The per-repo job aggregation filtered on the three `github.*` ref kinds, omitting `tangled.create-repo` -- the kind enrollment and resync actually enqueue. A create-repo job that was retrying or sitting at `failed` could therefore never move a repo to `behind` or `error`. One line plus a comment in `dashboard.get.ts`. The kind already carries the same `installationId`/`githubRepoId` payload shape, so the existing grouping is unchanged. No dependency on any other fix. Tests: full unit suite passes. No tests added: the change is a kind string inside the raw `sql` filter in the `dashboard.get.ts` route handler, and no route handler in this repo has unit coverage -- the suite targets utils and pure functions. The consumer of that query's output, `repoHealth()`, is already covered directly.