No actionable comments were generated in the recent review. 🎉

sequenceDiagram
  participant QueueOp as Queue Operation
  participant Queue as PriorityJobQueue
  participant Job as Job
  QueueOp->>Queue: complete, skip, cancel, or cancel_all
  Queue->>Queue: _promote_waiting_deps()
  Queue->>Job: check all dependencies completed
  alt Dependencies ready
    Queue->>Job: mark WAITING_DEPS as QUEUED
  end

_{Comment @coderabbitai help to get the list of available commands.}

queue.py:233 - move WAITING_DEPS jobs back to QUEUED once their dependencies finish, the way line 339 already does for WAITING_PRIORITY (nothing ever moves a job out of WAITING_DEPS. Two jobs linked by needs land at the same priority by default, so the dependent is stranded, the level never completes, and _dispatch_loop spins forever) pre-existing

queue.py:233 - move WAITING_DEPS jobs back to QUEUED once their dependencies finish, the way line 339 already does for WAITING_PRIORITY (nothing ever moves a job out of WAITING_DEPS. Two jobs linked by needs land at the same priority by default, so the dependent is stranded, the level never completes, and _dispatch_loop spins forever) pre-existing

Added _promote_waiting_deps() (mirrors the WAITING_PRIORITY -> QUEUED pattern at line 339) and call it from mark_completed, mark_skipped, cancel, and cancel_all before _check_priority_advance.
Only jobs at the current priority level are promoted.

Two things left.

1. cancel_all doesn't discard from _running_keys. READY jobs are cancellable and next_ready has already put the key in _running_keys, but the loop only adds to _completed_keys. On this branch: enqueue 3, one next_ready(), then cancel_all() gives running_count == 1 and pending_count == -1. It shows up as a phantom running job in the summary after stop_on_first_failure. Same discard(key) that mark_completed already does. test_cancel_all never calls next_ready, so it needs that case with running_count == 0 and pending_count >= 0.

2. holder.join(timeout=5) needs an assert not holder.is_alive() after it. Without it a stuck holder passes green and leaves a live thread that blocks pytest at exit. The consumer tests already do this.

Minor, not a blocker: checker.join(timeout=0.05) can't tell "blocked on the lock" from "not scheduled yet", so on a busy runner the guard passes without checking anything. An Event set just before the is_done read would close it.

Unrelated to your change: all_dependencies_met only checks _completed_keys, which fills on failure too, so a dependent runs after its dependency fails. It predates this PR. I'll open a separate issue.