fix(ci): streamline model checks and acceptance evidence

.github/workflows/ci.yml

@@ -206,9 +206,6 @@ jobs:
           path: models/qwen2.5-0.5b-instruct-q4_k_m.gguf
           key: qwen2.5-0.5b-q4-k-m-74a4da8c9fdb
 
-      - name: Fetch and Verify Pinned GGUF
-        run: ./scripts/fetch_real_test_models.sh --gguf-only
-
       - name: Run Real Model and Public Profile
         run: |
           ccache -z
@@ -273,20 +270,14 @@ jobs:
             ccache-${{ runner.os }}-${{ github.ref }}-
             ccache-${{ runner.os }}-
 
-      - name: Cache Pinned Whisper Models
+      - name: Cache Pinned Whisper Model
         uses: actions/cache@v4
         with:
-          path: |
-            models/ggml-base.bin
-            models/qwen2.5-0.5b-instruct-q4_k_m.gguf
-            models/ggml-tiny-q5_1.bin
-          key: whisper-models-${{ hashFiles('models/asset_manifest.json') }}
+          path: models/ggml-base.bin
+          key: whisper-e2e-${{ hashFiles('models/asset_manifest.json') }}
           restore-keys: |
-            whisper-models-
-            whisper-ggml-base-60ed5bc3dd14
-
-      - name: Fetch and Verify Pinned Whisper Model
-        run: ./scripts/fetch_real_test_models.sh --whisper
+            whisper-e2e-
+            whisper-ggml-base-
 
       - name: Run Real Whisper E2E and Public Profile
         run: |

doc/VERIFIABLE_SELECTION.md

@@ -39,7 +39,7 @@ Studio 的“另存为可运行方案”和“运行草稿”共用配置生成
 
 [models/asset_manifest.json](../models/asset_manifest.json) 包含 11 个权重/sidecar 条目的 SHA-256 与 8 个现有资产组合。权重、tokenizer、Kite 运行配置与视觉 projector 均纳入检查。
 
-- 下载命令为 `./scripts/fetch_real_test_models.sh --all`、`--kite`、`--whisper` 或 `--gguf-only`，精确 URL 和 SHA 统一从清单读取。
+- 下载命令为 `./scripts/fetch_real_test_models.sh --all`、`--kite`、`--whisper`、`--gguf-only` 或仅供 Whisper 真实模型测试使用的 `--whisper-e2e`，精确 URL 和 SHA 统一从清单读取。
 - 清单中的 Model/Backend 配置是可选择的起点；兼容性仍由当前执行文件的 Catalog 校验。
 - 清单路径相对资产目录；Pipeline 的 `models[].model_path` 相对宿主部署根；tokenizer、运行配置等 sidecar 路径相对实际模型所在目录解析。
 - 变更 tokenizer/运行配置路径后不会借用旧组合的校验结论，而会变为 `unregistered`。接入新资产时，补充清单中的 `artifacts`、`selections.paths` 与完整 `files`，再实际校验。

models/asset_manifest.json

@@ -7,8 +7,7 @@
       "download_groups": [
         "all",
         "gguf-only",
-        "kite",
-        "whisper"
+        "kite"
       ]
     },
     "bge_base_zh_v1.5.onnx": {
@@ -62,7 +61,8 @@
       "url": "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-base.bin",
       "download_groups": [
         "all",
-        "whisper"
+        "whisper",
+        "whisper-e2e"
       ]
     },
     "ggml-tiny-q5_1.bin": {

scripts/fetch_real_test_models.sh

@@ -8,7 +8,7 @@ MODEL_DIR="${PROJECT_ROOT}/models"
 MODE="all"
 
 if [[ $# -gt 1 ]]; then
-  echo "Usage: $0 [--all | --gguf-only | --kite | --whisper]"
+  echo "Usage: $0 [--all | --gguf-only | --kite | --whisper | --whisper-e2e]"
   exit 2
 fi
 if [[ $# -eq 1 ]]; then
@@ -17,8 +17,9 @@ if [[ $# -eq 1 ]]; then
     --gguf-only) MODE="gguf-only" ;;
     --kite) MODE="kite" ;;
     --whisper) MODE="whisper" ;;
+    --whisper-e2e) MODE="whisper-e2e" ;;
     *)
-      echo "Usage: $0 [--all | --gguf-only | --kite | --whisper]"
+      echo "Usage: $0 [--all | --gguf-only | --kite | --whisper | --whisper-e2e]"
       exit 2
       ;;
   esac

scripts/generate_acceptance_evidence.sh

@@ -53,7 +53,7 @@ import json
 import os
 
 evidence = {
-    "schema_version": 2,
+    "schema_version": 3,
     "scope": os.environ["EVIDENCE_SCOPE"],
     "generated_at_utc": datetime.datetime.now(
         datetime.timezone.utc).isoformat(),
@@ -65,7 +65,7 @@ evidence = {
     "worktree_clean": os.environ["WORKTREE_CLEAN"] == "true",
     "gates": {
         "canonical_run_all_tests": os.environ["CANONICAL_GATE"],
-        "full_address_undefined_sanitizer": os.environ["SANITIZER_GATE"],
+        "ci_runtime_address_undefined_sanitizer": os.environ["SANITIZER_GATE"],
         "real_c_abi_and_public_profile": os.environ["REAL_GATE"],
     },
 }

scripts/run_real_model_e2e.sh

@@ -39,7 +39,7 @@ echo "=================================================================="
 
 # 1. 拉取或校验固定提交的真实模型权重。
 if [[ "${MODE}" == "whisper" ]]; then
-  "${PROJECT_ROOT}/scripts/fetch_real_test_models.sh" --whisper
+  "${PROJECT_ROOT}/scripts/fetch_real_test_models.sh" --whisper-e2e
 elif [[ "${MODE}" == "all" ]]; then
   "${PROJECT_ROOT}/scripts/fetch_real_test_models.sh" --all
 else

tests/contract/architecture/test_quality_gate_contract.py

@@ -252,7 +252,8 @@ def check_real_model_contract(root, env, log):
         configurations.append((configure, preset, False, True))
         assert "-DENABLE_REAL_MODEL_TESTS=ON" in configure
         commands = [row["command"] for row in records]
-        assert commands[0] == ["fetch_real_test_models.sh", "--" + mode]
+        fetch_mode = "--whisper-e2e" if mode == "whisper" else "--" + mode
+        assert commands[0] == ["fetch_real_test_models.sh", fetch_mode]
         for row in records:
             if row["command"][0] in ("test_real_models_e2e", "alg_demo"):
                 assert row["executable"] == str(build / row["command"][0]), row
@@ -344,16 +345,28 @@ def main():
         workflow = (ROOT / ".github/workflows/ci.yml").read_text()
         assert "WHISPER_GATE_RESULT: ${{ needs.whisper-asr.result }}" in workflow
         assert "KITELLM_GATE_RESULT: ${{ needs.kite-llm.result }}" in workflow
+        assert "run: ./scripts/fetch_real_test_models.sh --gguf-only" not in workflow
+        assert "run: ./scripts/fetch_real_test_models.sh --whisper" not in workflow
+        manifest = json.loads((ROOT / "models/asset_manifest.json").read_text())
+        artifact_groups = {
+            group: {name for name, artifact in manifest["artifacts"].items()
+                    if group in artifact.get("download_groups", [])}
+            for group in ("whisper", "whisper-e2e")
+        }
+        assert artifact_groups["whisper-e2e"] == {"ggml-base.bin"}
+        assert artifact_groups["whisper"] == {"ggml-base.bin", "ggml-tiny-q5_1.bin"}
         evidence = root / "evidence.json"
         for state in ("success", "failure", "skipped", "cancelled"):
             evidence_env = {**os.environ, "WHISPER_GATE_RESULT": state,
                             "KITELLM_GATE_RESULT": "skipped"}
             result = run([str(root / "scripts/generate_acceptance_evidence.sh"), str(evidence),
                           "success", "failure", "cancelled"], env=evidence_env)
             assert result.returncode == 0, result.stdout + result.stderr
-            gates = json.loads(evidence.read_text())["gates"]
+            report = json.loads(evidence.read_text())
+            assert report["schema_version"] == 3
+            gates = report["gates"]
             assert gates == {"canonical_run_all_tests": "success",
-                             "full_address_undefined_sanitizer": "failure",
+                             "ci_runtime_address_undefined_sanitizer": "failure",
                              "real_c_abi_and_public_profile": "cancelled",
                              "whisper_asr_backend_and_real_profile": state,
                              "kitellm_private_release_and_real_gguf": "skipped"}

tests/e2e/real_models/test_real_models_e2e.cpp

@@ -23,9 +23,6 @@ class RealModelE2ETest : public ::testing::Test {
         std::filesystem::path(LLM_EDGEFLOW_PROJECT_SOURCE_DIR));
     model_root_ = project_root_ / "models";
     model_path_ = model_root_ / "qwen2.5-0.5b-instruct-q4_k_m.gguf";
-    ASSERT_TRUE(std::filesystem::is_regular_file(model_path_))
-        << "ENABLE_REAL_MODEL_TESTS requires pinned artifacts; run "
-           "./scripts/fetch_real_test_models.sh --gguf-only";
   }
 
   std::filesystem::path project_root_;
@@ -49,6 +46,9 @@ class RealModelE2ETest : public ::testing::Test {
 
 // 1. 真实 Qwen GGUF 物理前向与自回归 Token 生成测试
 TEST_F(RealModelE2ETest, RealQwenGgufTextGeneration) {
+  ASSERT_TRUE(std::filesystem::is_regular_file(model_path_))
+      << "ENABLE_REAL_MODEL_TESTS requires pinned artifacts; run "
+         "./scripts/fetch_real_test_models.sh --gguf-only";
   auto model = CreateModel();
   ASSERT_NE(model, nullptr);
 
@@ -80,6 +80,9 @@ TEST_F(RealModelE2ETest, RealQwenGgufTextGeneration) {
 
 // 2. 真实 Qwen 模型在 FixedBatchExecutor 定长对齐批推理压测
 TEST_F(RealModelE2ETest, RealQwenBatchExecutionWithPadding) {
+  ASSERT_TRUE(std::filesystem::is_regular_file(model_path_))
+      << "ENABLE_REAL_MODEL_TESTS requires pinned artifacts; run "
+         "./scripts/fetch_real_test_models.sh --gguf-only";
   auto model = CreateModel();
   ASSERT_NE(model, nullptr);
 
@@ -110,6 +113,9 @@ TEST_F(RealModelE2ETest, RealQwenBatchExecutionWithPadding) {
 
 // 3. 真实模型接入 Operator 全链路端到端验证
 TEST_F(RealModelE2ETest, RealModelOperatorEndToEnd) {
+  ASSERT_TRUE(std::filesystem::is_regular_file(model_path_))
+      << "ENABLE_REAL_MODEL_TESTS requires pinned artifacts; run "
+         "./scripts/fetch_real_test_models.sh --gguf-only";
   std::vector<std::string> sentences = {
       "李雷在微软北京研发中心负责AI大模型芯片开发。"};
   std::ifstream corpus(project_root_ / "data/corpus_entity_extract.txt");