#2418 serve: after lazy hydration, scoped poller pulls never take the fast path, and the login mint waits behind them on the sync gate

Description

On a serve whose datastore cache was hydrated lazily, every poller pull takes the slow path indefinitely. The OAuth login mint (and every gated WebSocket mutation) waits behind those pulls on the sync gate.

1. A lazy hydration pull (metadataOnly) sets lazyPullActive in the S3/GCS extension's sync state (gcs_cache_sync.ts:2507-2511).
2. While lazyPullActive is set, every pull that is not metadataOnly skips the commitSeq fast path (gcs_cache_sync.ts:2178).
3. Only an unscoped pull with no model context clears the flag (gcs_cache_sync.ts:2512-2519).
4. All three serve pollers (config, access-data, runtime-data) pass subdirs, so they are scoped and never clear the flag. Each poll every 30 s therefore does a full slow-path pull: assembling the index from every shard, then listing the subdir prefix. The runtime poller lists all of data/.
5. Each poller pull holds the one-permit FIFO sync gate (src/serve/sync_gate.ts), for up to POLLER_PULL_TIMEOUT_MS (120 s). The login mint (device_auth_handler.ts, withSyncGate) and the gated WebSocket mutations queue behind up to three such holds. They go ahead ungated only after GATE_WAIT_TIMEOUT_MS (150 s).

The result: on a large lazily hydrated datastore, login and mutation latency can be dominated by the gate wait even once the mint's own push is scoped (swamp-club#2408).

Where this applies

• ops.swamp-club.com hydrates lazily through the platform-swamp serve-lazy-hydration.patch. It currently runs core 375d7eca, which predates the sync gate. So this becomes the main login delay once platform-swamp moves its core pin to current main while keeping that patch.
• Current main's serve does not hydrate lazily at boot. But a cache hydrated lazily by the CLI (hydrationStrategy: lazy) and later served by swamp serve keeps lazyPullActive in its sync-state sidecar, so the same behaviour follows.

Possible directions

• Extension: let a scoped pull that covers the subdirs it was asked for use the fast path when the remote commitSeq has not moved, or track lazy hydration per subdir.
• Serve: bound the mint's gate wait separately, or shorten poller holds, for example by skipping a poll cycle while a gated request is queued.

Evidence

• Code references are from swamp-extensions at c47093989 (datastore/gcs/extensions/datastores/_lib/gcs_cache_sync.ts). The S3 extension has the same !scoped guard.
• The swamp-club#2408 MinIO repro did not hydrate lazily. Its pollers hit the fast path (pullChanged.fastpath ~5ms hit), so gate contention was not observed there. Reproducing this issue needs a lazily hydrated serve cache.