ci: stop the publish job's cache save from erroring on package scratch

[KeyCode17]

Not a px-cache bug

The v1.10.0 publish job succeeded but posted failure annotations:

Error: ENOENT: no such file or directory, opendir
'/home/runner/work/px-solver/px-solver/target/package/pxsolver-cache-1.10.0/tests/trybuild'

Named crates: pxsolver-cache, pxsolver-camoufox, pxsolver-native — three of the six published crates that ship a tests/ directory. Nothing is wrong with any of them, and no manifest change fixes it.

Root cause

cargo publish leaves scratch in target/package/<crate>-<ver>/. For a crate shipping tests, that scratch contains a directory literally named tests. rust-cache's save-time cleanup treats any directory named tests under target/ as a possible nested test workspace and recurses into tests/target and tests/trybuild — a carve-out for trybuild and macrotest artifacts. This workspace has no trybuild dependency, so neither path exists.

From Swatinem/rust-cache src/cleanup.ts, cleanProfileTarget:

if (path.basename(profileDir) === "tests") {
  try {
    cleanTargetDir(path.join(profileDir, "target"), packages, checkTimestamp);   // not awaited
  } catch {}
  try {
    cleanTargetDir(path.join(profileDir, "trybuild"), packages, checkTimestamp); // not awaited
  } catch {}

The calls are async but not awaited, so the rejection never reaches the surrounding catch. It escapes as an unhandled rejection — which is why it renders as a failure annotation on a job that succeeded, and why it named an arbitrary subset of the eligible crates rather than all six.

Fix

save-if: false on the publish job's cache step. Publish scratch is single-use and worth nothing in the cache, and pre-flight already populates the same key, so the job restores as before and simply never runs the save-time cleanup. That removes the failing code path rather than racing it (an rm -rf target/package cleanup step would only cover the paths we know about today).

pre-flight and all three ci.yml jobs are untouched — none of them run cargo package, so none can hit this.

Verification

Workflow YAML parses; the with: save-if: false lands on crates-publish only, pre-flight unchanged.

Note this cannot be proven by a re-run: the publish loop skips crates already at the target version, so cargo publish never runs and target/package/ is never created. The next real publish (1.11.0) is the empirical confirmation. The fix is sound by construction — no save step means no cleanup walk.