I really don't love the solution here, but it's not clear there is a better way:

• npm ci does not allow you to disable node_modules removal
• Users include npm ci in the build step, instead of the install step, which cause node_modules to attempt to be removed which then fails the build since it's a read-only mount when used as a cache directory.
• We could attempt to redirect various caches that use node_modules/.cache to something like ~/.node_modules_cache but each tool uses their own way of configuring the cache path, and attempting to modify all of this tooling feels like it's not worth the squeeze (in addition to be confusing to users).
• We do need to document how to workaround "system" catches by defining your own railpack.json. I have that in my todo. This will be confusing to users.
• It seems like there is not a way around the read-only mountpath that buildkit uses for cache directories.
• Ideally the user could disable adding the node_modules/.cache, but I think it's better to not create more knobs for users to tune and instead (a) try to do the right thing via ugly regex-style stuff like we've done in this PR and (b) enable the user to override everything via a railpack.json if they are outside the golden path.

What I decided on was "post processing" the build plan after it's been merged with the user-provided railpack.json. It's more complex, but if we attempted to omit NODE_MODULES_CACHE in the node provider:

• If a user supplied a railpack.json with npm ci they'd run into the same issue, which is odd
• --build-cmd also ran into the same issue

However, a very obvious case still fails:

{
  "scripts": {
    "build": "npm ci"
  }
}

Without writing a web of complex code to analyize the content of build this would fail, and even that's a fools errand since they could shell out to anything that could attempt to rm -rf node_modules.

I still think the complexity here is worth it since it handles the common case and still gives advanced users the option to (a) understand npm ci should run in install and (b) write their own railpack.json to workaround the issue.

fixes #255