fix: revoke stale per-repo grants (#874)

[ionfwsrijan]

Problem

Per-repo grants were only reconciled when a repo already had a grant, so stale installation_user_repos rows (revoked access, removed members) were never deleted ? leading to phantom repository access and stale grants.

Fix

• New pure reconcileRepoGrants(existing, proposed) diff returns { toUpsert, toDelete } and runs unconditionally.
• Stale repo-grant rows are deleted even when the proposed set is empty (all existing rows become stale), decoupled from the highestRepoGrant gate that still controls junction pushes.

Files changed

• src/lib/maintainer/discover.ts
• src/inngest/functions/maintainer-discover.ts

Testing

• Full MergeShip suite passes.
• Tests cover unconditional reconciliation, empty-proposed deletion, and unchanged-grant no-ops.

Closes #874

[vercel]

@ionfwsrijan is attempting to deploy a commit to the codersogs-3057's projects Team on Vercel.

A member of the Team first needs to authorize it.

[github-actions]

Hey @ionfwsrijan

You have 7 open PRs right now. The limit is 3 at a time.

Please get your existing PRs merged or closed before opening new ones:

• fix: usage week aligns with UTC calendar week (#875) #880 - fix: usage week aligns with UTC calendar week ([BUG] Usage page's "week XP" is a rolling 7×24h window, not a calendar week — inconsistent with the calendar-day "today" and with the platform's weekly analytics #875)
• fix: prevent self-merge recommended-XP loop (#873) #878 - fix: prevent self-merge recommended-XP loop ([BUG] Recommended-merge XP has no self-merge guard — a repo owner can claim and merge their own repo's issues for up to 400 XP, bypassing the anti-abuse rule enforced on unrecommended merges #873)
• fix: reconcile merged-PR XP across all installs and pages (#872) #877 - fix: reconcile merged-PR XP across all installs and pages ([BUG] reconcileMergedPrXp never scans beyond the first 50 installs, first 20 repos, or the first 100 closed PRs — lost merge XP is permanently unrecoverable on busy repos #872)
• fix: exact maintainer status check with negative-cache lockdown (#871) #876 - fix: exact maintainer status check with negative-cache lockdown ([BUG] isUserMaintainer truncates the grant check to 20 junction rows and caches the result for 1h — maintainers with many installs are wrongly locked out #871)
• fix: maintainer queue pagination with in-SQL filters and exact counts #868 - fix: maintainer queue pagination with in-SQL filters and exact counts

This PR will remain open but won't be reviewed until you're under the limit. See our Contributing Guidelines for details.

[ionfwsrijan]

@jakharmonika364 @Soumya-codr Please review this

[jakharmonika364]

The diff-computation logic (reconcileRepoGrants) is correct and well tested, but the write side has a real bug: swap the installation_user_repos .insert(...) for .upsert(..., { onConflict: 'installation_id,user_id,repo_full_name' }) so permission-level changes on already-granted repos actually apply instead of silently failing and leaving stale grants in place - otherwise this PR trades one stale-grant bug for another.