feat(app): add selection-aware comments in fullscreen

[aryeko]

Summary

This PR adds selection-aware follow-up editing to the fullscreen Excalidraw widget, so users can point the model at specific diagram elements instead of only giving whole-diagram feedback.

It introduces:

• fullscreen element selection tracking
• a selection badge in the editor toolbar
• anchored Comment actions for selected elements
• structured widget context for selected elements, recent edits, and recent comments
• tests for selection and widget-context serialization/store behavior

Problem

Today users can manually edit diagrams in fullscreen, but they cannot easily tell the model “this box” or “these two nodes” when asking for follow-up changes. That makes iterative refinement much less precise.

What Changed

Selection-aware context

• Track selected Excalidraw elements in fullscreen
• Exclude bound text from the visible count/context
• Resolve labels from bound text where available
• Send structured selection context to the model for future turns

Anchored comments

• Show a contextual Comment affordance near the current selection
• Let users send follow-up requests scoped to the selected element(s)
• Only enable this when the host supports both ui/message and ui/update-model-context
• Only persist a comment in widget context after the host accepts the message

Widget context refactor

• Replace the old edit-context flow with a more structured widget-context controller
• Track:
  • checkpoint id
  • current selection
  • recent edit summaries
  • recent accepted comments
  • last action metadata
• Improve the visible text payload sent via updateModelContext

Editor/persistence cleanup

• Split scene parsing, selection anchoring, viewport handling, and persistence into focused modules
• Preserve checkpoint/local persistence behavior for fullscreen edits
• Keep fullscreen editor state in sync when a new final tool result arrives

Why This Is Useful

This makes the app feel much more conversational and iterative:

• users can select a single node and ask for a targeted change
• the model gets structured context instead of inferring intent from a generic follow-up
• follow-up requests are safer because failed sends no longer pollute future context

Try It

Example follow-ups once the diagram is open in fullscreen:

• Select the Auth Service element and ask: use green
• Select one node and ask: make this label clearer
• Select two elements and ask: align these and add spacing

Demo

• Live MCP endpoint: https://excalidraw-mcp-nine.vercel.app/mcp
• Demo: attached trimmed recording showing select -> comment -> updated result
  

Testing

Ran:

• npm test
• npm run build

Added tests for:

• selection detail/summary behavior
• widget context store transitions
• widget context serialization

Verified

Verified manually:

• selecting a single element updates the badge and selection context
• bound text is excluded from the visible count/context
• clicking Comment opens the anchored composer near the selection
• sending a selection-scoped request works end-to-end
• failed sends do not appear as accepted recent comments in widget context
• exiting fullscreen preserves user edits
• new final tool results sync correctly into the fullscreen editor

Notes

Happy to split this if maintainers would prefer the selection badge, widget-context refactor, and anchored comments to land separately.

[vercel]

@aryeko is attempting to deploy a commit to the Excalidraw Team on Vercel.

A member of the Team first needs to authorize it.

[aryeko]

@antonpk1 Any chance you can have a look? I think this feature will be very useful and make will make excelidraw mcp more usable and advanced than others (eg draw.io)

If the refactor is too much I can focus the pr to the functionality only