Refresh Subsystem

Choose the changed contract first; do not load every refresh document or reference for a narrow task.

Route context

Read fragility only when the task touches a named failure-prone path. Use its headings to select the relevant section rather than loading it for ordinary DTO or local snapshot projection work.

Core contracts

• backend.RefreshCoordinator owns refresh/catalog lifecycles, aggregate routing, telemetry, governor/spill state, and the shared global container-log limiter. Refresh may read ClusterRuntimeManager and invalidate ResourceGateway; neither dependency may call back into Refresh.
• Per-cluster initialization order is informer factory and permission checker, permission preflight, ordered domain registration, snapshots/queues/streams, manager start, then revalidation. Publish aggregate request/response and stream routing only after their owning aggregates are ready. Wails owns the /api/v2 service route and named-stream registration; do not add an application-owned loopback server.
• buildRefreshSubsystemForSelection is the construction chokepoint for startup, selector-open, recovery, and governor re-warm. Per-cluster readiness, invalidation, and lifecycle wiring belong there.
• Refresh scopes target one cluster. Aggregate display fans out above refresh state; handlers route results rather than merging clusters.
• The shared refresh-domain contract owns cross-layer metadata. Generated types are regenerated, never hand-edited.
• Snapshot and stream rows share projection helpers and parity tests. Stream identity crosses the wire in the top-level full ref.
• New table/list domains require change-signal coverage. Polling is a stream-down fallback unless a documented conditional producer requires poll augmentation.
• Snapshot cache keys, invalidation, source clocks, signal clocks, and query revisions are one ordering contract; trace producer through rendered consumer before changing any of them.
• Metrics-interval and global container-log-limit settings arrive through write-only sinks after the Preferences lock is released. Never read Preferences from refresh code or acquire a subsystem lock while holding the limiter's leaf lock.

Workflow

1. Identify domain, scope, producer, signal source, cache owner, and every frontend consumer.
2. Prove registration/permission ordering and both sides of any readiness gate.
3. Write the failing regression test at the contract boundary.
4. Change backend and frontend mappings together when the shared contract moves.
5. Exercise restricted RBAC and multiple connected clusters when affected.
6. Check diagnostics, teardown, and fallback behavior before the root final validation gate.

Focused checks

Choose packages/specs matching the change:

mise exec -- go test ./backend/refresh/snapshot ./backend/refresh/system
mise exec -- go test ./backend/refresh/resourcestream/...
mise exec -- npm run test --prefix frontend -- refresh streaming
mise exec -- npm run typecheck --prefix frontend

For runtime wedges, capture a goroutine dump before modifying synchronization; see docs/workflows/goroutine-dump.md. Verify payload-shape claims against the actual snapshot endpoint when the local app is reachable.