Allow restoring the store cache without saving it (e.g. a save input) #56

Summary

Allow a job to restore the pnpm store cache without saving it, e.g. save: false (or cache: 'restore' alongside the existing boolean). There is currently no way to opt out of the post-step save while keeping the restore.

Why

The action assumes a workflow shape where saving is cheap because few jobs do it. That breaks down on a fan-out — a deploy matrix, or a reusable workflow called once per target — where many jobs share one lockfile and one runtime, and therefore one cache key.

In that shape only the first job to finish can usefully save. Every other job:

• rebuilds and compresses the whole store in its post step;
• loses the reservation and logs Failed to save: Unable to reserve cache with key …, another job may be creating this cache;
• discards the upload.

On a fan-out of around a hundred jobs, with a store of several hundred MB, that is roughly 18s of job time per job spent producing something guaranteed to be thrown away, plus a warning per job drowning out real ones. The work is pure loss: the content is identical to what the winner already published.

What these workflows want is one designated job that warms the store and publishes it, with everything downstream restoring read-only.

Why the existing inputs do not cover it

• cache: false also disables the restore, so the fan-out installs cold — much worse.
• install: false does not affect caching.
• There is no equivalent of actions/cache/restore's split, or of lookup-only.

Working around it means abandoning the action's caching entirely and hand-rolling actions/cache/restore plus a single actions/cache/save, which gives up the store-path detection and key construction this action exists to provide.

Sketch

# one upstream job warms and publishes the store
- uses: pnpm/setup@v2
  with:
    cache: true

# the fan-out restores it and never saves
- uses: pnpm/setup@v2
  with:
    cache: true
    save: false

save: false would skip registering the post step (or make it a no-op), leaving restore, the lockfile-verification cache and cache-hit untouched.

Relationship to #43

#43 moves saves onto a run-scoped key. That does not help this case — GITHUB_RUN_ID and GITHUB_RUN_ATTEMPT are per run, so every job in a matrix or reusable workflow computes the same save key and races it exactly as now, while entries accumulate per run. Commented there with detail. The two are orthogonal: this request removes the redundant saves by construction, and would compose with whatever keying #43 settles on.