fix: deflake zsh-fork approval test (#25669)

Fixes this flake:
https://github.com/openai/codex/actions/runs/26773809591/job/78919970410?pr=25659

This test is about zsh-fork subcommand approval behavior, not workspace
sandboxing, so it now runs with `DangerFullAccess` to avoid macOS
sandbox setup failures before the second subcommand approval.

codex-rs/app-server/tests/suite/v2/turn_start_zsh_fork.rs

@@ -535,12 +535,11 @@ async fn turn_start_shell_zsh_fork_subcommand_decline_marks_parent_declined_v2()
             }],
             cwd: Some(workspace.clone()),
             approval_policy: Some(codex_app_server_protocol::AskForApproval::UnlessTrusted),
-            sandbox_policy: Some(codex_app_server_protocol::SandboxPolicy::WorkspaceWrite {
-                writable_roots: vec![workspace.clone().try_into()?],
-                network_access: false,
-                exclude_tmpdir_env_var: true,
-                exclude_slash_tmp: true,
-            }),
+            // This test is about execve-intercept approval propagation, not
+            // workspace sandboxing. Using full access avoids macOS sandbox
+            // setup failures that can terminate the parent shell before the
+            // second subcommand approval is observed.
+            sandbox_policy: Some(codex_app_server_protocol::SandboxPolicy::DangerFullAccess),
             model: Some("mock-model".to_string()),
             effort: Some(codex_protocol::openai_models::ReasoningEffort::Medium),
             summary: Some(codex_protocol::config_types::ReasoningSummary::Auto),