fix(integration): pin codex model + demote false pinned-baseline parity blocker

.github/scripts/build_integration_review_pack.py

@@ -409,7 +409,18 @@ def compute_verdict(slots: list[Slot], parity: list[ParityResult]) -> Verdict:
             )
 
     for pr in parity:
-        if pr.status == "fail":
+        if pr.status == "fail" and pr.kind == "pinned-baseline":
+            # The pinned-baseline reward-band gate currently feeds a NATIVE HF
+            # leaderboard baseline to the Harbor-schema + git-pinned checker, which
+            # structurally false-fails (missing Harbor fields / pin mismatch) — NOT
+            # a real reward regression. Quarantine it (visible, non-blocking) until
+            # check_skillsbench_harbor_parity gains a native-vs-native baseline mode
+            # (tracked follow-up). Within-PR docker/daytona parity still hard-blocks.
+            quarantines.append(
+                f"parity {pr.kind} (advisory — gate needs native-baseline mode): "
+                f"{pr.pair_id} — {pr.detail}"
+            )
+        elif pr.status == "fail":
             blockers.append(f"parity {pr.kind} fail: {pr.pair_id} — {pr.detail}")
         elif pr.status == "quarantine":
             quarantines.append(f"parity {pr.kind}: {pr.pair_id} — {pr.detail}")

.github/workflows/integration-final-review.yml

@@ -514,6 +514,11 @@ jobs:
           # codex_review.py uses for the CLI (dropping the DeepSeek base URL) so
           # the judge clobber never leaks into codex auth.
           CODEX_API_KEY: ${{ secrets.OPENAI_API_KEY }}
+          # Pin the codex composer to a model the OPENAI_API_KEY is PROVEN to serve
+          # (the codex-acp matrix cells use gpt-5.4-nano). Otherwise `codex exec`
+          # falls back to its built-in default (gpt-5.x-codex), which the key may
+          # not be entitled to -> model_not_found -> 'codex unavailable'.
+          CODEX_MODEL: gpt-5.4-nano
           DETERMINISTIC_VERDICT: ${{ steps.pack.outputs.deterministic_verdict }}
         run: |
           set +e

.github/workflows/integration-scope.yml

@@ -502,6 +502,9 @@ jobs:
           # via CODEX_API_KEY — codex_review.py isolates the CLI env (real key +
           # default OpenAI base) so the judge clobber never leaks into codex auth.
           CODEX_API_KEY: ${{ secrets.OPENAI_API_KEY }}
+          # Pin the codex composer to a model the key is proven to serve (else
+          # `codex exec` uses its built-in default, which the key may not serve).
+          CODEX_MODEL: gpt-5.4-nano
         run: |
           set +e
           uv run python .github/scripts/codex_review.py \

tests/test_build_review_pack.py

@@ -39,6 +39,28 @@ class SkipTest(Exception):
     """Raised to skip a test under the stdlib runner / signal pytest.skip."""
 
 
+def test_pinned_baseline_parity_fail_demotes_to_quarantine() -> None:
+    # Regression: the pinned-baseline reward-band gate currently false-fails
+    # (a native HF leaderboard baseline run through the Harbor-schema + git-pinned
+    # checker), so a pinned-baseline FAIL must DEMOTE to a quarantine — visible but
+    # non-blocking — not a hard 'not mergeable'.
+    pinned_fail = pack_mod.ParityResult(
+        "pinned-baseline", "pinned-baseline", "fail", "missing Harbor field(s)"
+    )
+    v = pack_mod.compute_verdict([], [pinned_fail])
+    assert v.verdict == pack_mod.VERDICT_QUARANTINES
+    assert not v.blockers
+    assert v.quarantines
+
+    # ...but a REAL within-PR docker/daytona parity FAIL still hard-blocks.
+    within_fail = pack_mod.ParityResult(
+        "sandbox-parity(x)", "within-pr", "fail", "incomplete parity pair"
+    )
+    v2 = pack_mod.compute_verdict([], [within_fail])
+    assert v2.verdict == pack_mod.VERDICT_NOT_MERGEABLE
+    assert v2.blockers
+
+
 # ------------------------------------------------------------------
 # Fixtures: hand-built flat rollouts + a matrix plan.
 # ------------------------------------------------------------------