Whole-file loading still failed after adding caching because the public proxy interrupted large GLB response bodies. Download validated 1 MiB ranges, retry transient failures up to three attempts, and persist completed chunks so reloads resume from their last boundary.
Add regression cases observed failing before the change for interrupted downloads, cross-reload resume, and temporary HTTP failures. Verify range integrity, full-cache reuse, permanent errors, and storage fallback.
Persist versioned models and validated network metadata in IndexedDB. Show scene download progress and recoverable errors with the existing frosted material.
Connect existing skeletons to pending navigation and route boundaries while preserving the map. Previously only dynamic imports showed fallbacks, and rejected model promises prevented retries.
Verified: 32 focused unit tests; 4 navigation and scene browser checks; 7 cache browser checks including invalid-metadata recovery; isolated production build and type checks.
The scene previously returned early when realtime rows were absent and selected map-tier assets for network mode, hiding SCADA markers and station interiors and floors.
Use SCADA timestamps as valid frames, render node-aligned monitor markers, and pin the full network context in the manifest regression test. Keep the timeline close action inside its drag header.
Keep the control panel geometry stable when toggling building context, and use accessible switch controls to prevent focus-driven movement. Add browser coverage for camera framing, opacity, responsive layout, and the default 6x pipe scale.
The previous fix only prioritized authStore inside the route guard, so an unauthenticated NextAuth session or a suppressed access-context 401 still collapsed into an empty permission set. Propagate both authentication signals before authorization so expired sessions consistently reauthenticate.
Use project-scoped endpoints and run IDs for current project, SCADA metadata, historical time series, migrated DMA results, and sensor placement runs.
The regressions recurred because tests mocked pre-interceptor and new-only payload shapes, while history queries relied on implicit global scheme state and unbounded request fan-out. Cover post-interceptor pages, migrated result_rows, explicit run_id routing, multi-element requests, request limits, and cancellation.
Runtime model checks previously mapped every authenticated endpoint failure to models_unavailable. Treat request failures as service outages and retry failed checks after five seconds; empty model configurations remain distinct.
Keep the virtual-window measurement wrapper as a flex column so AgentTurn alignSelf rules remain active. Add regression coverage for the alignment context.
jiang
TJWater CI