancellation, sparse Spreadsheet calculation,
-   revisioned session fallback, and the shared Spreadsheet parity fixtures.
+2. JavaScript fallback tests for protocol parity, 100,000-block linear
+   pagination, safe page-prefix reuse, no-Worker operation, cancellation,
+   sparse Spreadsheet calculation, revisioned session fallback, dense and
+   sparse worksheet iteration, and the shared Spreadsheet parity fixtures.
 3. A raw generated-WASM ABI smoke test that registers both shipped fonts,
    proves the Latin face lacks CJK glyphs, resolves them through the ordered
    fallback face, verifies mixed-face line metrics, initializes a Spreadsheet
    session, and recalculates a dirty formula chain from a cell patch.
-4. Browser checks for real Worker/WASM/font loading, shaped-line parity with
-   browser line boxes at non-100% zoom, real per-grapheme fallback diagnostics,
-   explicit unresolved-glyph fallback, page-view reflow, web-view clearing,
-   page counts, one physical sheet per measured page, visible page gaps, nested
-   and RTL list flow, undo behavior, and slide-relative element alignment.
+4. Browser checks for real layout and XLSX-import Worker/WASM/font loading,
+   shaped-line parity with browser line boxes at non-100% zoom, real
+   per-grapheme fallback diagnostics, explicit unresolved-glyph fallback,
+   page-view reflow, web-view clearing, page counts, one physical sheet per
+   measured page, visible page gaps, nested and RTL list flow, undo behavior,
+   and slide-relative element alignment.
 
 Presentation group serialization tests inspect generated slide and layout
 OOXML, nested group order, identity child-coordinate transforms, unique
@@ -1278,7 +1597,10 @@ Presentation thumbnail tests use a controlled intersection observer to prove
 that long decks retain absolute keyboard reachability while bounding mounted
 buttons and full scenes, releasing scenes that leave the window, preserving
 focus after deletion, and reconnecting the observer when switching between
-normal and sorter views.
+normal and sorter views. A deterministic A3S Test suite imports the 1,000-slide
+fixture, proves the 18-button first and final windows, reaches slide 1,000 with
+`End`, performs two object nudges, and captures final focus, accessibility,
+console, and page-error evidence.
 
 Run the focused kernel checks with:
 
@@ -1293,4 +1615,18 @@ Build and verify the Playground performance boundary with:
 bun run playground:build
 bun run playground:bundle-check
 bun run playground:visual
+bun run test:e2e:large-documents:check
+bun run test:e2e:large-documents
+bun run performance:large-documents
+bun run performance:large-document-edits
+bun run test:e2e:spreadsheet-large:check
+bun run test:e2e:spreadsheet-large
+bun run performance:large-spreadsheets
+bun run performance:large-spreadsheet-edits
+bun run test:e2e:large-pdf:check
+bun run test:e2e:large-pdf
+bun run performance:pdf
+bun run test:e2e:large-presentation:check
+bun run test:e2e:large-presentation
+bun run performance:presentation
 ```

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@a3s-lab/office",
-  "version": "0.12.0",
+  "version": "0.13.0",
   "description": "Open-source collaborative browser editors for documents, Markdown, spreadsheets, presentations, and PDFs.",
   "keywords": [
     "office",
@@ -89,10 +89,23 @@
     "playground:visual": "playwright test --config playwright.config.ts",
     "playground:visual:spreadsheet-ribbon": "playwright test visual-tests/spreadsheet-ribbon.functional.spec.ts --config playwright.config.ts",
     "playground:visual:update": "playwright test --config playwright.config.ts --update-snapshots",
+    "performance:fixtures": "bun .a3s-test/performance/generate-fixtures.ts",
+    "performance:large-document-edits": "bun .a3s-test/performance/generate-fixtures.ts --documents-only && bun .a3s-test/performance/benchmark-document-edit.ts http://127.0.0.1:4175/ 3",
+    "performance:large-documents": "bun .a3s-test/performance/generate-fixtures.ts --documents-only && bun .a3s-test/performance/benchmark.ts --url http://127.0.0.1:4175/ --scenarios document-text,document-table --runs 3 --timeout-ms 180000",
+    "performance:large-spreadsheet-edits": "bun .a3s-test/performance/generate-fixtures.ts --spreadsheets-only && bun .a3s-test/performance/benchmark-spreadsheet-edit.ts http://127.0.0.1:4175/ 3",
+    "performance:large-spreadsheets": "bun .a3s-test/performance/generate-fixtures.ts --spreadsheets-only && bun .a3s-test/performance/benchmark.ts --url http://127.0.0.1:4175/ --scenarios spreadsheet --runs 5 --timeout-ms 180000",
+    "performance:pdf": "bun .a3s-test/performance/generate-fixtures.ts --pdf-only && bun .a3s-test/performance/benchmark-pdf.ts --url http://127.0.0.1:4175/ --runs 3 --timeout-ms 120000",
+    "performance:presentation": "bun .a3s-test/performance/generate-fixtures.ts --presentations-only && bun .a3s-test/performance/benchmark-presentation.ts --url http://127.0.0.1:4175/ --runs 3 --timeout-ms 120000 --modes default,content-visibility",
     "test": "rstest",
     "test:e2e": "bun run test:e2e:fixtures && a3s-test run tests/e2e/word-page-color.acl --json && a3s-test run tests/e2e/word-page-setup-phone.acl --json && a3s-test run tests/e2e/word-paragraph-layout-phone.acl --json && a3s-test run tests/e2e/word-list-formatting-phone.acl --json && a3s-test run tests/e2e/word-font-picker-phone.acl --json && a3s-test run tests/e2e/word-comments-drawer.acl --json && a3s-test run tests/e2e/word-comment-windowing.acl --json && a3s-test run tests/e2e/word-citations-phone.acl --json && a3s-test run tests/e2e/word-navigation-search.acl --json && a3s-test run tests/e2e/word-navigation-windowing.acl --json && a3s-test run tests/e2e/word-find-replace-phone.acl --json && a3s-test run tests/e2e/word-page-navigation.acl --json && a3s-test run tests/e2e/word-page-windowing.acl --json && a3s-test run tests/e2e/word-ai-question.acl --json && a3s-test run tests/e2e/word-picture-insert.acl --json && a3s-test run tests/e2e/word-picture-alt-text-phone.acl --json && a3s-test run tests/e2e/word-track-changes-phone.acl --json && a3s-test run tests/e2e/word-formatting-revision.acl --json && a3s-test run tests/e2e/word-paragraph-formatting-revision.acl --json && a3s-test run tests/e2e/word-review-conflict-phone.acl --json && a3s-test run tests/e2e/word-revision-windowing.acl --json && a3s-test run tests/e2e/word-table-phone.acl --json && a3s-test run tests/e2e/word-table-borders.acl --json && a3s-test run tests/e2e/word-theme-table.acl --json && a3s-test run tests/e2e/word-styled-table.acl --json && a3s-test run tests/e2e/word-multi-page-table.acl --json && a3s-test run tests/e2e/word-cross-reference-phone.acl --json && a3s-test run tests/e2e/word-bookmark-links-phone.acl --json && a3s-test run tests/e2e/word-notes-phone.acl --json && a3s-test run tests/e2e/word-fields-phone.acl --json && a3s-test run tests/e2e/spreadsheet-phone-rename.acl --json && a3s-test run tests/e2e/spreadsheet-phone-find.acl --json && a3s-test run tests/e2e/spreadsheet-phone-context-menu.acl --json && a3s-test run tests/e2e/spreadsheet-phone-task-pane.acl --json && a3s-test run tests/e2e/spreadsheet-maximum-sparse.acl --json && a3s-test run tests/e2e/presentation-cut-paste-focus.acl --json && a3s-test run tests/e2e/presentation-presenter-view.acl --json && a3s-test run tests/e2e/presentation-phone-chart-pane.acl --json && a3s-test run tests/e2e/presentation-phone-comments.acl --json && a3s-test run tests/e2e/pdf-thumbnail-keyboard.acl --json && a3s-test run tests/e2e/pdf-page-drawer-phone.acl --json && a3s-test run tests/e2e/markdown-phone-modes.acl --json && a3s-test run tests/e2e/office-sidebar-breakpoint.acl --json && a3s-test run tests/e2e/office-docs-navigation.acl --json && a3s-test run tests/e2e/collaboration-playground-entry.acl --json && a3s-test run tests/e2e/collaboration-document-comments.acl --command-timeout-ms 120000 --json && a3s-test run tests/e2e/collaboration-document-suggestions.acl --command-timeout-ms 120000 --json && a3s-test run tests/e2e/collaboration-participants.acl --json && a3s-test run tests/e2e/collaboration-pdf-annotations.acl --json && a3s-test run tests/e2e/collaboration-spreadsheet-cells.acl --json && a3s-test run tests/e2e/collaboration-presentation-elements.acl --json",
     "test:e2e:check": "bun run test:e2e:fixtures && a3s-test check tests/e2e/word-page-color.acl --json && a3s-test check tests/e2e/word-page-setup-phone.acl --json && a3s-test check tests/e2e/word-paragraph-layout-phone.acl --json && a3s-test check tests/e2e/word-list-formatting-phone.acl --json && a3s-test check tests/e2e/word-font-picker-phone.acl --json && a3s-test check tests/e2e/word-comments-drawer.acl --json && a3s-test check tests/e2e/word-comment-windowing.acl --json && a3s-test check tests/e2e/word-citations-phone.acl --json && a3s-test check tests/e2e/word-navigation-search.acl --json && a3s-test check tests/e2e/word-navigation-windowing.acl --json && a3s-test check tests/e2e/word-find-replace-phone.acl --json && a3s-test check tests/e2e/word-page-navigation.acl --json && a3s-test check tests/e2e/word-page-windowing.acl --json && a3s-test check tests/e2e/word-ai-question.acl --json && a3s-test check tests/e2e/word-picture-insert.acl --json && a3s-test check tests/e2e/word-picture-alt-text-phone.acl --json && a3s-test check tests/e2e/word-track-changes-phone.acl --json && a3s-test check tests/e2e/word-formatting-revision.acl --json && a3s-test check tests/e2e/word-paragraph-formatting-revision.acl --json && a3s-test check tests/e2e/word-review-conflict-phone.acl --json && a3s-test check tests/e2e/word-revision-windowing.acl --json && a3s-test check tests/e2e/word-table-phone.acl --json && a3s-test check tests/e2e/word-table-borders.acl --json && a3s-test check tests/e2e/word-theme-table.acl --json && a3s-test check tests/e2e/word-styled-table.acl --json && a3s-test check tests/e2e/word-multi-page-table.acl --json && a3s-test check tests/e2e/word-cross-reference-phone.acl --json && a3s-test check tests/e2e/word-bookmark-links-phone.acl --json && a3s-test check tests/e2e/word-notes-phone.acl --json && a3s-test check tests/e2e/word-fields-phone.acl --json && a3s-test check tests/e2e/spreadsheet-phone-rename.acl --json && a3s-test check tests/e2e/spreadsheet-phone-find.acl --json && a3s-test check tests/e2e/spreadsheet-phone-context-menu.acl --json && a3s-test check tests/e2e/spreadsheet-phone-task-pane.acl --json && a3s-test check tests/e2e/spreadsheet-maximum-sparse.acl --json && a3s-test check tests/e2e/presentation-cut-paste-focus.acl --json && a3s-test check tests/e2e/presentation-presenter-view.acl --json && a3s-test check tests/e2e/presentation-phone-chart-pane.acl --json && a3s-test check tests/e2e/presentation-phone-comments.acl --json && a3s-test check tests/e2e/pdf-thumbnail-keyboard.acl --json && a3s-test check tests/e2e/pdf-page-drawer-phone.acl --json && a3s-test check tests/e2e/markdown-phone-modes.acl --json && a3s-test check tests/e2e/office-sidebar-breakpoint.acl --json && a3s-test check tests/e2e/office-docs-navigation.acl --json && a3s-test check tests/e2e/collaboration-playground-entry.acl --json && a3s-test check tests/e2e/collaboration-document-comments.acl --json && a3s-test check tests/e2e/collaboration-document-suggestions.acl --json && a3s-test check tests/e2e/collaboration-participants.acl --json && a3s-test check tests/e2e/collaboration-pdf-annotations.acl --json && a3s-test check tests/e2e/collaboration-spreadsheet-cells.acl --json && a3s-test check tests/e2e/collaboration-presentation-elements.acl --json",
     "test:e2e:fixtures": "bun scripts/create-e2e-fixtures.ts",
+    "test:e2e:large-documents": "bun .a3s-test/performance/generate-fixtures.ts --documents-only && a3s-test run tests/e2e/word-large-document-windowing.acl --json",
+    "test:e2e:large-documents:check": "a3s-test check tests/e2e/word-large-document-windowing.acl --json",
+    "test:e2e:large-pdf": "bun .a3s-test/performance/generate-fixtures.ts --pdf-only && a3s-test run tests/e2e/pdf-large-windowing.acl --json",
+    "test:e2e:large-pdf:check": "a3s-test check tests/e2e/pdf-large-windowing.acl --json",
+    "test:e2e:large-presentation": "bun .a3s-test/performance/generate-fixtures.ts --presentations-only && a3s-test run tests/e2e/presentation-large-windowing.acl --json",
+    "test:e2e:large-presentation:check": "a3s-test check tests/e2e/presentation-large-windowing.acl --json",
     "test:e2e:initial-focus": "a3s-test run tests/e2e/office-editor-initial-focus.acl --json",
     "test:e2e:initial-focus:check": "a3s-test check tests/e2e/office-editor-initial-focus.acl --json",
     "test:e2e:docs": "a3s-test run tests/e2e/office-docs-navigation.acl --json",
@@ -117,8 +130,12 @@
     "test:e2e:writer-ribbon:check": "a3s-test check tests/e2e/word-ribbon-collapse.acl --json",
     "test:e2e:spreadsheet-ribbon": "a3s-test run tests/e2e/spreadsheet-ribbon-alignment.acl --json",
     "test:e2e:spreadsheet-ribbon:check": "a3s-test check tests/e2e/spreadsheet-ribbon-alignment.acl --json",
+    "test:e2e:spreadsheet-cell-fill": "a3s-test run tests/e2e/spreadsheet-cell-fill.acl --json",
+    "test:e2e:spreadsheet-cell-fill:check": "a3s-test check tests/e2e/spreadsheet-cell-fill.acl --json",
     "test:e2e:spreadsheet-maximum-sparse": "a3s-test run tests/e2e/spreadsheet-maximum-sparse.acl --json",
     "test:e2e:spreadsheet-maximum-sparse:check": "a3s-test check tests/e2e/spreadsheet-maximum-sparse.acl --json",
+    "test:e2e:spreadsheet-large": "bun .a3s-test/performance/generate-fixtures.ts --spreadsheets-only && a3s-test run tests/e2e/spreadsheet-large-controlled-editing.acl --json",
+    "test:e2e:spreadsheet-large:check": "a3s-test check tests/e2e/spreadsheet-large-controlled-editing.acl --json",
     "test:e2e:writer-shortcuts": "a3s-test run tests/e2e/word-wps-shortcuts.acl --json",
     "test:e2e:writer-shortcuts:check": "a3s-test check tests/e2e/word-wps-shortcuts.acl --json",
     "test:e2e:writer-layout": "a3s-test run tests/e2e/word-insert-page-layout.acl --json",
@@ -140,6 +157,7 @@
     "@embedpdf/fonts-hebrew": "1.0.0",
     "@embedpdf/fonts-latin": "1.0.0",
     "@embedpdf/fonts-sc": "1.0.0",
+    "@embedpdf/models": "2.14.3",
     "@embedpdf/pdfium": "2.14.3",
     "@embedpdf/react-pdf-viewer": "2.14.3",
     "@fortune-sheet/core": "^1.0.4",