test: Add test case for regression in npm/cli (#203)

[TrevorBurnham]

I looked into #203. What's happening is that Arborist is providing npa with a path where the # character has already been URL-encoded, like this:

npa('file:/test_%23dir')

Prior to #200, the result looked like this:

{
  type: 'directory',
  name: null,
  rawSpec: 'file:/test_%23dir',
  fetchSpec: '/test_#dir',
  saveSpec: 'file:/test_#dir',
}

After #200, the result looks like this:

{
  type: 'directory',
  name: null,
  rawSpec: 'file:/test_%23dir',
  fetchSpec: '/test_%23dir',
  saveSpec: 'file:/test_%23dir',
}

I'm not sure what the best fix for this is, but I hope this (failing) test case helps. I believe the key change is around this line:

npm-package-arg/lib/npa.js

Line 316 in fbbf22e

resolvedUrl = new URL(rawSpec, `${pathToFileURL(path.resolve(where))}/`)

cc @wraithgar @reggi

[wraithgar]

This is a tricky one. I think supporting this is the wrong direction. This was a very deep-seated bug that took quite a bit of research to figure out, even going in and pinging the maintainers of Node.js itself who were well versed in the ecma spec on these things. The connection between a file: url, a path, and other kinds of urls is ... pretty complicated. URI encoding doesn't overlap with file urls, and there are also special functions like url.pathToFileURL.

At the end of the day, making the tool do the right thing was the bugfix. It should require valid input. This should have been fixed here originally instead of worked around in npm itself.

I'm going to tentatively close this based on the fact that npm/cli#8115 exists.