fix(deno_facade): allow non-MainModule resolution when referrer is "."

[0x5457]

What kind of change does this PR introduce?

Bug fix

What is the current behavior?

EmbeddedModuleLoader::resolve rejects any resolution where referrer == "."
unless kind == ResolutionKind::MainModule:

if kind != ResolutionKind::MainModule {
    return Err(generic_error(format!(
        "Expected to resolve main module, got {:?} instead.",
        kind
    )));
}

This breaks Deno's CJS→ESM bridge. When a CommonJS module synchronously
require()s a pure-ESM module, loadESMFromCJS re-enters resolve with
ResolutionKind::Import and the same "." referrer, and the worker boots
with:

Error: Expected to resolve main module, got Import instead.
    at loadESMFromCJS (node:module:777)

fix #703

What is the new behavior?

After the fix, require("./message.mjs") and require("chalk") (a
"type": "module" package) both load correctly from a CJS entrypoint.

[nyannyacha]

Hello @0x5457 😋
Thank you for submitting this.

Have you tried running deno compile in Deno v2.1.4?
The module loader implementation in the edge-runtime codebase reflects the EmbeddedModuleLoader implementation in the deno compile subcommand path of Deno v2.1.4 codebase, so it appears to behave consistently; however, deno run seems to work without error, which suggests an inconsistency.

I think this change is acceptable, but do you have any special comments regarding the above?

[0x5457]

Hi @nyannyacha, thanks for the review!

I haven't tried deno compile on v2.1.4 directly, but I noticed that upstream Deno hit the same issue and fixed it in denoland/deno#28379. That PR removes the exact same if kind != ResolutionKind::MainModule { ... } guard from EmbeddedModuleLoader::resolve in cli/rt/run.rs, its description explicitly mentions "fixes a bug where require esm wasn't working in deno compile", which matches what we're seeing here.

https://github.com/denoland/deno/pull/28379/changes#diff-c30a4785f54c02866be5d07ef0ce993d0518eeea81a85344ed953de6bbcf852cL156-L167

[nyannyacha]

Some tests seem a bit flaky due to third-party sources, but as the fix itself appears to be fine, so LTGM.
Thank you for your efforts! 😄