Manually execute `ng new` deps postinstall scripts

首先检查会自动运行 npm install 的 ng new 流程，并确定当前如何处理 postinstall 脚本。完成标准是禁用广泛的自动 postinstall 执行，仅调用所需的已知包，并添加用于检测缺失 postinstall 依赖的覆盖测试。

🚀 Feature request

Command (mark with an x)

• new

Description

Currently, ng new will automatically run npm install which (if the user has not disabled it), automatically runs all postinstall scripts. This can be a vulnerability since any compromised package in the NPM dependency graph could add a postinstall step to install malware on developer machines.

Describe the solution you'd like

We could reduce the attack surface by disabling postinstall on the automatic npm install and then manually invoke the postinstall for a known set of required packages. Only 3 packages currently use postinstall steps, so limiting execution to just those would significantly reduce the attack surface for a potential supply chain attack.

One possible concern is for dependencies which add a required postinstall step in the future. We can pretty easily add a test to make sure we aren't missing any postinstall steps from our transitive dependencies, though this inherently breaks abstraction somewhat. Adding a postinstall step is (somewhat debate-ably) a breaking change, so any package which adds one in the future should require a major version bump where we have an opportunity to allowlist it.

The one edge case I can think of is if we have:

ng-new-app@0.0.0 -> package-a@^1.0.0 -> package-b@^1.0.0

And package-b gets a new postinstall step in v2.0.0. However, package-a may be able to manage the breakage without violating their own public API (or maybe doesn't notice the new postinstall step) and simply bumps to v1.0.1. This would immediately be pulled in to the next ng new command and fail. I think such a scenario would actually be a bad patch release for package-a, since adding a required postinstall step is fundamentally a breaking change. We would rely on NPM package maintainers to make the right semver-compatible decisions for a somewhat nuanced case, but this is probably better than the alternative.