fix: violations and AI referrals are neither sent twice nor lost during an outage

[cport1]

For WebDecoy/app#1245, item 7.

• Violations sent twice: every request that records a violation drains the spool at shutdown, and the cron drains too. Two drains at once read the same oldest rows and sent them twice.
  • Drains now take an atomic add_option() lock. A lock older than 60s, left by a request that died, is taken over.
• Violations lost in an outage: each refused send spent an attempt, so two refused drains discarded the spool.
  • A refusal (no answer, 429, 5xx) now keeps the rows and their attempts and starts the shared 60s pause.
  • Only a rejected batch (other 4xx) spends attempts.
• AI referrals lost on 429: the batch was deleted.
  • It's now kept and resent under the same id, which ingest deduplicates, and flushing waits out the pause.

Tests: php tests/run.php gives 196 passed (7 new, in ViolationDrainTest.php).

• Mutation-checked: removing the lock, the refusal handling, or the referral 429 handling each fails a test.
• The test add_option stub now fails on an existing option, as WordPress does. The existing suite still passes.