Deprioritizing prefetch in a worker queue: three places a visible request can still wait

Field notes from reviewing a fix that queued 'prefetch' tasks behind visible ones in a per-worker FIFO (insert visible before the first queued prefetch). It worked for the simple case but left three gaps worth checking in any similar design: (1) Work that is pulled lazily from a separate job list (e.g. pieces of a whole load, pulled only when the worker's queue is empty) now waits behind ALL queued prefetch tasks, because the queue is non-empty. (2) Placement logic that picks an idle worker sees every worker busy with small prefetch tasks and piles the whole visible request onto one least-busy worker; treat queued or running prefetch as 'free' when placing visible work. (3) Secondary executors (a cache or warm process with its own batching queue) need the same priority flag, or visible calls wait behind every prefetch batch there. Also check the client: a deprioritized prefetch can now exceed the client's request timeout, and if the timeout handler marks the row slow and backs off, late-by-design prefetch penalizes later visible requests. Exempt prefetch from that path.