Cover the unprovisioned-model category and fix a wrong annotation

Greptile (5/5, safe to merge) named the two gaps on this PR.

UPSTREAM_MODEL_UNAVAILABLE had no tests. It is the taxonomy's most dangerous
silent failure -- unclassified, a model that is configured but not deployed
makes every call 404, the agent writes no answer, and each case is recorded as
an informative task failure, i.e. the harness scoring the candidate down for a
model that does not exist. Tests now pin the provider error codes it matches,
its terminating/not-a-sample/counts-toward-invalidity policy, and the
deliberate narrowness: a bare "does not exist" must stay unmatched, or it would
swallow "loading container: file does not exist".

load_build was annotated `-> dict` while returning `(dict, Path)`; the caller
unpacks two values. Corrected rather than merely added.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

Author: varunursekar

harness-engineering-bench/scripts/rescore_candidate.py

@@ -48,7 +48,7 @@ def log(message: str) -> None:
     print(f"[rescore] {message}", flush=True)
 
 
-def load_build(benchmark: str) -> dict:
+def load_build(benchmark: str) -> tuple[dict, Path]:
     import yaml  # provided by the vero environment
 
     path = BENCH_ROOT / benchmark / "baseline" / "build.yaml"

vero/tests/test_v05_error_taxonomy.py

@@ -20,6 +20,45 @@ def test_classify_signal_recognizes_infrastructure_categories():
     )
 
 
+def test_classify_signal_recognizes_an_unprovisioned_upstream_model():
+    """A configured-but-not-deployed model must not be blamed on the candidate.
+
+    Left unclassified this is the worst silent failure in the taxonomy: every
+    call 404s, the agent writes no answer, and each case is recorded as an
+    informative task failure -- the harness scoring the candidate down for a
+    model that does not exist.
+    """
+    for signal in (
+        "DeploymentNotFound",
+        "model_not_found",
+        "The API deployment for this resource does not exist",
+        "openai.NotFoundError: model_not_found",
+    ):
+        assert (
+            classify_signal(signal) is ErrorCategory.UPSTREAM_MODEL_UNAVAILABLE
+        ), signal
+
+    # Deliberately narrow: a bare "does not exist" is a container/file problem,
+    # not a missing model, and matching it here would swallow real infra errors.
+    assert (
+        classify_signal("loading container: file does not exist")
+        is not ErrorCategory.UPSTREAM_MODEL_UNAVAILABLE
+    )
+
+
+def test_unprovisioned_model_policy_is_terminating_and_not_a_sample():
+    """It is a permanent misconfiguration, so it stops the run and scores nothing."""
+    p = policy(ErrorCategory.UPSTREAM_MODEL_UNAVAILABLE)
+    assert p.retryable is False  # every remaining case fails identically
+    assert p.terminating is True
+    assert p.is_informative_sample is False  # never scored as a candidate failure
+    # Unlike auth, still counts toward invalidity: if the terminating path is
+    # ever bypassed, the aggregate must come out invalid rather than averaging
+    # a shrinking set of survivors.
+    assert p.counts_toward_invalidity is True
+    assert p.diagnostic_code == "upstream_model_unavailable"
+
+
 def test_classify_signal_leaves_task_failures_unclassified():
     # The benign "produced no answer" marker and a candidate's own bug are not
     # infrastructure — the case-level classifier turns both into task failures.