Dashboard: unified degraded summary banner (lists affected endpoints + Retry)

[chaizhenhua]

Context

PR #222 added per-query 5xx soft-degrade on the dashboard: when one of providers/models/agents/MCP list endpoints returns a 5xx, the dashboard renders ? for the affected count and a "degraded" pill, instead of blanking the page. WorkloadCard distinguishes route_absent (404) vs store_unavailable (503). The reviewer noted that the dashboard still lacks a single place that summarises "what is degraded right now" — it's spread across small pills next to each value.

Problem

In practice an operator wants:

• a top-of-page banner when any dashboard subquery is degraded
• the list of which endpoints are affected (/v1/agents, /v1/providers/list, etc.)
• the last-attempted-at time
• a single Retry-all button

Today they have to scan every card looking for the small ? indicator.

Proposal

1. Extract a useDashboardHealth hook that aggregates the degraded state of every subquery used by the dashboard (config-list providers / models / agents / MCP; runtime-stats; run-counts).
2. Add a <DashboardDegradedBanner> component at the top of the dashboard that renders only when one or more subqueries are in a degraded / error state. It lists affected endpoints by name, the last error message (or HTTP status), the last fetch time, and a Retry button that invalidates the affected queries.
3. Keep the per-card pills as the in-context indicator — the banner is the summary.

Acceptance

• Single banner aggregates degrade state
• Each row in the banner names the endpoint + status + last-fetched timestamp
• "Retry" invalidates affected queries; on success the banner row disappears
• A11y: banner has role="status"; transitions from clean → degraded fire an aria-live="polite" announcement
• Test: induce 503 on two subqueries → banner shows both rows + retry recovers

Notes

This is the response to the reviewer's "排障效率仍一般" concern. Doesn't change the per-card behaviour; just adds the page-level summary.