feat: configure canvas pages sent for note analysis

Canvas analysis previously sent every page. Add a Magic Notes setting for
1-8 canvas pages, defaulting to 1 for new and historical settings. Select
the first N pages in current canvas order before capture and use the same
text/image range for saved entries, drafts and canvas-source todos.

Keep full 50-page saving/export, expose the same setting through config
tools, and use the shared settings select with an inline help icon.
Include pagination, persistence, production IPC and isolated Electron
regressions, including text-only input and scaled page boundaries.

Validation: focused feature tests and isolated Electron regressions passed;
the final compact UI passed 14 tests. Typecheck and lint were reported
passing by the coordinating agent. No full-suite pass is claimed because
the completed full test command exposed no summary. No live model calls
were made for this commit preparation.

Release note: 魔法笔记新增“发送画布页数”设置，可选 1～8 页，默认 1 页，按当前画布顺序发送前 N 页的文字和图片；标题旁问号可查看说明，保存和导出完整画布不受影响。

docs/features/magic-notes/README.md

@@ -61,8 +61,11 @@ changing saved content or export dimensions; see the
 [canvas zoom reference](../../../src/renderer/src/magic-canvas/README.md#view-zoom).
 
 Drafts, saved entries and todos sourced from canvases support AI comments using
-the default model's image-input capability. Image-capable models receive complete
-page captures and extracted text. Text-only models receive an explicitly labelled
+the default model's image-input capability. The Canvas pages to send setting
+selects the first 1-8 pages in current canvas order, defaulting to 1. It counts
+pages in one canvas, not entries, and leaves full-canvas saving and export intact.
+Image-capable models receive complete captures and extracted text for that prefix.
+Text-only models receive the same page range as an explicitly labelled
 text fallback; a purely visual canvas entry or draft reports that an image-capable
 default model is required. Canvas comments are requested manually in `immediate`
 mode, or after a manual save in `after-save-auto` mode. Analysis failure does not

docs/features/magic-notes/progress.md

@@ -1,5 +1,56 @@
 # Magic Notes Progress
 
+## 2026-09-19: Compact Page Count Setting
+
+- The page-count select now uses the existing `.field` settings control styling.
+  A title-adjacent `CircleHelp` icon uses the shared `icon-button` and native
+  `title` hint; the full description is screen-reader-only and remains associated
+  with both controls through `aria-describedby`. The compact row uses flex and
+  keeps the label and help icon together with inline-flex, including at narrow
+  widths. Control styling and the native hint mechanism remain shared. Page
+  selection and persistence behavior are unchanged.
+- `npx vitest run src/renderer/src/ApplicationCenter.test.tsx` passed 14 tests.
+  Focused ESLint and a temporary TypeScript configuration rooted at
+  `ApplicationCenter.tsx` with its imported dependencies passed. No full suite,
+  full typecheck, Electron/live-model test or model request was run in this UI pass.
+- The commit preparation had only inspected status, diffs and history. The index
+  remained empty and HEAD remained `ecf09f4`; no staging or commit was performed.
+  The later same-line adjustment passed the same 14-test UI suite before commit.
+- The coordinating agent subsequently reported full typecheck and lint passing.
+  The first full test command timed out at 200 seconds; the second completed but
+  exposed only the RUN banner without a summary. No full-suite pass is claimed.
+
+## 2026-09-19: Canvas Pages to Send
+
+- Implemented the page-count preference and input selection described in
+  [Canvas Analysis and Comments](./technical-design.md#canvas-analysis-and-comments).
+  Application settings and configuration tools persist the same field; historical
+  missing values read as 1. All three production analysis handlers use the saved
+  count for text and images. Saving retains the complete canvas.
+- Focused verification passed **15 files / 271 tests**:
+  `npx vitest run src/main/application-settings-store.test.ts src/main/magic-notes/magic-note-analyzer.test.ts src/main/magic-notes/canvas-ipc.integration.test.ts src/renderer/src/ApplicationCenter.test.tsx src/renderer/src/MagicNotesWorkspace.test.tsx src/renderer/src/MagicCanvasEditor.test.tsx src/renderer/src/MagicCanvasThumbnail.test.tsx src/renderer/src/magic-canvas src/shared/goodbuddy-config-contracts.test.ts src/shared/magic-note-pages.test.ts src/main/goodbuddy-config-service.test.ts src/renderer/src/UpdateSettingsSection.test.tsx`.
+  Coverage includes 1/8 boundaries, invalid counts, missing historical fields,
+  restart persistence, settings events, config-tool writes, 50-page source
+  preservation, current page-array order, selected PDF/object/flow text, missing
+  flow text and selected pages with no readable text.
+- `npx vitest run tests/magic-notes-analysis.electron.test.ts` passed **2 tests**,
+  one per image capability. The real settings form changes the default from 1 to
+  2 through production preload/IPC and the file-backed settings store. Workspace,
+  Quill and Fabric then analyze saved entries, edited saves, new drafts, new saves
+  and canvas-source todos through registered IPC/analyzer/SQLite. Explicit page
+  breaks and automatic overflow both exclude later-page text. Browsing page three
+  at 75% zoom still sends pages one and two. This exposed scaled DOMRect rounding
+  at column starts; the corrected boundary retains each page's first character.
+- `npx vitest run src/renderer/src/App.test.tsx -t application` passed **5 tests**
+  with 279 unrelated cases filtered out. Focused ESLint and `git diff --check`
+  passed. Full `npm test`, typecheck and lint were not run, per user request.
+- Electron uses an isolated feature window, real application code and temporary
+  data, with DOM actions/Quill edits and substituted model output. It does not
+  validate a packaged full App, physical input devices or a live provider.
+  Real model calls: **0**. Analysis uses desktop `createDefaultModelRuntime`;
+  deployed gbagent and desktop-to-Agent execution are unaffected. Existing user
+  changes were retained; no commits or user-data writes were made by this task.
+
 ## 2026-09-19: Canvas Content and History Preservation
 
 - Page-break deletion now removes only the embed, retaining existing paragraph

docs/features/magic-notes/technical-design.md

@@ -75,9 +75,8 @@ An abort listener calls core `destroy()` immediately: PDF render cancellation an
 the Fabric abort happen synchronously, before waiting for capture to settle.
 The `finally` path awaits the same destruction promise and removes the temporary
 surface, without destroying twice or waiting for capture before cancellation.
-AI capture retains its existing
-all-page behavior. Main canvas page height, wheel handling and file storage are
-unchanged.
+AI capture uses the configured page prefix described under Canvas Analysis and
+Comments. Main canvas page height, wheel handling and file storage are unchanged.
 
 Background refresh preserves dirty editing content and its original optimistic
 revision, including externally deleted entries. Clean editors adopt refreshed
@@ -249,18 +248,39 @@ model profile's `supportsImageInput` (or the resolved default setting when no
 profile is selected). Main resolves settings again for the actual request. This
 does not change an Agent runtime or desktop-to-Agent protocol.
 
+`magicNoteCanvasPageCount` is an application preference: an integer from 1 to 8,
+defaulting to 1 when absent in historical settings. The application center and
+`goodbuddy_config` application.update/get share the settings schema, store and
+change event. It counts pages within the canvas being analyzed, never entries.
+The first N pages follow the current `content.pages` array order, independent of
+the viewed page, page IDs or PDF source page numbers. Fewer than N pages selects
+the whole canvas. Saving and PDF export retain the 50-page canvas limit.
+
+Renderer and Main share `selectMagicNotePages`. The core selects the prefix before
+compositing, so it never captures 50 pages just to discard 49. Analysis capture
+returns the selected pages' Quill text measured from the actual CSS columns;
+binary searches locate the character boundaries for explicit and automatic page
+breaks. Text-only requests skip image compositing but still read page text.
+The transient `canvasPageText` IPC field carries these page IDs and text, without
+changing saved canvas content. Main reads the persisted page count for all three
+handlers, selects page text, PDF text, objects and images in the same order, and
+ignores captures outside the prefix. A partial canvas with flow requires matching
+page text; missing data fails instead of including the full cross-page flow.
+Prompts state the selected and total page counts. A changed setting applies to
+the next analysis; existing comments are not automatically regenerated.
+
 | Default model capability | Analysis input and result |
 | --- | --- |
-| Image input supported | Renderer captures all pages; Main checks page IDs and PNG/JPEG encoding. Images are ordered by page alongside flow, PDF and object text. Comments carry `inputMode: canvas-images`. |
-| No image input | No page capture is required. Extracted text is used with an explicit instruction that handwriting, images and layout were not seen. Comments carry `inputMode: text-fallback`. |
-| No image input and no extracted text in an entry/draft | Analysis fails with guidance to switch to an image-capable default model; no visual understanding is claimed. |
+| Image input supported | Renderer captures the selected prefix; Main requires each selected page's PNG/JPEG capture and sends it alongside that page's flow, PDF and object text. Comments carry `inputMode: canvas-images`. |
+| No image input | Only text from the selected prefix is sent, with an explicit instruction that handwriting, images and layout were not seen. Comments carry `inputMode: text-fallback`. |
+| No image input and no extracted text in the selected entry/draft pages | Analysis fails with guidance to switch to an image-capable default model; no visual understanding is claimed. |
 
 Todo analysis includes its title and instructions plus the source canvas context,
 and uses the same capability and capture rules. Non-canvas content keeps the text
 analysis path. All analysis remains read-only with no tool calls.
 
 Saved-entry requests accept `expectedRevision`; todo requests accept
-`sourceEntryRevision`. Nonempty `canvasImages` requires the corresponding revision,
+`sourceEntryRevision`. Nonempty `canvasImages` or `canvasPageText` requires the corresponding revision,
 even when the resolved model uses text fallback. Main loads the source after
 resolving settings and checks the supplied revision before model invocation.
 Existing rich-text callers may omit these fields. Entry writeback retains its