[mathiasrw]

Closes #788

The first PR of a new train. #789 and #790 go on top of it.

Change

• Before, fs.statSync threw everything but ENOENT, and the run crashed with a stack trace. Now:
  • A path that is not there, or has a file where a folder should be (ENOENT, ENOTDIR), goes on to be globbed, as before.
  • A glob that can not be looked at as a path, like a long brace glob, goes on to be globbed.
  • A plain path that can not be looked at, like a name that is too long or a symlink loop, is a ❌ error for that path (exit code 2), and the run goes on with the other files. So it is not hidden as 0 files found.
• When the glob itself fails (like a.md/x, where a.md is a file), it is an error (exit code 2), and the run goes on with the files that can be found:

> rexreplace foo bar a.md/x
 ❌ Not all paths could be accessed
    ENOTDIR: not a directory, scandir '…/a.md'

 🟡 0 files found

• The same for the paths from --glob-pipe. A glob of --exclude-glob that fails is a panic (exit code 3), as the files it should keep out can not be known.

With Bun and Node the long brace glob from the issue now just works: the files it can find are found.

From the review of 23 October.

Tests

yarn test-ci passes: the demo on Node and Bun, and 1028 Bun tests. test/fix/788.test.js is new, and all its tests fail without the fix.

After the review of 24 October

• A glob that meets a folder that can not be read, like one with no access, is now an error (Not all paths could be accessed: EACCES …), and the run goes on with the files that can be found. The exit code is 2, like for a file that can not be read. Before, one locked folder anywhere in '**/*.md' stopped the whole run, and no file was changed.
• With --halt it still stops at once.
• A glob of --exclude-glob that fails is still a panic, as the files it should keep out can not be known, and they must not be changed.
• The title now says what is fixed.

After the review of 24 October

• The error says Not all paths could be accessed: EACCES ….