fix(dashboard): pin --vv-offset-top to resize/focus, ignore pan scrolls

iOS fires visualViewport scroll events at 60fps during a pan with the
keyboard up. Routing those through React state and into the .chat-thread
translateY(--vv-offset-top) transform amplified the pan into a visible
judder + ~300px shift + body background exposure.

useMobileKeyboard now uses two listeners: a full update (resize +
focusin/focusout) that re-snapshots all metrics including offsetTop, and
a scroll-only update that updates only height/keyboardOpen. offsetTop
is therefore frozen between keyboard open/close events — the transform
correctly compensates for iOS's initial visualViewport shift on focus
without following pan-time movement.

Restores the translateY anchor (so the thread isn't off-screen on
first focus) while keeping the swipe-jitter fix.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

docs/architecture.md

@@ -471,6 +471,12 @@ Implemented in `agent-heartbeat.ts`:
 - Canonical replication/write-coordination contract: [`docs/shared-mesh-protocol.md`](./shared-mesh-protocol.md)
   - Defines protocol versioning, write classes, quorum/ack semantics, lease epochs/fencing, offline queue/replay, reconciliation outcomes, restart recovery hooks, and degraded-read staleness metadata.
   - Existing `/api/mesh/sync` and settings-sync payloads remain the active exchange primitives while follow-on runtime tasks implement full v1 coordinator/quorum behavior.
+- Distributed task-ID allocation (`packages/core/src/distributed-task-id.ts`) is the first mesh-aware coordinated write primitive.
+  - Durable state lives in SQLite tables `distributed_task_id_state` (prefix sequence + authoritative committed count) and `distributed_task_id_reservations` (reservation lifecycle rows).
+  - Reserve/commit/abort execute under a process-local lock and a single SQLite transaction. Lazy reservation expiry cleanup runs inside those same transactions.
+  - Default reservation TTL is `15 * 60 * 1000` ms (15 minutes). Expired/aborted reservations are **burned IDs** and are never reissued.
+  - `committedClusterTaskCount` from allocator state is the only authoritative cluster-wide committed-task count. Local task-row counts and ID suffix math are not authoritative.
+  - Mesh allocator write routes (`/api/mesh/task-ids/reserve|commit|abort`) return `503` when the coordinator node is unreachable; they never fall back to local-only cluster ID issuance.
 - Process lifecycle ownership:
   - `fn serve` / `fn dashboard` start a single process-level `PeerExchangeService` and stop it during shutdown.
   - `CentralCore.startDiscovery()` is invoked from CLI startup only after HTTP bind completes so discovery advertises the actual listening port.

docs/multi-project.md

@@ -35,6 +35,7 @@ Peer/mesh coordination spans core + engine, with startup ownership in CLI proces
 - `NodeDiscovery` and `NodeConnection` in `@fusion/core` handle discovery and remote node connectivity/auth primitives.
 - `PeerExchangeService` in `@fusion/engine` coordinates node-to-node sync/exchange workflows.
 - Canonical replication semantics live in [`docs/shared-mesh-protocol.md`](./shared-mesh-protocol.md). That protocol separates strongly coordinated shared state from append-only streams, queued replay classes, and node-local runtime state.
+- Distributed task-ID allocation is one strongly coordinated shared-state path: reserve/commit/abort are coordinator-mediated writes, and cluster-wide committed task totals come from allocator `committedClusterTaskCount` state (not per-node local task counts).
 - `runServe()` and `runDashboard()` (CLI) own process-level mesh service lifecycle:
   - start one process-wide `PeerExchangeService` instance
   - call `CentralCore.startDiscovery()` only after the HTTP server is listening and the real bound port is known