fix(eve): resolve MCP connection url from re-imported module at runtime

[amotarao]

Closes #521.

Summary

defineMcpClientConnection connections now honor the URL produced by the re-imported authored module at runtime instead of the manifest snapshot captured at build time. Env-driven endpoints — most notably Vercel Services bindings, whose target host is only injected at request time — resolve to the correct URL on cold start.

• resolveConnectionDefinition prefers resolvedRecord.url when the re-imported module exports a non-empty string. When it omits or blanks url, it falls back to the compiled manifest value, so existing static-URL connections are unaffected.
• The connection:<url> secondary source key now uses the runtime URL, matching the key stamped by defineMcpClientConnection on the same re-imported module.

Authoring stays unchanged — the current shape is enough:

const base = process.env.BACKLOG_MCP_BASE ?? "http://backlog-mcp.invalid";
export default defineMcpClientConnection({
  url: `${base}/mcp`,
  description: "...",
});

At build the placeholder URL is baked into the manifest; at runtime the module is re-imported with the real binding env in place and its computed URL wins. No new API surface, no new marker on the compiled manifest.

I intentionally did not add the url: (ctx) => … resolver form proposed in the issue — the reproducer (Vercel Services bindings) is a module-scoped env problem, not a per-caller one, and this fix closes it without expanding the API. Happy to layer the function form on top separately if a per-caller URL use case shows up.

Test plan

• pnpm typecheck
• pnpm oxlint --fix / pnpm oxfmt
• pnpm guard:invariants
• pnpm test:unit (4381 passed)
• Added two resolve-connection.test.ts cases: runtime URL preferred, manifest fallback when the module blanks url.

[vercel]

@amotarao is attempting to deploy a commit to the Vercel Team on Vercel.

A member of the Team first needs to authorize it.