[vboctor]

Summary

• New `sync_errors` journal table: per-entity fetch failures are recorded and skipped, not aborting the run; retried (up to 3 attempts) on the next sync.
• Cursor persisted after every batch — interrupted full syncs resume from the last batch instead of restarting.
• MantisHub initial sync uses a `snapshotCeiling` pinned at run start (`filter_updated_before`) so DESC pagination stays stable across resumes; list endpoint uses `select=id,updated_at`; `PAGE_SIZE=25` everywhere; per-issue detail fetches run 5-way concurrent.
• MantisHub users phase aborts after 10 fetch failures and emits shell entities for the remaining users from data already collected during the issues phase. User failures are never journaled.
• `startSyncJob` was double-calling `ensureSyncJob` and silently exiting because the guard saw the synchronously-seeded entry as already active — fixed by claiming the seeded entry by status.
• UI: error counter shown in progress card; `counter-deleted` is grey, new `counter-failed` is red. Status messages drop the collection-name and inline counter redundancy.

Test plan

• `bun run ci` passes locally (markdown lint, typecheck, all test suites, UI build, worker build)
• Full sync of a large MantisHub project completes; cursor advances per batch and resume works after a kill
• Inject a 5xx for a single issue and confirm it lands in `sync_errors` (visible as "N failed" in the UI) without aborting
• Re-run sync and confirm the journaled entry retries; after 3 failed attempts, retries stop
• Users-phase abort triggers when most user fetches return 403; shell entities are emitted for the rest
• Click "Sync" — UI no longer hangs on "starting"; status moves through "syncing" → progress messages

🤖 Generated with Claude Code