feat: add soak runner with budget guards and invariant scoreboard

tools/soak/scoreboard.py

@@ -0,0 +1,134 @@
+"""压测记分板: 六条硬不变量断言 + 报告产出(findings §4;吃自己遥测的狗粮)。
+
+输入口径三类(M2 计划 T11): 不变量 2/3/5/6 = 纯函数(遥测行迭代器);
+不变量 1(记账归零)= 活 Redis/后端检查(async);不变量 4(RSS)= 纯函数
+(run_soak 周期采样 json)。断言函数抛 AssertionError 即不变量被击穿。
+"""
+
+from __future__ import annotations
+
+import sqlite3
+from collections import Counter
+from datetime import datetime
+from pathlib import Path
+from typing import Any
+
+Row = dict[str, Any]
+
+
+def load_rows(*db_paths: Path | str) -> list[Row]:
+    """合并全部 worker 遥测库的 llm_calls 行(dict 形态)。"""
+    rows: list[Row] = []
+    for path in db_paths:
+        conn = sqlite3.connect(path)
+        conn.row_factory = sqlite3.Row
+        try:
+            rows.extend(dict(r) for r in conn.execute("SELECT * FROM llm_calls"))
+        finally:
+            conn.close()
+    return rows
+
+
+# —— 硬不变量(纯函数) ——
+
+
+def inv_rows_match_calls(rows: list[Row], *, expected_calls: int, tolerance: int = 0) -> None:
+    """不变量 2a: 遥测行数 == 发出的请求数(±取消双记的已知语义容差)。"""
+    assert abs(len(rows) - expected_calls) <= tolerance, (
+        f"遥测行数 {len(rows)} ≠ 请求数 {expected_calls}(容差 {tolerance})"
+    )
+
+
+def inv_call_ids_unique(rows: list[Row]) -> None:
+    """不变量 2b: call_id 无重复。"""
+    dupes = [cid for cid, n in Counter(r["call_id"] for r in rows).items() if n > 1]
+    assert not dupes, f"call_id 重复: {dupes[:5]}"
+
+
+def _minute_bucket(created_at: str) -> int:
+    return int(datetime.fromisoformat(str(created_at)).timestamp()) // 60
+
+def inv_rpm_never_exceeded(rows: list[Row], per_source_rpm: dict[str, int]) -> None:
+    """不变量 3: 按遥测时间戳重算,任一分钟桶内单源请求数 ≤ RPM 配置。
+
+    口径 = 固定分钟窗口(与限流器 `int(sec/60)` 同源);滑动 60s 窗会对
+    "窗尾+窗头"的合法背靠背流量误报,不采用(findings §4 条 3 的执行口径)。
+    """
+    buckets: Counter[tuple[str, int]] = Counter(
+        (r["source_name"], _minute_bucket(r["created_at"]))
+        for r in rows
+        if per_source_rpm.get(r["source_name"], 0) > 0
+    )
+    breaches = {
+        key: n for key, n in buckets.items() if n > per_source_rpm[key[0]]
+    }
+    assert not breaches, f"RPM 击穿: {dict(list(breaches.items())[:5])}"
+
+
+def inv_rss_stable(samples_mb: list[float], *, max_growth_mb: float) -> None:
+    """不变量 4: 进程 RSS 首末差 < 阈值(窗口计数器/租约表无无界增长)。"""
+    assert len(samples_mb) >= 2, "RSS 采样不足"
+    growth = samples_mb[-1] - samples_mb[0]
+    assert growth < max_growth_mb, f"RSS 增长 {growth:.1f}MB ≥ 阈值 {max_growth_mb}MB"
+
+
+def structured_success_rate(rows: list[Row], *, session_suffix: str) -> float:
+    """不变量 5 的量: 指定场景行的最终成功率(error 为空即成功)。"""
+    scoped = [r for r in rows if str(r.get("session_id") or "").endswith(session_suffix)]
+    if not scoped:
+        return 0.0
+    ok = sum(1 for r in scoped if not r.get("error"))
+    return ok / len(scoped)
+
+
+async def inv_accounting_zeroed(limiter, sources: list[str]) -> None:
+    """不变量 1: 结束后全部源 inflight == 0(无泄漏租约)。活后端检查。"""
+    for name in sources:
+        stats = await limiter.source_stats(name)
+        assert stats.inflight == 0, f"源 {name} 结束后 inflight={stats.inflight}(租约泄漏)"
+
+
+# —— 报告 ——
+
+
+def _percentile(values: list[float], q: float) -> float:
+    if not values:
+        return 0.0
+    ordered = sorted(values)
+    return ordered[min(int(len(ordered) * q), len(ordered) - 1)]
+
+
+def render_report(run_id: str, rows: list[Row], verdicts: list[tuple[str, str]]) -> str:
+    """结构化 Markdown 报告(不变量 6: 成本/延迟基线产出)。"""
+    errors = Counter(str(r["error"]).split(":")[0] for r in rows if r.get("error"))
+    latencies = [float(r["latency_ms"]) for r in rows if not r.get("error")]
+    ttfts = [float(r["ttft_ms"]) for r in rows if r.get("ttft_ms") is not None]
+    costs = [float(r["cost"]) for r in rows if r.get("cost") is not None]
+    tokens = sum(int(r["prompt_tokens"]) + int(r["completion_tokens"]) for r in rows)
+    cache_hits = sum(1 for r in rows if r.get("cache_hit"))
+    lines = [
+        f"# Soak 报告: {run_id}",
+        "",
+        "## 不变量裁决",
+        *(f"- {name}: {verdict}" for name, verdict in verdicts),
+        "",
+        "## 规模",
+        f"- 遥测行数: {len(rows)};缓存命中: {cache_hits};总 token: {tokens}",
+        f"- 错误分布: {dict(errors) or '无'}",
+        "",
+        "## 延迟基线(成功行)",
+        f"- 总时长 p50/p95: {_percentile(latencies, 0.5):.0f} / {_percentile(latencies, 0.95):.0f} ms",
+        f"- TTFT p50/p95: {_percentile(ttfts, 0.5):.0f} / {_percentile(ttfts, 0.95):.0f} ms",
+        "",
+        "## 成本",
+        f"- 累计 cost: {sum(costs):.4f}(有单价行 {len(costs)}/{len(rows)})",
+    ]
+    return "\n".join(lines) + "\n"
+
+
+def write_report(run_id: str, content: str, out_dir: Path | str = "tests/outputs/soak") -> Path:
+    out = Path(out_dir)
+    out.mkdir(parents=True, exist_ok=True)
+    path = out / f"{run_id}.md"
+    path.write_text(content, encoding="utf-8")
+    return path