ci: fit the Rust caches back inside the 10GB quota (#5532)

The repository has 18 rust-cache families of 1.2-3.1GB each against GitHub's
fixed 10GB per-repo quota. Measured usage sat at 9.72GB, 9.96GB and 11.63GB on
three samples, so LRU eviction is continuous and the main lanes lose: ci-test
and ci-e2e were repeatedly absent from the surviving entries. That is what
makes "Setup Rust environment" bimodal — 0.7-3.4 minutes warm against
11.8-20.9 minutes cold.

Four changes, all cache-only. No job builds or tests anything different.

Collapse ci.yml's nine cache sites into four keys, each with exactly one
writer: ci-dev (test-and-lint writes; ILM, debug-binary, e2e-tests and e2e-full
read), ci-feat-rio (rio-v2 lint writes, its debug-binary reads), ci-feat-proto
(the swift leg writes for both protocol legs), and ci-uring, which stays alone.
e2e-full is easy to miss here: it already shared ci-e2e with e2e-tests and both
saved, so without an explicit 'false' it would have become a second unnamed
writer of ci-dev. rio/swift/sftp are deliberately NOT merged — measured at
2365/2307/1200MB they are not near-identical, and none is a superset of the
others.

Give each writer a superset warm-up on main. A writer's own steps are not
automatically a superset of its readers': clippy emits metadata only, the
nextest pass excludes e2e_test, and no lint lane enables e2e-test-hooks, whose
different feature resolution yields a different -Cmetadata. Without these
builds the readers would restore a cache missing precisely what they need.
Guarded to main, so the PR critical path is unaffected.

Stop writing tag-scoped caches in build.yml. A cache saved on refs/tags/X can
only be restored by a re-run of that same tag, so each release cycle wrote up
to 12 unreadable 1-2GB entries that evicted the hot lanes. Tag builds still
restore the main-scoped cache. The one real cost is that re-running a failed
leg of the same tag now falls back to main's cache.

Flip the composite action's cache-save-if default to 'false' and make the input
mandatory in practice. audit.yml was relying on the old "true" default: every
PR touching Cargo.toml or Cargo.lock saved a second, PR-scoped copy (~843MB
measured, job 91048468127) that pushed main-scoped lanes out of the quota. The
fail-safe direction is a cold cache, not a stolen quota slice.
scripts/security/check_cache_save_if.sh now asserts every call site states it,
wired into audit.yml next to the existing pin check.

While there, cargo-deny stops pulling the full setup composite. It compiles
nothing, so apt, protoc, flatc and nextest were pure overhead — but it does run
cargo metadata, and Cargo.toml pins datafusion and s3s as git dependencies that
must be materialised into ~/.cargo/git, so the cache itself stays.

Refs: rustfs/backlog#1598, rustfs/backlog#1600

.github/workflows/build.yml

@@ -254,7 +254,15 @@ jobs:
           target: ${{ matrix.target }}
           cache-shared-key: build-${{ matrix.target }}
           github-token: ${{ secrets.GITHUB_TOKEN }}
-          cache-save-if: ${{ github.ref == 'refs/heads/main' || startsWith(github.ref, 'refs/tags/') }}
+          # main only. A cache saved on refs/tags/X is scoped to that tag: no
+          # other tag, no main run and no PR can restore it, so every release
+          # cycle wrote up to 12 entries of 1-2GB (preview tag plus final tag,
+          # six legs each) that nobody could read, evicting the hot lanes from
+          # the repo-wide 10GB quota. Tag builds still restore the main-scoped
+          # cache, since default-branch caches are readable from every ref.
+          # The one real cost: re-running a failed leg of the same tag no longer
+          # finds that tag's own warm cache and falls back to main's.
+          cache-save-if: ${{ github.ref == 'refs/heads/main' }}
           install-cross-tools: ${{ matrix.cross }}
 
       - name: Download static console assets