#3308 swamp serve lists every wait-key record on each continuation pass once any pending signal has lapsed

Problem

Follow-up to swamp-club#3212, raised by the adversarial review of PR 3021. That PR made swamp serve run the pending-signal sweep with every continuation pass (sweepHeldSignals in src/serve/continuation_sweep_service.ts, default every 30 seconds), so pending records are cleaned up where they gather.

sweepPendingSignals in src/domain/workflows/signal_wait_cleanup.ts returns early only when there are no pending addresses at all. A lapse is never deleted: by design one stays per slot until the slot is claimed. So once a key has had one early signal lapse and its next run does not come (a retired workflow, or one that runs monthly), every pass from then on:

• lists the whole wait-pending/ prefix,
• calls listKeyRecords() over the whole store, which is one read per key record of every workflow,
• reads highestKeyRecord and highestPendingRecord for each slot.

Nothing is wrong with the result. On a control-plane store backed by S3 or GCS it is a steady stream of requests that never stops and grows with the number of keyed workflows. findTakerOf in src/domain/workflows/early_signal.ts has the same shape (it lists every key record of every workflow), but only runs on a rare arrival path.

Suggested fix

• Skip a slot whose highest record is already a lapse with nothing below it before listKeyRecords() is called.
• Build the set of claims that read a slot lazily, only when a slot that is not live turns up.
• Or run this part of the sweep on a longer cadence than the continuation pass.
• Consider a per-key listing for findTakerOf instead of listKeyRecords().

Assert on reads done, not on time: a test with only lapsed slots should make no call to listKeyRecords.