module: Refine and unify exports resolution error handling

[guybedford]

This PR includes a number of refinements to the error handling of exports resolutions in the ES module resolver:

• Exports resolution error messages and codes are fully unified between the CJS and ESM resolver implementations, with the only difference that the ESM resolver errors include an imported from ... part since they don't have natural error stacks like require() does.
• The MODULE_NOT_FOUND errors from "exports" resolution are split up into three separate errors: ERR_INVALID_PACKAGE_TARGET, ERR_INVALID_MODULE_SPECIFIER, and ERR_PACKAGE_PATH_NOT_EXPORTED. Previously we wanted to try to maintain unification on the MODULE_NOT_FOUND error, but we already have ERR_INVALID_PACKAGE_CONFIG and ERR_INVALID_MODULE_SPECIFIER to track, such that this unification has become restricting I think.
• ERR_INVALID_PACKAGE_TARGET is thrown whenever the exports resolution value in the package.json itself is invalid. This includes values that fail validation like invalid:url or backtrack below the package boundary like ../../../outside.
• ERR_INVALID_MODULE_SPECIFIER is extended to throw for cases like "exports": { "./sub/": "./sub/" } where the user attempts to require eg pkg/sub/../../../asdf or pkg/sub/node_modules/... etc. The error thus distinguishes that the fault is with the requestor, not the package.json itself.
• ERR_PACKAGE_PATH_NOT_EXPORTED is thrown whenever trying to import an export from a package that is not defined. These are effectively the encapsulation errors for "exports".

The main semantic changes in this PR are then the following:

1. Conditional exports fallbacks (as introduced by logical conditional exports ordering in module: implement logical conditional exports ordering #31008) are then restricted only to apply to ERR_PACKAGE_PATH_NOT_EXPORTED errors, and not the other classes of errors. This much more clearly ensures that only those conditions not defined to match fall through. For a contrived example: "exports": { "node": { "node": "not:valid" }, "default": "./valid.js" } would before this PR always return the valid.js path, but after this PR always throw an ERR_INVALID_PACKAGE_TARGET error for the not:valid path. On the other hand, "exports": { "node": { "never": "./never.js" }, "default": "./valid.js" } would fall through the "never" path fine. Array fallbacks continue to fall back on the validation errors though as designed.
2. Whenever "exports" is set, we no longer fall back to trying the main. If there is no main defined by exports - either because there is no "." entry, or because there is no conditional resolution for the "." entry (eg "exports": { "never": "./never.js" }), the ERR_PACKAGE_PATH_NOT_EXPORTED error will immediately throw without doing any "main" or "index.js" lookups. This makes "exports" exhaustive in its definition of the package resolution, and simplifies the error handling consistency for exports in turn.

Since its original proposal, "exports" has very much turned into a full replacement for "main". Having self-resolve entirely reliant on this field being set is also in line with this. I understand (2) may be the most controversial part of this PR with the biggest user-facing consequences, so I expect we should discuss this further in the next meeting. This PR can be adjusted further based on any resolutions. The logic connects around the error handling hence the monolith here.

In addition, this PR fixes #31510 and the incorrect use of XOR instead of bit shift described in #31008 (review).

Checklist

• make -j4 test (UNIX), or vcbuild test (Windows) passes
• tests and/or benchmarks are included
• documentation is changed or added
• commit message follows commit guidelines

[guybedford]

//cc @nodejs/modules-active-members

[guybedford]

Thanks @addaleax for the review here, would appreciate your advice re how to properly do the check handling.

[guybedford]

@addaleax I've added all the checks you mentioned here now. Better docs / info around how exactly to these sorts of cases properly would definitely be useful for future reference.

[nodejs-github-bot]

CI: https://ci.nodejs.org/job/node-test-pull-request/29200/

[nodejs-github-bot]

CI: https://ci.nodejs.org/job/node-test-pull-request/29201/

[guybedford]

Landed in 58de9b4.

[guybedford]

The CITGM failures were tracked down to causes unrelated to this PR in nodejs/citgm#785 (comment).

[codebytere]

@guybedford if this is something that should be in v12.x can you please manually backport it? Feel free to swap the label with a dont-land if that's not the case.

[guybedford]

@codebytere thanks for letting me know about this - backport PR at #32287.