Merge pull request #53 from FluffyAIcode/AgentMemory/v030-pr-n1-remove-verifier-doubles-8e7f

PR-N1: remove verifier-protocol test doubles from Linux CI gate

Author: FluffyAIcode

.github/workflows/ci.yaml

@@ -72,7 +72,23 @@ jobs:
           # PYTHONPATH route avoids a setuptools build step in CI.
           PYTHONPATH: .:sdks/python
         run: |
-          pytest \
+          # PR-N1 (ADR 0008) split: this gate covers ONLY
+          # verifier-independent code. The Linux runner cannot load
+          # real Qwen3 weights, and PR-N1 retired the FakeVerifier
+          # test double that previously stood in for them.
+          # Verifier-dependent modules — currently
+          # ``inference_engine.session.coordinator`` and
+          # ``inference_engine.session.generator`` — move to the
+          # tests/integration/ suite, gated on Mac M4 / CUDA hosts.
+          #
+          # Coverage is invoked via ``coverage run -m pytest`` rather
+          # than ``pytest --cov=`` to avoid a torch+pytest-cov race
+          # at conftest-import time on the hosted Linux runner
+          # (PR-N1 hit "module functions cannot set METH_CLASS or
+          # METH_STATIC" / SIGSEGV when --cov= started the tracer
+          # before torch's C extension finished initializing).
+          # Same pattern Mac M4 reviewer scripts use.
+          coverage run -m pytest \
             tests/inference_engine/server/ \
             tests/inference_engine/memory/ \
             tests/inference_engine/scheduler/ \
@@ -82,19 +98,13 @@ jobs:
             tests/sdk/python/ \
             tests/training/repr_align/ \
             tests/backends/mlx/test_env.py \
-            --cov=inference_engine.server \
-            --cov=inference_engine.memory \
-            --cov=inference_engine.scheduler \
-            --cov=inference_engine.pipeline \
-            --cov=inference_engine.session \
-            --cov=inference_engine.bench \
-            --cov=kakeya \
-            --cov=training.repr_align \
-            --cov-report=term \
-            --cov-report=xml:coverage.xml \
-            --cov-fail-under=100 \
             --junitxml=junit.xml \
             -v
+          coverage report \
+            --include='inference_engine/server/*,inference_engine/memory/*,inference_engine/scheduler/*,inference_engine/pipeline/*,inference_engine/session/store.py,sdks/python/kakeya/*,training/repr_align/*' \
+            --fail-under=100
+          coverage xml -o coverage.xml \
+            --include='inference_engine/server/*,inference_engine/memory/*,inference_engine/scheduler/*,inference_engine/pipeline/*,inference_engine/session/store.py,sdks/python/kakeya/*,training/repr_align/*'
 
       - name: Upload coverage artifact
         if: always()

results/platform-tests/smoke-all-prs-1780370637.junit.xml

@@ -0,0 +1,103 @@
+#!/usr/bin/env bash
+# Mac M4 review aid for PR-N1 (no-test-doubles cleanup, scope =
+# verifier-protocol mirror classes).
+#
+# PR-N1 retired FakeVerifier, _LyingVerifier, _RegressingVerifier,
+# and _LyingFakeVerifier from the Linux test tree. Their dispatch /
+# state-mirror tests moved to tests/integration/test_coordinator_real.py
+# and tests/integration/test_generator_real.py, where they run
+# against the real Qwen3-0.6B SinkWindowVerifier instead of a
+# hand-coded mirror. This script runs that integration suite on
+# Apple Silicon and produces the JSON evidence reviewers commit.
+#
+# Produces 1 artifact:
+#
+#   results/platform-tests/pr-n1-mac-integration-tests-<unix>.json
+#     pytest -m integration tests/integration/ — coordinator and
+#     generator integration tests against real Qwen3 + the existing
+#     INV-3 byte-exact GA gate. Acceptance: all tests pass.
+#
+# Usage (from repo root, on Mac M4 / arm64):
+#
+#     bash scripts/review_pr_n1_on_mac.sh
+#
+# Then commit:
+#
+#     git add results/platform-tests/pr-n1-mac-*
+#     git commit -m "Mac M4 review evidence for PR-N1"
+#     git push
+
+set -euo pipefail
+
+ROOT="$(cd "$(dirname "$0")/.." && pwd)"
+cd "$ROOT"
+
+stamp="$(date +%s)"
+out_dir="results/platform-tests"
+mkdir -p "$out_dir"
+
+junit="$out_dir/pr-n1-mac-integration-tests-${stamp}.junit.xml"
+report="$out_dir/pr-n1-mac-integration-tests-${stamp}.json"
+
+echo "==> integration suite (PR-N1 migrated tests + INV-3 GA gate)"
+PYTHONPATH=.:sdks/python python3 -m pytest \
+    -m integration \
+    tests/integration/ \
+    --junitxml="$junit" \
+    -v
+
+PYTHONPATH=.:sdks/python python3 - "$junit" "$report" <<'PY'
+import json
+import platform
+import sys
+import xml.etree.ElementTree as ET
+
+junit_path, out_path = sys.argv[1:3]
+jr = ET.parse(junit_path).getroot()
+
+testsuites = list(jr.iter("testsuite"))
+total_tests = sum(int(ts.get("tests", "0")) for ts in testsuites)
+total_failures = sum(int(ts.get("failures", "0")) for ts in testsuites)
+total_errors = sum(int(ts.get("errors", "0")) for ts in testsuites)
+total_skipped = sum(int(ts.get("skipped", "0")) for ts in testsuites)
+
+cases = []
+for tc in jr.iter("testcase"):
+    cases.append({
+        "classname": tc.get("classname"),
+        "name": tc.get("name"),
+        "time": float(tc.get("time", 0.0)),
+        "outcome": (
+            "failed" if tc.find("failure") is not None
+            else "errored" if tc.find("error") is not None
+            else "skipped" if tc.find("skipped") is not None
+            else "passed"
+        ),
+    })
+
+report = {
+    "schema_version": 1,
+    "kind": "pr_n1_mac_integration_tests",
+    "host": {
+        "platform": platform.platform(),
+        "machine": platform.machine(),
+        "python": platform.python_version(),
+    },
+    "junit": {
+        "tests": total_tests,
+        "failures": total_failures,
+        "errors": total_errors,
+        "skipped": total_skipped,
+        "cases": cases,
+    },
+}
+with open(out_path, "w", encoding="utf-8") as fh:
+    json.dump(report, fh, indent=2)
+print(f"  -> {out_path}")
+PY
+
+echo
+echo "==> Done. Commit:"
+echo "    git add $out_dir/pr-n1-mac-*"
+echo "    git commit -m 'Mac M4 review evidence for PR-N1'"
+echo "    git push"