Upstreams: mark check-up findings on the upstream's row, drop the Check-up tab

[fylorn]

The Check-up tab sat beside three configuration tabs (upstreams, proxies, price sheets) it had nothing in common with, and half of it repeated figures shown elsewhere: requests and failures in the 24-hour cell, TTFT and speed in their own column and on the overview. The part only it had was easy to miss behind a fourth tab.

Now

• The tab and its range picker are gone (CheckupTab.tsx, Checkup.i18n.ts).
• Three findings from the last 7 days appear as amber notes next to the upstream's name, after the existing problems (disabled, credential rejected, sign in again, circuit open): Model name differs, Input reported high/low, Low cache reads. Hovering shows the evidence and sample sizes; when several notes stack, each gets a heading, since the row only has room for the first.
• The flagging rules in checkup.ts are unchanged; discrepancies() collects what to mark. The failure-rate flag is removed because the 24-hour cell already covers failures.
• The findings don't count toward "needs attention" in the page header, since forwarding is unaffected. A disabled upstream shows only "Disabled".
• GET /upstreams/health is called with no window (core's default is the last 7 days) and refetched after requests finish, at most every 30 s.
• README (en/zh) bullet rewritten to match.

Checked: pnpm typecheck, pnpm test (835), pnpm build. Also checked in the screenshot harness with injected findings, zh and en: labels on the right rows, tooltip contents and headings.

The site's Lite docs (features.md, keep-api-keys-from-relays.md) still describe the tab; they change with the next release.

🤖 Generated with Claude Code