Destroy orphan BlockedExecution rows when the job class no longer resolves

[julik]

Fixes #784.

Summary

Fixes a DelegationError raised on every SolidQueue::Dispatcher::ConcurrencyMaintenance tick when a BlockedExecution row references a Job whose class_name no longer safe_constantizes — e.g. because the concurrency-limited ActiveJob class was renamed or removed between deploys.

Job#acquire_concurrency_lock already gates on concurrency_limited? (which checks job_class.present?), but BlockedExecution#acquire_concurrency_lock — invoked from the dispatcher-driven release path — went straight to Semaphore.wait(job), which calls job.concurrency_limit, which is delegated to a nil job_class:

ActiveSupport::DelegationError: concurrency_limit delegated to job_class, but job_class is nil
  app/models/solid_queue/job/concurrency_controls.rb:11:in 'concurrency_limit'
  app/models/solid_queue/semaphore.rb:86:in 'limit'
  app/models/solid_queue/semaphore.rb:58:in 'attempt_creation'
  app/models/solid_queue/semaphore.rb:46:in 'wait'
  app/models/solid_queue/blocked_execution.rb:65:in 'acquire_concurrency_lock'
  app/models/solid_queue/blocked_execution.rb:49:in 'block (2 levels) in release'
  ...
  lib/solid_queue/dispatcher/concurrency_maintenance.rb:40

Because the exception is raised inside BlockedExecution#release's transaction, the row is never destroyed or promoted, and the dispatcher re-picks the same orphans forever — flooding error reporters and starving any legitimately blocked jobs that share the concurrency key.

Changes

• BlockedExecution#release: if job.job_class is nil, destroy the orphan row and emit orphaned: true in the release_blocked instrumentation payload. Otherwise fall through to the existing acquire/promote/destroy path.
• BlockedExecution#set_expires_at: fall back to SolidQueue.default_concurrency_control_period when the class does not resolve (guards the same latent bug at create time).
• Job#job_class promoted from private to public so BlockedExecution can consult it.
• New unit test covering the orphan release path.

Test plan

• bin/rails test test/models/solid_queue/blocked_execution_test.rb — new test passes.
• bin/rails test test/models/ — full model suite green (97 runs, 585 assertions, 0 failures).
• CI across mysql/postgres/sqlite.

[julik]

@rosa any chance getting a look at this?

[rosa]

Hey @julik, sorry for the delay! I've been on-call for a couple of weeks, and working on a lot of other stuff, but I'm on vacation this week and catching up with PRs. I have a few in the queue before this one, so it'll take a few days more.

[rosa]

Hey @julik! Sorry again for the delay. I'm looking into this one now. I wonder: would it make sense, instead of just discarding these jobs, to mark them as failed instead, so you have the chance of doing something with them? Or discarding if you don't need them, that's always an option.

[julik]

I will check - I think anything but having things "occupying a slot" will work

[julik]

Done - hope this works ;-)

[rosa]

Thanks @julik! I've pushed a couple of small tweaks to your solution, but basically this is it. Thank you!