feat(recovery): add exact job retry CAS

[rldyourmnd]

Outcome

Add a separate recover-provider-job-retry operation for an exact job-bound provider circuit after its external root cause has been repaired.

It requires all of:

• GARM stopped and execution as the garm service account
• exact scale-set:<entity>:<id>:job:<job> key and matching entity/scale-set arguments
• record attempts == 3, terminal state and exact recoverable error class
• CAS on exact updated_at plus the existing one-minute recovery grace
• private bounded queue journal schema 4
• the same exact job currently live in queued|acquiring|acquired|assigned
• dry-run by default; atomic apply increments the retry journal generation

The existing domain recovery remains unchanged and still rejects concrete keys. Invalid/absent/running queue proof performs zero mutation.

Verification

• make verify
• focused package tests and race tests