Show a loading Model picker on Video Gen and cache the model list

[atomantic]

Summary

The Video Gen Model select is populated from GET /api/video-gen/status, which shells out to python and rebuilds the hardware-aware model list on every call. The field was hidden until that landed, so it popped into the middle of the form a second or two after the page rendered.

• The field holds its place. ModelSelect gains a loading prop — the same <select>, with its options swapped for a disabled Loading models… placeholder and aria-busy set. The Model label stays put, so nothing below it jumps when the list arrives.
• The list is cached per session. client/src/lib/videoGenStatusCache.js stores the model-shaping slice of the payload (models, defaultModel, systemMemoryGb) in sessionStorage, so a revisit paints the real picker immediately while the probe revalidates behind it.
• The cache can't report python health. Only the whitelisted fields are stored — connected, reason, missingPackages, pythonPath, byovRuntimes, runtime are dropped rather than guarded — and the read is marked stale: true. The connectivity pill and the "install missing packages" banner wait for the live probe, so an interpreter the user just fixed (or broke) is never described from storage.

Session scope, not local: the model registry and the python environment both move with an upgrade or an install.

Test plan

• client/src/pages/VideoGen.modelLoading.test.jsx (new, 3 cases): the field renders disabled with the placeholder while /status is in flight and becomes the real picker when it resolves; a second mount paints the cached list before the probe returns, and only the three model-shaping keys are persisted; a cached entry carrying a failed probe still shows Checking… with no install banner until the live probe lands.
• client/src/pages/VideoGen* + client/src/components + client/src/lib/index.test.js: 448 files / 4962 tests pass.
• biome lint --error-on-warnings clean on the changed files.
• Four pre-existing suite failures on this repo (a11yConventions, responsiveGridConventions, FableLoomStory.browser, ChiefOfStaff) reproduce identically on the base commit and are unrelated.

Follow-ups filed

• Serve the Video Gen model list off /video-gen/models instead of the python-probe /status payload #5835 — serve the model list off /video-gen/models (no python probe) so even a cold load doesn't wait, which would let the cache be deleted.
• Extract the shared VideoGen page test harness (5 near-identical mock scaffolds) #5836 — extract the shared VideoGen page test harness (now 5 near-identical copies).