perf(ui): stop full trace re-download on complete & window-focus

[XinweiHe]

The trace detail view re-downloads the entire trace twice unnecessarily: once when the live stream signals completion (invalidateQueries), and again on window-focus (React Query default staleTime/refetchOnWindowFocus). For a large trace this means repeatedly pulling tens of MB — the user-visible "slow to refresh." A completed trace is immutable, and final spans already arrive via the live stream merge.

Tasks

• In frontend/ui/src/features/traces/hooks/use-trace-stream.ts, remove the invalidateQueries on trace_complete; rely on the merged streamed spans.
• For the trace-detail query (features/traces/hooks/index.ts / TraceViewerPanel.tsx), set refetchOnWindowFocus: false and a long/Infinity staleTime.
• Keep live-streaming merge intact for in-progress traces.

Acceptance

• Opening a completed trace and switching tabs back triggers no new GET …/traces/{id} request.
• In-progress traces still stream and complete correctly.

Files: frontend/ui/src/features/traces/hooks/use-trace-stream.ts, frontend/ui/src/features/traces/hooks/index.ts, TraceViewerPanel.tsx, cross-check app/providers.tsx

Part of #1040.

[Adarsh-Raj-Jaiswal]

Hi @XinweiHe , I would like to work on this one kindly assign it to me.

[XinweiHe]

Deprioritized / parked — superseded by #1041 (gzip) + the P2 lazy-load work (#1046).

Why: The unique gain here — avoiding a full trace re-download on trace_complete and on window-focus — is largely redundant once gzip cuts that download ~8× (#1041, interim) and P2 makes the trace-detail payload sub-MB structurally (#1046). So this change attacks a problem two other items already cover from better angles.

Risk that tips the balance: the invalidateQueries(["trace", …]) on trace_complete is a safety net — its original comment was "refetch to get the final consistent state from ClickHouse." With the known live-completion root-ends-first subtlety (streaming completion can fire before all spans land), the merged SSE stream may not always equal final state. Removing the net to save an already-getting-cheaper download is a poor trade.

Resolution: once P2 makes the payload small, the safety-net refetch becomes cheap, so we most likely just keep it rather than remove it. Revisit this issue only if/when we fix the live-completion signal. Parking for now — removed from Wave 1.