Reduce ranking spills and repeated maintenance after lease loss

[CountableNewt]

Production ranking stopped publishing during repeated Coordinator lease loss. A matched diagnostic window recorded roughly 1.77 GB of candidate-query temporary writes with only 6.16 MB of new WAL. This change reduces candidate spill and prevents maintenance work from repeatedly blocking publication.

• Materialize compact ranking keys before hydrating selected payloads, preserving the original ordering, moderation, limits and statement snapshot.
• Preserve graph startup grace/backoff across lease-host replacement. Keep its six-hour successful cadence.
• Move disposable mention/profile/metadata expiration out of publication into independently paced 500-row batches, preserving active fetches and required live-item metadata.
• Attribute lease failures using allowlisted service names, query IDs and bounded transaction/query ages; no SQL text or parameters in telemetry.
• Retain the earlier incremental rollup experiment with both database tracking and runtime activation disabled. Its forced-collision benchmark still fails the CPU gate; this PR does not activate it or claim its savings.

Validation: 342 Wire tests with PostgreSQL integration, 152 OperationsCore tests serially, 11 IndexingWorker tests, and two independent-cleanup runtime tests passed. Candidate legacy-query parity covers 32 baseline/external/language/limit combinations. At 100,000 synthetic rows and unchanged 4 MB work_mem, temporary writes fell from about 479 MB to 18 MB and runtime fell about 45%; buffer lookups increased due to selected-key hydration. Production performance remains a rollout gate. Full parallel Operations tests exposed existing shared-fixture lock contention; serial execution passed.

Rollout: Development first, required CI and service/migration checks, then Production. Keep the 16 GB cap, durability, backup policy and replica counts unchanged while comparing publication freshness, actionable queue age, p95 latency, temporary writes and I/O pressure. The bundled experimental migration installs disabled triggers; verify they remain disabled. Worker/Redis aggregation work is a separate follow-up.

Refs TSW-92, TSW-114.

[linear-code]

TSW-114

TSW-92