Yarn mislinks devDependencies when NODE_ENV=production

[STRML]

Tested this with v0.17.9 and with v0.17.10, Node 6.9.1.

The following package.json:

{
  "devDependencies": {
    "mocha": "^3.2.0"
  },
  "dependencies": {
    "markdown-it": "^8.2.0"
  }
}

Produces this .bin directory when running yarn:

$ yarn
yarn install v0.17.9
warning No license field
[1/4] 🔍  Resolving packages...
[2/4] 🚚  Fetching packages...
[3/4] 🔗  Linking dependencies...
[4/4] 📃  Building fresh packages...
✨  Done in 0.51s.
~/g/f/yarn-link (master|…) $ ll node_modules/.bin
total 24
lrwxr-xr-x  1 samuelreed  staff    19B Dec  1 11:09 _mocha -> ../mocha/bin/_mocha
lrwxr-xr-x  1 samuelreed  staff    33B Dec  1 11:09 markdown-it -> ../markdown-it/bin/markdown-it.js
lrwxr-xr-x  1 samuelreed  staff    18B Dec  1 11:09 mocha -> ../mocha/bin/mocha

But this incorrect .bin when running env NODE_ENV=production yarn:

$ env NODE_ENV=production yarn
yarn install v0.17.9
warning No license field
[1/4] 🔍  Resolving packages...
success Already up-to-date.
✨  Done in 0.10s.
~/g/f/yarn-link (master|…) $ ll node_modules/.bin
total 24
lrwxr-xr-x  1 samuelreed  staff   102B Dec  1 11:10 _mocha -> ../../../../../Library/Caches/Yarn/npm-mocha-3.2.0-7dc4f45e5088075171a68896814e6ae9eb7a85e3/bin/_mocha
lrwxr-xr-x  1 samuelreed  staff    33B Dec  1 11:10 markdown-it -> ../markdown-it/bin/markdown-it.js
lrwxr-xr-x  1 samuelreed  staff   101B Dec  1 11:10 mocha -> ../../../../../Library/Caches/Yarn/npm-mocha-3.2.0-7dc4f45e5088075171a68896814e6ae9eb7a85e3/bin/mocha

This can cause these bins to fail, for example with nsp, which fails resolving its dependency, subcommand.

Same results on v0.17.10, OSX 10.11, and on Ubuntu 14.04. I can confirm this is a regression from v0.16.1.

Might be related to #1379, #761.

[sebmck]

Fixed via #2116.

[joeyespo]

I'm seeing this happen again on Heroku when NODE_ENV=production and NPM_CONFIG_PRODUCTION=false. (See heroku/heroku-buildpack-nodejs#337)

@kittens It's not clear how #2116 fixes the problem. Happen to know off the top of your head where in that PR the fix is for this specific issue? (I'd be happy to PR a new fix for this regression, but digging through that large commit isn't giving any clarity as to how it was fixed the first time.)

[STRML]

This is still broken with the same reproduction as in the original ticket.