Reconsider pnpm's global virtual store for dev worktrees: disk and delete time

[JustYannicc]

#22765 (opt-in pnpm global virtual store for dev worktrees) was closed because pnpm 12 took most of the install-speed win. That weighed install speed only. The bigger cost for agent-heavy work is disk and delete time, when one machine keeps dozens of worktrees of this repo alive at once.

We measured on macOS (APFS), warm store, one worktree each, on an isolated volume so other writes couldn't skew it:

per worktree	per-project virtual store	global virtual store
physical disk	~160 MB	~5.5 MB
entries under node_modules	~372k	~2.4k
install	119–137 s	~2.6 s
rm -rf node_modules	143–153 s	~0.14 s

The shared store is a one-time ~7.6 GB. du makes per-project stores look like ~6–7 GB each because APFS clones are counted per copy; the physical cost is the ~160 MB above, plus 372k files' worth of metadata that every delete has to walk. With ~100 worktrees, that's ~15 GB and minutes of I/O per cleanup.

On the risks listed in the close: we run it unconditionally in our fork, CI included, so type resolution can't differ from CI. Vite, Metro and desktop packaging follow the store links through one shared helper.

Proposal: reconsider the global virtual store for dev worktrees, either on by default or behind the existing opt-in from #22765. If you're open to it, we'll send a focused PR scoped however you prefer.

[juliusmarminge]

Closing as a duplicate of #26926, which tracks the same request. Please follow along there.