[Feature] More control in supportedArchitectures behavior

[wojtekmaj]

• I’d be willing to implement this feature (contributing guide)
• This feature is important to have in this repository; a contrib plugin wouldn’t do

Describe the user story

I’m using the checkoutable cache feature to make business-critical projects independent from the npm infrastructure.

To make my project work on all platforms I need to support:

• macOS, arm64
• Windows, arm64
• Linux, x64 / arm64

I configure this via:

supportedArchitectures:
  cpu:
    - arm64
    - x64
  os:
    - darwin
    - linux
    - win32
  libc:
    - glibc

This currently results in all possible combinations being fetched, which is excessive (already tracked in #4604).

The bigger issue is that both the cache and the actual node_modules get populated with all these combinations. This becomes especially painful as more native dependencies appear in my node_modules.

Example:
• biomejs: 29.3 MB when installed via npm
• biomejs: 199.7 MB when installed via Yarn with the above configuration

Describe the solution you’d like

Add a setting that changes the behavior of supportedArchitectures so that .yarn/cache conforms to the full list, but yarn install only installs packages for the current architecture.

Example:

supportedArchitecturesBehavior: 'all' # default
supportedArchitecturesBehavior: 'cache-only' # new

This would still allow reproducible, cross-platform offline installs, but without bloating node_modules for the active environment.

Describe the drawbacks of your solution

• Slightly more complexity in install logic
• Unclear how this would behave for other node linkers than node-modules and pnpm; needs to be documented. But is it really a problem in PnP, that modules working on other archs will be available?

Describe alternatives you’ve considered

• Manually pruning node_modules post-install (fragile and error-prone)