[alexeygrigorev]

• tokens.py: ensure_campaign_recipient_tokens(locked=True) skips the second SELECT FOR UPDATE when the sender already holds the row lock. Queries/email (sqlite marginal, test-measured): 10 -> 9 (Postgres +1 FOR SHARE in both).
• campaign_sender.py: persistent sender thread pool (Django connections are per-thread) so DB connections survive across batches; close_old_connections hygiene at drain end. Campaign is loaded once per batch (already from Cut per-message send CPU ~60% (GIL-bound worker: 30 -> 45+ msg/s) #73; asserted in test).
• Batch log line now carries per-phase sum/p50/p95 (lock_sql, render_mime, limiter_wait, ses_call, commit), emails, and worker CPU (process_time); no PII.
• ses.py: thread-local build/limiter timings; conftest: close pool before test DB drop.
• SES call stays inside the transaction (out of scope).