`Expand` fails to work recursively with some combinations

[ranweiler]

The Expand evaluator canonicalizes Path-typed values eagerly, since #22. Canonicalization is an I/O operation, and fails if the resulting path does not exist on-disk. As a result, if a Path-valued string is expanded to a value that containers another Expand placeholder (expandable variable), the expansion will fail.

This is exhibited by an important use case that currently fails:

#[test]
fn test_setup_dir_and_target_exe() -> Result<()> {
    let expand = Expand::new()
        .setup_dir("/my/setup")
        .target_exe("{setup_dir}/my-fuzzer.exe");

    let target_exe = expand.evaluate_value("{target_exe}")?;
    assert_eq!(target_exe, "/my/setup/my-fuzzer.exe.);

    Ok(())
}

In the above, the internal expansions should be:

1. {target_exe}
2. {setup_dir}/my-fuzzer.exe
3. /my/setup/my-fuzzer.exe

But, since the intermediate state of (2) is an ExpandedValue::Path and does not exist on disk due to the unexpanded path segment, the call to canonicalize() in Expand::replace_value() fails.

The fix to this must support the above test case, and maybe reasonably abandon canonicalization (or at least have a mode that permits canonicalization failures). It must also avoid expansion cycles that would lead to stack overflows or infinite loops.

AB#41312795

[Porges]

@ranweiler Is there any reason to canonicalize during expansion and not once at the end?

[ranweiler]

@ranweiler Is there any reason to canonicalize during expansion and not once at the end?

I think that would work fine (though I'm not sure we really should continue to require that the final paths exist at expansion time).

The trick is just that canonicalize() is currently called for the value of placeholder variables assignments of type ExpandValue::Path. But its caller, evaluate_value(), doesn't know if the ultimately-returned String is itself actually a Path. It truly isn't, in general. It might be an env var assignment, or a UUID, &c. So we could just change the API and require that callers canonicalize as appropriate, rather than promising to canonicalize values. We could perhaps add a helper expand_path() that calls expand_value() and then canonicalizes for callers.