app-server: add turn-scoped MCP metadata by server

[mzeng-openai]

Why

App-server clients need a way to pass turn-scoped metadata to model-initiated tool calls for a particular MCP server, such as request location metadata for search, without exposing the broader responsesapiClientMetadata bag to every MCP server used during the turn.

What changed

• Add experimental turn/start.mcpMetaByServer, keyed by configured MCP server name, and carry it through turn startup and active-turn steering.
• Merge each configured metadata object into outbound _meta only for model-initiated calls routed through the matching server. An entry for an aggregate server such as codex_apps therefore applies to connector tools routed through that server.
• Keep responsesapiClientMetadata scoped to Responses API requests rather than forwarding it to MCP requests.
• Protect Codex-owned MCP request metadata keys, including turn metadata, app/plugin/thread/sandbox fields, and rollout trace correlation ids, from client override.
• Document the experimental API and add protocol, core unit, and model-triggered MCP integration coverage.

Verification

• cargo test -p codex-protocol
• cargo test -p codex-app-server-protocol
• cargo test -p codex-rollout-trace
• cargo test -p codex-core --lib mcp_tool_call_request_meta
• cargo test -p codex-core --lib turn_metadata_state_merges_client_metadata_without_replacing_reserved_fields
• cargo test -p codex-core --lib steer_input_returns_active_turn_id
• cargo test -p codex-core mcp_tool_call_receives_turn_metadata_for_selected_server
• cargo test -p codex-app-server --test all turn_start_updates_sandbox_and_cwd_between_turns_v2
• cargo check -p codex-mcp-server -p codex-thread-manager-sample
• Generated stable and experimental app-server schema fixtures with just write-app-server-schema and just write-app-server-schema --experimental; stable fixtures remain unchanged because the new field is experimental.

Known issue

• cargo test -p codex-core aborts in thread_manager::tests::resume_and_fork_do_not_restore_thread_environments_from_rollout with a stack overflow; the same test reproduces when run in isolation, while the feature-specific core tests above pass.