[moeghassi]

Summary

GetPackageJSON() in src/node_modules.cc treated every negative return
from ReadFileSync() identically: it negative-cached the path and reported the
manifest as absent. As a result a denied read (EACCES/EPERM, or an
anti-malware/EDR block that surfaces as a failed open) is indistinguishable from
ENOENT, so resolution silently falls back to index.js and the package loads
anyway.

This means an on-endpoint scanner cannot stop a require()/import of a
quarantined package by denying/quarantining its package.json — Node treats the
denied manifest as "no manifest" and runs the code.

Change

• Only ENOENT and ENOTDIR now mean "there is no package.json here"
  (unchanged behavior, still negative-cached).
• Any other read error is treated as a security-relevant signal and throws
  ERR_ACCESS_DENIED instead of falling back — so a denied/quarantined manifest
  aborts resolution for every package style (default index.js, "main", and
  "exports"), not just exports-based ones. The failure is intentionally
  not negative-cached so the deny isn't latched for the process lifetime.
• TraverseParent and GetPackageScopeConfig now stop walking parent scopes
  when an exception is pending, so the thrown error propagates instead of being
  swallowed by continuing up the tree.

Adds a POSIX regression test (chmod(0) the manifest → asserts
ERR_ACCESS_DENIED and that the index.js fallback did not run; skipped on
Windows and when running as root).

Notes for reviewers

• This is a behavior change to module resolution (fail-open → fail-closed on
  non-ENOENT read errors). It may warrant a semver-major label: users with
  a legitimately unreadable ancestor directory (EACCES while walking up for
  package-type detection) would now get a thrown error instead of silent
  fallback. I'm happy to scope the throw more narrowly (e.g. only the target
  package's own manifest, not ancestor type walks) if preferred.
• cc @nodejs/loaders @nodejs/security-wg — feedback on the desired posture and
  scope is very welcome.

I was unable to run a full local build to compile-check (the environment lacks
the ClangCL toolset required for the current main); CI will be the source of
truth here.

[mcollina]

lgtm

I’m unsure about the semversiness of this, as it’s a substantial behavior change.

[mcollina]

unfortunately it needs a rebase now

[moeghassi]

@mcollina rebased.