Stop a cache-warning handoff from reading dune's make test as C/C++ - #385

What

A real failed CI run of ocaml/dune (run 29585452292) was classified cpp_build_failure 0.53 — but nothing in the log is a C/C++ compile. dune's own cram/blackbox tests fail through make test (a test-case diff, make: *** [test] Error 1).

Why the Crystal fix missed it

The Crystal fix demotes the bare GNU make recipe line while cpp_build_failure is the best match — but that deferral never runs if a benign warning first outscores it. The same dune job twice logs the benign cache save-warning Failed to save: Unable to reserve cache …, giving artifact_or_cache_failure two matched signals to cpp's one. So the make-recipe deferral was skipped, and the downstream benign-warning handoff — which yields to "the best rule that saw a real error" — then resurrected that same make-only cpp_build_failure as the verdict.

The fix

The benign-warning handoff now refuses to hand a verdict to a rule that is itself ambiguous-only (the bare make recipe or the RSpec-style summary line — the lines that name no ecosystem). The save-warning stays best and the run settles on unknown, the honest ceiling. Resource/network symptoms name their own category and stay eligible, so a buildkite job OOM-killed at exit code 137 still lands runner_resource_exhaustion; a genuine C/C++ build trips a real toolchain signal and is untouched.

Verification

• examples/real-world/dune-29585452292-excerpt.log: cpp_build_failure 0.53 → unknown 0.15
• Regression test tests/test_dune_make_not_cpp.py (incl. a guard that a real gcc error beside the same cache warning still classifies cpp_build_failure)
• Zoo benchmark unchanged: 223/223 top-1, and all 21 prior real-world logs unchanged (buildkite OOM fixture still runner_resource_exhaustion 0.71)
• Benchmark doc + CHANGELOG updated