Avoid duplicate codex_apps tool refresh on force refetch

[mzeng-openai]

Why

app/list(force_refetch = true) creates a fresh temporary McpConnectionManager so it can read current codex_apps tools. That fresh manager already performs an uncached tools/list during startup and writes the result into the Codex Apps tools cache.

The old force-refetch branch then immediately called hard_refresh_codex_apps_tools_cache() on that same fresh manager, which issued a second uncached tools/list against the same server connection before any caller could observe the first result. That duplicated a live inventory request without making the data any fresher.

This is separate from the repeated in-process inventory reads that can happen during normal codex exec startup. Those reads reuse the already-loaded tool snapshot; the redundant network call was specific to the fresh-manager force-refetch flow.

Relevant code:

• list_accessible_connectors_from_mcp_tools_with_environment_manager

What changed

• For force_refetch, wait for the freshly-created temporary manager's codex_apps startup to complete, then reuse list_all_tools() from that manager instead of forcing a second live refresh.
• Keep the existing fallback behavior when codex_apps is not ready: continue with the startup/cached tools path rather than changing visible behavior.
• Extend the existing app-list integration coverage to count tools/list calls and assert:
  • the warm request performs one inventory fetch
  • the later force-refetch request adds exactly one more fetch, not two

Regression coverage:

• list_apps_force_refetch_patches_updates_from_cached_snapshots

Verification

• cargo test -p codex-app-server list_apps_force_refetch_patches_updates_from_cached_snapshots
• cargo test -p codex-core connectors
• just fix -p codex-core