Should we test for the sympton rather than the specific detail? In this case, its an extra rebuild

Why is only comparing first() sufficient?

The reason this approach didn't work is pretty much due to #10358. When Cargo computes fingerprint for a Unit of work, it asks env::var which doesn't contain LIBRARY_PATH or PATH because that is set by Cargo. rerun-if-env-changed only captures environment variables set by users via shell like MY_VAR=1 cargo build. Variables added by Cargo only available when consulting the Command struct of the build script execution via Command::get_envs (like what has done in f91dc78).

As a consequence, if you open the fingerprint JSON file of the build script execution (normally at target/debug/.fingerprint/foo-559e30894cda9925/run-build-script-build-script-build.json), you'll see

{
  "rustc": 11851488446658826000,
  // ...
  "local": [
    {
      "RerunIfEnvChanged": {
        "var": "DYLD_FALLBACK_LIBRARY_PATH",
        "val": null
      }
    }
  ],
  "rustflags": [],
  // ...
}

which shows that Cargo failed to capture the initial library search path set by itself.

There is another way to observe this bug: set the LEVEL env var in the test to a higher number like 5. Any subsequent Cargo invocation after the first one should be fresh. Before doing this verification you need to change std::env!("LEVEL") to std::env::var!("LEVEL") because the former will be captured by rustc dep-info, which trigger a rebuild.

Before doing this verification you need to change std::env!("LEVEL") to std::env::var!("LEVEL") because the former will be captured by rustc dep-info, which trigger a rebuild.

I didn't expect such a difference.

I decide to to get the "LEVEL" from the command line argument.