fix: scope connection health refreshes

[saga-agent]

Issue

When a user opens an integration with multiple saved connections and clicks Check now on one connection, the UI can briefly make the other connections look like they received the same health-check result.

Example user-visible behavior:

• click Check now on Account A
• Account A correctly updates to Healthy or Expired
• unrelated rows, like Account B or Account C, can also appear to update to that same status in the live UI
• after a full refresh, only Account A shows the new persisted health verdict, which makes the in-page state look wrong or stale

The server-side saved verdict is scoped to the checked connection. The confusing part is the client-side refresh/reconciliation path after the manual check.

Fix

• remove broad connections mutation reactivity from the manual per-row Check now path
• apply the returned verdict directly to the clicked row's local probe state
• explicitly refresh both connection read surfaces afterward:
  • the owner-scoped connection list used by account rows and summaries
  • the all-connections list used by provider-account grouping

This keeps sibling rows from sharing the clicked row's mutation result while still preventing stale reads after the persisted last_health changes.

Test plan

• bun run --cwd packages/react typecheck
• bun run format:check
• E2E_SELFHOST_URL=http://100.106.134.13:45774 ../node_modules/.bin/vitest run --project selfhost scenarios/health-checks-ui.test.ts

Notes

This PR is intentionally opened in the fork against review/pr-1311-base, not against upstream. Do not upstream yet.