fix(onnx): fetch pinned baseline when CI clone is shallow

[FeelTheBeats]

The self-hosted CI sync does a depth-1 fetch, so the pinned baseline commit (20b105e) is absent and git rev-parse fails with rc=128. Resolve the baseline on demand via git fetch and cover it with a regression test.

[github-actions]

🤖 AI Code Review

共审查 2 个变更文件

📁 benchmarks/bench_onnx_operators.py

Review: benchmarks/bench_onnx_operators.py

---

🟡 git 参数注入风险 — resolve_commit 中 git fetch 未使用 --end-of-options，而 rev-parse 已正确使用了：

# 当前
git("fetch", "--depth=1", "origin", ref)

# 建议
git("fetch", "--depth=1", "--end-of-options", "origin", ref)

若 ref 以 - 开头（例如 --tags），会被 git fetch 解析为参数而非 refspec，可能改变行为。虽然实际传入的是 baseline_ref（commit hash/tag），但防御性处理一致更好——rev-parse 已经做了，fetch 也应该做。

---

💭 fetch 后再次失败时无诊断上下文 — 如果 fetch 成功但第二次 rev-parse 仍然失败（例如 ref 在远端也不存在），抛出的异常不会提示"已尝试 fetch"。考虑加一条 stderr 日志或用 logger.warning 标注 fetch 阶段，方便 CI 排障时区分"本地缺失"和"远端也缺失"。

---

💭 git fetch 失败时的静默传播 — 如果 fetch 因网络问题失败，异常直接向上传播，用户看到的是 raw CalledProcessError。对内部 benchmark 脚本可接受，但如果此模块被外部调用，值得考虑包装一层带上下文的异常。

---

整体评价：逻辑正确，docstring 清晰，"先试后拉再重试"的三层 fallback 设计合理。主要问题是 fetch 调用缺少 --end-of-options，与同函数中 rev-parse 的防御风格不一致，建议统一。

---

📁 tests/test_onnx_operator_benchmark.py

🟡 Fragile mock state — The any(call[0] == "fetch" for call in calls) check conflates call history with state. If the logic ever grows (e.g., a fetch handler that itself calls rev-parse), this becomes a footgun. A simple boolean flag would be clearer:

def test_missing_baseline_commit_is_fetched_from_origin(monkeypatch):
    calls = []
    fetched = False

    def fake_git(*arguments, binary=False):
        nonlocal fetched
        calls.append(arguments)
        if arguments[0] == "fetch":
            fetched = True
        if arguments[0] == "rev-parse" and not fetched:
            raise subprocess.CalledProcessError(128, ["git", *arguments])
        return bench.BASELINE if arguments[0] == "rev-parse" else ""

🟡 No call-count assertion — The test passes if resolve_commit fetches once and retries, or skips the first failed rev-parse entirely. Add a check that rev-parse was called exactly twice (once failing, once succeeding) to lock down the expected flow:

assert sum(1 for c in calls if c[0] == "rev-parse") == 2

💭 binary kwarg coupling — The mock assumes bench.git accepts binary as keyword-only. If the real function signature changes, this mock silently mis-routes calls. Consider def fake_git(*args, **kwargs) to be resilient to signature drift.

---