fix(granola): add incremental sync with early termination

[aculich]

Summary

Reduces API calls and sync time for Granola integration by:

1. Passing updated_since parameter to the API (works if API supports it)
2. Early termination after 20 consecutive unchanged documents (fallback)

Fixes #435

Problem

The Granola sync fetches ALL documents from the API on every sync cycle, even when most documents haven't changed. With 500 meetings, this means ~50 API calls with 1-second delays = 50+ seconds per sync, even when nothing changed.

Solution

Dual strategy with graceful fallback:

1. Try updated_since parameter - If the Granola API accepts this parameter, it will return only documents updated since the last sync, drastically reducing API calls.

2. Early termination fallback - If the API ignores the parameter (returning all docs anyway), the sync stops after hitting 20 consecutive documents that haven't changed since the last sync. This works because the API typically returns documents in a consistent order.

Changes

apps/x/packages/core/src/knowledge/granola/sync.ts:

1. Add EARLY_TERMINATION_STREAK = 20 constant
2. Modify getDocuments() to accept optional updatedSince parameter
3. Pass state.lastSyncDate to API request as updated_since
4. Track unchangedStreak counter in sync loop
5. Break out of pagination when streak reaches threshold

Testing

1. Enable Granola integration with an account that has many meetings
2. Start the app and observe console logs for sync behavior
3. On first sync: all documents should be fetched (no early termination)
4. On subsequent syncs: should see "Early termination: 20 consecutive unchanged docs" log
5. If a new Granola meeting is created, it should be picked up on next sync

Impact

• First sync: No change in behavior (all docs fetched)
• Subsequent syncs: Dramatically faster when no changes (~2-3 API calls vs ~50+)
• API rate limit consumption significantly reduced

Made with Cursor

[vercel]

@aculich is attempting to deploy a commit to the RowBoat Labs Team on Vercel.

A member of the Team first needs to authorize it.

[Bortlesboat]

This still looks like the right direction for #435. The issue description itself already proposes basically this two-part approach: try updated_since, then fall back to early termination when the API ignores it.

Right now the main blocker seems to be that the branch is conflicted with main, not that the idea is unsound. If the author is still around, a rebase would make this much easier to evaluate without reopening the design discussion from scratch.