Require exact fast-mlsirm GPU JUnit evidence

.github/workflows/fast-mlsirm-ci.yml

@@ -173,14 +173,53 @@ jobs:
           from pathlib import Path
           from xml.etree import ElementTree
 
+          required = {
+              ("tests.test_marginal_parity", "test_marginal_gpu_agrees_with_cpu_loosely"),
+              ("tests.test_bifactor_gpu_high_q", "test_bifactor_gpu_parity_q121"),
+              ("tests.test_bifactor_gpu_high_q", "test_bifactor_gpu_parity_q241"),
+              ("tests.test_bifactor_gpu_high_q", "test_bifactor_gpu_parity_q481"),
+              ("tests.test_bifactor_gpu_high_q", "test_bifactor_gpu_parity_q241_wide_items_metal_workgroups"),
+              ("tests.test_bifactor_gpu_high_q", "test_bifactor_cpu_q121_vs_q241_agree"),
+              ("tests.test_bifactor_bootstrap_benchmark", "test_joint_bootstrap_cpu_vs_gpu_wall_time_q121"),
+          }
+
+          def count(node, field):
+              value = node.attrib.get(field, "")
+              if not value.isascii() or not value.isdecimal():
+                  raise SystemExit(f"GPU evidence has invalid {field} count")
+              return int(value)
+
           root = ElementTree.parse(Path("gpu-junit.xml")).getroot()
-          suites = [root] if root.tag == "testsuite" else list(root.findall("testsuite"))
-          tests = sum(int(suite.attrib.get("tests", "0")) for suite in suites)
-          skipped = sum(int(suite.attrib.get("skipped", "0")) for suite in suites)
-          if tests < 1:
+          if root.tag == "testsuite":
+              suites = [root]
+          elif root.tag == "testsuites" and all(child.tag == "testsuite" for child in root):
+              suites = list(root)
+          else:
+              raise SystemExit("GPU evidence has unsupported JUnit layout")
+          if not suites:
               raise SystemExit("GPU evidence did not execute a test")
-          if skipped != 0:
-              raise SystemExit(f"GPU evidence contained {skipped} skipped test(s)")
+          identities = []
+          for suite in suites:
+              if any(child.tag not in {"testcase", "properties", "system-out", "system-err"} for child in suite):
+                  raise SystemExit("GPU evidence has unsupported suite children")
+              cases = list(suite.findall("testcase"))
+              if list(suite.iter("testcase")) != cases:
+                  raise SystemExit("GPU evidence has nested testcase rows")
+              if count(suite, "tests") != len(cases):
+                  raise SystemExit("GPU evidence header/testcase count mismatch")
+              for field, tag in (("failures", "failure"), ("errors", "error"), ("skipped", "skipped")):
+                  actual = sum(len(list(case.iter(tag))) for case in cases)
+                  if count(suite, field) != actual or actual != 0:
+                      raise SystemExit(f"GPU evidence contained invalid or nonzero {field}")
+              identities.extend((case.get("classname"), case.get("name")) for case in cases)
+          if len(identities) != len(required) or set(identities) != required:
+              raise SystemExit("GPU evidence requires each exact selector once, without extras")
+          if root.tag == "testsuites":
+              for field in ("tests", "failures", "errors", "skipped"):
+                  if field in root.attrib:
+                      expected = len(identities) if field == "tests" else 0
+                      if count(root, field) != expected:
+                          raise SystemExit("GPU evidence aggregate header count mismatch")
           PY
 
   fuzz:

tests/test_fast_mlsirm_self_hosted_workflows.py

@@ -1,4 +1,8 @@
 from pathlib import Path
+import textwrap
+from xml.etree import ElementTree as ET
+
+import pytest
 
 
 ROOT = Path(__file__).resolve().parents[1]
@@ -10,6 +14,43 @@ def workflow_text(name: str) -> str:
     return (WORKFLOWS / name).read_text(encoding="utf-8")
 
 
+GPU_IDENTITIES = (
+    ("tests.test_marginal_parity", "test_marginal_gpu_agrees_with_cpu_loosely"),
+    ("tests.test_bifactor_gpu_high_q", "test_bifactor_gpu_parity_q121"),
+    ("tests.test_bifactor_gpu_high_q", "test_bifactor_gpu_parity_q241"),
+    ("tests.test_bifactor_gpu_high_q", "test_bifactor_gpu_parity_q481"),
+    ("tests.test_bifactor_gpu_high_q", "test_bifactor_gpu_parity_q241_wide_items_metal_workgroups"),
+    ("tests.test_bifactor_gpu_high_q", "test_bifactor_cpu_q121_vs_q241_agree"),
+    ("tests.test_bifactor_bootstrap_benchmark", "test_joint_bootstrap_cpu_vs_gpu_wall_time_q121"),
+)
+
+
+def _gpu_suite(identities=GPU_IDENTITIES) -> ET.Element:
+    suite = ET.Element("testsuite", tests=str(len(identities)), failures="0", errors="0", skipped="0")
+    for classname, name in identities:
+        ET.SubElement(suite, "testcase", classname=classname, name=name)
+    return suite
+
+
+def _run_gpu_guard(xml: ET.Element, tmp_path, monkeypatch) -> None:
+    step = workflow_text("fast-mlsirm-ci.yml").split("      - name: Reject skipped GPU evidence\n", 1)[1]
+    script = textwrap.dedent(step.split("python - <<'PY'\n", 1)[1].split("          PY\n", 1)[0])
+    (tmp_path / "gpu-junit.xml").write_bytes(ET.tostring(xml))
+    monkeypatch.chdir(tmp_path)
+    exec(compile(script, "fast-mlsirm-ci.yml::gpu-junit-guard", "exec"), {})
+
+
+def test_gpu_guard_accepts_only_the_exact_successful_seven(tmp_path, monkeypatch):
+    """Execute the producer guard against complete and false-admission evidence."""
+    _run_gpu_guard(_gpu_suite(), tmp_path, monkeypatch)
+    partial = _gpu_suite(GPU_IDENTITIES[:1])
+    skipped = _gpu_suite()
+    ET.SubElement(skipped[0], "skipped")
+    for xml in (partial, skipped):
+        with pytest.raises(SystemExit):
+            _run_gpu_guard(xml, tmp_path, monkeypatch)
+
+
 def test_fast_mlsirm_reusable_workflows_keep_isolated_runner_and_gate_contracts():
     """Require every reusable fast-mlsirm job to use the isolated runner boundary."""
     for name in ("ci", "codeql", "statistical-studies", "hourly-pr-governance", "pypi-gap-guard"):