[beinan]

Dedicated catch-up processes can time out claiming a free table because startup and each claim sweep every running task. This change makes dedicated startup, admission and completion target-local; ordinary recovery batches healthy-claim checks and reads running tasks in bounded pages. Blocked queue candidates are rejected before allocating leases, while final ownership comparisons remain authoritative.

Start lease renewal before publishing a claim, verify our unique token after a lost transaction response, and let an already admitted claim attempt finish across the retry deadline. This prevents a successful claim from being discarded by the outer timeout. Storage fencing, merge concurrency and retry budgets are unchanged.

Validation:

• Workspace test build, master all-target check, formatting and diff checks pass.
• Master tests: 63 ordinary tests and all 67 etcd-backed HA tests pass on native local etcd (4 test threads). An earlier x86-emulated etcd run suffered multi-second RPC latency and 11 timeout/lease-sensitive failures; the complete suite passed after switching to the host architecture.
• Isolated staging etcd: target isolation, live/expired ownership, paginated recovery and unique-token lost-response tests pass.
• Three staging samples with 80 healthy running tasks: old sequential recovery scan 6.31–6.67 s (median 6.44 s), batched recovery 0.43–0.46 s (median 0.43 s), complete exact-target claim 0.58–0.60 s (median 0.59 s). These are admission/control-plane timings, not end-to-end merge throughput; the old scan excludes the additional work of claiming.
• Self-review confirmed final claim/recovery CAS predicates remain authoritative; a stale preflight cannot grant ownership or requeue a newly claimed task. CI must pass before merge; production acceptance also requires subsequent WAL publication.