Potential deadlock in WalkPool/SchemaWalkPool WaitForCompletion (missed wakeup)

[califlower]

Summary

We're seeing intermittent test hangs/timeouts when doctor is used via vacuum rule execution. Goroutines end up blocked in WaitForCompletion() waiting on sync.Cond.

From inspection of v0.0.49, both worker pools can miss a wakeup:

• model/high/v3/walk_pool.go (WalkPool.worker)
• model/high/v3/schema_pool.go (SchemaWalkPool.processItem)

Both decrement inFlight and call cond.Broadcast() without holding mu.

Reason (as far as I can tell)

WaitForCompletion() does:

1. lock mu
2. check inFlight > 0
3. call cond.Wait()

Workers do:

1. inFlight.Add(-1)
2. if zero, call cond.Broadcast() (currently unlocked)

Because the condition transitions and broadcast are not synchronized with the same lock used by waiters, there is a missed-wakeup window between the waiter's condition check and entering Wait(). If inFlight reaches zero and broadcast happens in that window, waiter can sleep indefinitely.

Trace

The stuck goroutines consistently include:

• github.com/pb33f/doctor/model/high/v3.(*WalkPool).WaitForCompletion
• github.com/pb33f/doctor/model/high/v3.(*DrContext).WaitForCompletion
• github.com/pb33f/doctor/model/high/v3.(*Document).Walk
• caller waiting in github.com/daveshanley/vacuum/motor.ApplyRulesToRuleSet

Proposed fix

Guard Broadcast() with mu in both pool implementations:

if p.inFlight.Add(-1) == 0 {
    p.mu.Lock()
    p.cond.Broadcast()
    p.mu.Unlock()
}

Applied to:

• model/high/v3/walk_pool.go
• model/high/v3/schema_pool.go

Happy to open a PR with this change if you'd like.