perf(covgate): instrument the build-cache pre-warm

[ben-miru]

What

Make covgate's build-cache pre-warm instrumented, matching what the coverage runs actually compile. Two passes instead of one plain pass:

• cover paths (the measured packages): go test -run=^$ -cover <pkgs> — -cover defaults to instrumenting each package for itself, matching covgate's per-package -coverpkg=<pkg>.
• plain paths (the ./tests/... dirs): go test -run=^$ <dirs> — covgate compiles integration test packages as plain deps, so these stay non-instrumented.

Why

The pre-warm compiled non-instrumented test binaries; covgate's coverage runs compile instrumented ones. The saved build cache therefore stored both artifact sets — in production (backend) the gobuild-test cache roughly doubled (~2.2 GB → ~4.3 GB), which pushed the repo back over the 10 GB Actions cache quota (re-triggering LRU eviction → colder restores → high run-to-run variance), and the warm output was only partly reused by the instrumented runs.

Instrumenting the warm pass so it produces the same objects covgate consumes should let the cache hold ~one artifact set and let covgate reuse the pre-warmed instrumented objects instead of rebuilding them.

Prototype — what to measure

This is a prototype; the proof is a backend CI run once this is released and bumped:

• gobuild-test cache size should fall back toward ~2.2 GB (the headline metric — resolves the quota regression).
• Cache restore time should drop with the smaller cache.
• Cold-run penalty (pre-warm + covgate on a cache-cold PR) should ease, since covgate reuses more.
  If the cache doesn't shrink meaningfully, revert — the plain pre-warm is simpler.

Tests / validation

• PrewarmBuild tests updated to the two-arg form (empty/empty → nil; build error in either pass propagates).
• collectWarmPaths split tested: cover = measured pkgs, plain = de-duplicated ./tests/... dirs.
• Existing no-op prewarm seams updated to func([]string, []string) error.
• go build, go vet, go test, repo lint, and the covgate self-gate (covgate 100%, gocover 93.8%) all pass; no .covgate floors lowered.

Gate (parallelism > 1 && len(pkgs) > 1), the pre-warm duration logging (v0.2.1), and "Total time" semantics are unchanged.

🤖 Generated with Claude Code