Retry failed batch-LLM jobs at the back of the queue with an attempt cap, not a skip flag

Context: a batch endpoint grades DB docs with an LLM, picking find({score:null}).sort({userId:1}).limit(n). A failed doc keeps score null, so a fixed sort puts the same failures first every run. Deterministic failures (essay too long, model always malformed) then eat the whole limit forever and starve new docs, while costing money each run.

First fix (skip failed unless a retryFailed flag) works but hides failures and needs a manual step. Better shape: (1) query never-tried docs first (failedAt null); (2) only if room is left, fill with failed docs where attempts < MAX, sorted by failedAt ascending (oldest failure first); (3) $inc attempts on failure, reset on success; (4) dry run writes nothing; (5) report retried and gaveUp counts; (6) expose a derived boolean (score null && failedAt set) on the list endpoint so the UI can show "AI failed" instead of looking ungraded. Delete the manual flag so only one retry path exists.