fix(ssr): class property keys hoisting matching imports

[SunsetFi]

In an SSR environment, vite will incorrectly hoist imports whose names match class property keys.

This was first identified in vitest:
Reproduction: https://github.com/SunsetFi/vitest-class-method-name-collision/tree/main
Issue: vitest-dev/vitest#10093
PR: vitest-dev/vitest#10094

This PR was made on the request of @sheremet-va, over in the vitest PR for this issue.

I do not have a reproduction at hand, but the class props test in ssrTransform.spec.ts demonstrate the bug occurring:

    const __vite_ssr_import_0__ = await __vite_ssr_import__("vue", {"importedNames":["remove","add"]});


    // Incorrect: Should not be hoisted, as the class identifiers are keys, not references
    const add = __vite_ssr_import_0__.add;
    const remove = __vite_ssr_import_0__.remove;
    class A {
      remove = 1
      add = null
    }

This PR provides 2 code changes, both related:

• Fix isRefIdentifier incorrectly indicating non-computed class property keys are references
• Roll back the fix from Bad code transformation for class fields if an import has the same name vitest-dev/vitest#303

Based on our research, the initial reason for the identifier references getting hoisted was introduced due to #2221. In order to reference a variable in the 'extends' directive of a class, the variable must not be a property path:

// Wrong
class MyClass extends __vite_ssr_import_0__.MySuper {}

The fix in #2221 was to extract it to its own variable:

// The intent behind the fix
const MySuper = __vite_ssr_import_0__.MySuper;
class MyClass extends MySuper {}

However, at some point, it was noted that vitest was mangling class property keys as if they were imports: vitest-dev/vitest#303.

The fix for this was implemented in #6261. However, the fix itself seems to function only indirectly, and it introduced the bug seen in vitest-dev/vitest#10093. The effect of this fix was to treat all class property key identifiers as if they were super references in class declarations. This consequently triggers the variable hoisting behavior on all matches. The root cause, that class property keys were treated as references to identifiers, was never resolved, and still surfaces through the hoisting mechanism kicking in when it sees an import that coincidentally matches the name of a class property key.

The appropriate fix for vitest-dev/vitest#303 seems to be instead changing isRefIdentifier to correctly specify that non-computed class property key identifiers are not references.

[SunsetFi]

Unfortunately, I am totally unable to put together a reproduction to this. I am not too familiar with SSR in vite, and try as I might, I cannot get ssrTransform to be invoked, either by build or by dev mode.

If anyone knows the conditions that this function runs, I can put together a reproduction.

[sapphi-red]

The change makes sense to me. I think we don't need a reproduction for this 👍