design(server): avoid title-worker starvation while sessions are active

Problem

Automatic session-title generation uses a two-worker coordinator. Each worker currently owns the full lifecycle: claim, provider generation, and title-result persistence.

The immediate correctness fix in #1308 prevents a title snapshot save from overwriting a live chat's newer conversation by waiting until the session has no active run. That wait is safe for history, but a title result can then occupy a coordinator worker for the full duration of a later chat run.

Two long-running sessions can therefore consume both workers while they wait to persist titles. Other sessions' title work queues behind them; once the bounded queue fills, title submissions are dropped as best effort. Chat execution remains correct, but server-wide title generation throughput becomes coupled to unrelated chat duration.

Desired outcome

Design title generation so bounded provider-generation capacity is not held while waiting for an active session to become safe to update. Preserve the no-lost-update guarantee for conversation history, durable title-attempt accounting, bounded resource use, and clean shutdown/restart behavior.

Design directions to evaluate

• A durable title-only metadata record/projection separated from the conversation aggregate.
• Aggregate revision / compare-and-swap semantics with safe retry/merge for every session writer.
• A bounded, durable deferred-commit mechanism driven by session-settlement notifications.

Do not solve this by spawning an unbounded goroutine per blocked title commit; that only relocates the resource-exhaustion and restart-recovery problem.

Context

• Immediate correctness repair: #1308
• Title coordinator: internal/adapter/server/title_coordinator.go
• Terminal persistence/title submission: internal/adapter/server/service.go (persistRun)