ops-tune-workbench-plugin 0.0.0-stage → 0.0.1-beta.2

package/skills/tool-ops-tune/ops-mechanism/scripts/case_backfill.py

@@ -0,0 +1,170 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+# ----------------------------------------------------------------------------------------------------------
+# Copyright (c) 2026 Huawei Technologies Co., Ltd.
+# This program is free software, you can redistribute it and/or modify it under the terms and conditions of
+# CANN Open Software License Agreement Version 2.0 (the "License").
+# Please refer to the License for details. You may not use this file except in compliance with the License.
+# THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
+# INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
+# See LICENSE in the root of the software repository for the full text of the License.
+# ----------------------------------------------------------------------------------------------------------
+"""case_backfill.py — E 级案例回灌草稿生成（REVIEW2 G9 / 反馈 P1-2 ②）
+
+设计对齐：
+  - io-contracts.md 已冻结 case_candidate.json 契约（opstune.case-candidate/1.0，
+    Step 5 产出，「经评审后入库」），但案例卡的入库一直是人工从 JSON 誊写
+    Markdown——格式漂移、警示列漏写、证据等级漏标都是人工誊写的常态错误。
+  - 本脚本把候选 JSON 转成 ops-mechanism/cases/ 案例卡 Markdown **草稿**：
+    固定字段全带上 + E 级标注 + 警示列模板 + 「评审后入库」水印——仍是草稿，
+    评审与合入是人工动作（与 calibrate.py 同款纪律：脚本不替代评审）。
+
+用法：
+  python3 case_backfill.py --input tuning_artifacts/gelu/case_candidate.json \
+      --cases-dir ../cases -o cases/ascendc/case_gelu_ub_fusion.draft.md
+  # 多条目输入（数组形态）也支持：逐条产出 draft
+
+输出草稿字段（对齐 ops-mechanism SKILL.md 案例卡 schema）：
+  算子类别 / DSL / 类型（PR|非PR） / 来源可信度（E 级） / 来源链接 + 验证日期 /
+  瓶颈证据 / 优化机制 / 原文性能数字 / 适用范围与警示
+"""
+
+from __future__ import annotations
+
+import argparse
+import datetime
+import json
+import sys
+from pathlib import Path
+
+
+def as_candidates(data) -> list[dict]:
+    """单条目 dict 或 entries 数组都接受。"""
+    if isinstance(data, list):
+        return data
+    if isinstance(data, dict):
+        if "entries" in data and isinstance(data["entries"], list):
+            return data["entries"]
+        return [data]
+    return []
+
+
+def scene_fp_text(fp: dict) -> str:
+    if not isinstance(fp, dict):
+        return str(fp)
+    keys = ("soc", "shape", "dtype", "format", "cann_version", "timing_method")
+    return " / ".join(f"{k}={fp[k]}" for k in keys if fp.get(k) is not None) or str(fp)
+
+
+def draft_md(c: dict, idx: int) -> str:
+    fp = c.get("scene_fingerprint", {})
+    measured = c.get("speedup_measured")
+    est = c.get("speedup_est")
+    est_text = (
+        f"[{est[0]}, {est[1]}]"
+        if isinstance(est, (list, tuple)) and len(est) == 2
+        else str(est)
+    )
+    rollback = c.get("rollback_happened")
+    lines = [
+        f"# 案例卡（草稿 {idx}）：{c.get('operator', '?')} — {c.get('mechanism', '?')}",
+        "",
+        "> ⚠️ **E 级回灌草稿——评审后入库**：本文件由 case_backfill.py 从 case_candidate.json",
+        "> 自动生成，字段按 ops-mechanism SKILL.md 案例卡 schema 组装；警示列与适用边界",
+        "> 为模板占位，评审时必须补全/核实后才可摘掉草稿水印。",
+        "",
+        "| 字段 | 值 |",
+        "|------|-----|",
+        f"| 算子类别 | {c.get('operator', '?')}（{c.get('dsl', 'Ascend C')}） |",
+        "| 类型 | 非 PR（Agent 实测回流） |",
+        "| 来源可信度 | **E 级**（本工作流 verdict→机制→实测闭环；speedup 只绑定当时口径指纹，不外推） |",
+        f"| 来源链接 + 验证日期 | case_candidate.json @ {c.get('generated_at', '?')}；"
+        f"生成于 {datetime.date.today().isoformat()} |",
+        f"| 瓶颈证据 | rule_id `{c.get('bound_rule', '?')}`（verdict 命中，"
+        f"口径指纹：{scene_fp_text(fp)}） |",
+        f"| 优化机制 | {c.get('mechanism', '?')} |",
+        f"| 原文性能数字 | speedup **{measured if measured is not None else '—'}**（实测，"
+        f"KEEP 轮；预估区间 {est_text}） |",
+        f"| 回滚 | {'**发生过**（见 notes）' if rollback else '无'} |",
+        "",
+        "## 适用范围与警示（评审时补全，占位模板）",
+        "",
+        "- [ ] 口径边界：上述 speedup 仅在「{soc} / {shape} / {dtype} / CANN {cann_version}」"
+        "下成立，**禁止外推**到其他 shape/dtype/soc；".format(
+            soc=fp.get("soc", "?"),
+            shape=fp.get("shape", "?"),
+            dtype=fp.get("dtype", "?"),
+            cann_version=fp.get("cann_version", "?"),
+        ),
+        "- [ ] 触发条件：命中 {rule} 时的哪个反证/severity 档位下适用本机制；".format(
+            rule=c.get("bound_rule", "?")
+        ),
+        "- [ ] 反例边界：何种场景该机制会劣化（回滚轮次的 notes 是线索）；",
+        f"- [ ] 原始备注：{c.get('notes', '（无）')}",
+        "",
+        "## 入库前检查单（ops-mechanism SKILL.md 检索规则）",
+        "",
+        "- [ ] 警示列已补全（反例与收益案例同等收录，警示列必填）",
+        "- [ ] 验证日期已核对",
+        "- [ ] 与既有同型案例（≤3）合并或分立已决策",
+        "- [ ] 摘除本草稿头部「草稿」水印",
+    ]
+    return "\n".join(lines) + "\n"
+
+
+def main() -> int:
+    ap = argparse.ArgumentParser(
+        description="case_candidate.json → E 级案例卡草稿（回灌 27-1 的半自动段）"
+    )
+    ap.add_argument(
+        "--input", required=True, help="case_candidate.json（单条目或 entries 数组）"
+    )
+    ap.add_argument(
+        "--cases-dir",
+        default=None,
+        help="ops-mechanism/cases/ 目录（提示用；草稿默认不入库，评审后人工移入）",
+    )
+    ap.add_argument(
+        "-o",
+        "--output",
+        required=True,
+        help="草稿输出路径（多条目时为前缀，逐条 .draft-N.md）",
+    )
+    args = ap.parse_args()
+
+    data = json.loads(Path(args.input).read_text(encoding="utf-8"))
+    cands = as_candidates(data)
+    if not cands:
+        print("ERROR: 输入无有效条目", file=sys.stderr)
+        return 1
+    schema = data.get("schema") if isinstance(data, dict) else None
+    if schema and schema != "opstune.case-candidate/1.0":
+        print(
+            f"WARN: schema {schema} != opstune.case-candidate/1.0，仍按字段尽力解析",
+            file=sys.stderr,
+        )
+
+    out = Path(args.output)
+    written = []
+    if len(cands) == 1:
+        out.parent.mkdir(parents=True, exist_ok=True)
+        out.write_text(draft_md(cands[0], 1), encoding="utf-8")
+        written.append(out)
+    else:
+        for i, c in enumerate(cands, 1):
+            p = out.parent / f"{out.stem}.draft-{i}.md"
+            p.parent.mkdir(parents=True, exist_ok=True)
+            p.write_text(draft_md(c, i), encoding="utf-8")
+            written.append(p)
+
+    for p in written:
+        print(f"draft → {p}")
+    print(
+        "下一步：人工评审（检查单在草稿尾部）→ 补全警示列 → 移入 cases/ 并摘水印"
+        + (f"（目标目录参考：{args.cases_dir}）" if args.cases_dir else "")
+    )
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())

package/skills/tool-ops-tune/ops-mechanism/scripts/log_gap.py

@@ -0,0 +1,145 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+# ----------------------------------------------------------------------------------------------------------
+# Copyright (c) 2026 Huawei Technologies Co., Ltd.
+# This program is free software, you can redistribute it and/or modify it under the terms and conditions of
+# CANN Open Software License Agreement Version 2.0 (the "License").
+# Please refer to the License for details. You may not use this file except in compliance with the License.
+# THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
+# INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
+# See LICENSE in the root of the software repository for the full text of the License.
+# ----------------------------------------------------------------------------------------------------------
+"""log_gap.py — ops-family 知识缺口计数写入（REVIEW2 G9 / 反馈 P1-2 ③，提前骑行项）
+
+设计对齐：
+  - ops-family SKILL.md 承诺「未收录族返回并**计数上报**（知识缺口感知回路的输入，
+    落盘 references/knowledge_gaps.json）」——此前该文件六条缺口 count 全 0，
+    无任何脚本写入：计数是纸面承诺（反馈 §4：「log_gap.py 提前——0.5 人日，
+    让 knowledge_gaps 计数从 M1 起开始积累，否则 P1-7 的季度排产承诺持续空转」）。
+  - 本脚本就是写入载体：路由时发现族未收录 / 族内文档为骨架 → 本条命令落一条
+    gap 记录（count+1 / first_seen 登记 / note 追加），季度排产（27-4）读它。
+
+用法：
+  python3 log_gap.py --family fa --kind family_missing [--note "本次场景：FA 混合绑核"]
+  python3 log_gap.py --list
+
+gap 条目语义（与 knowledge_gaps.json 现有 schema 一致）：
+  family    算子族名（fa/elementwise/sort/simt/matmul/mc2/...）
+  kind      family_missing | guide_missing | guide_skeleton | no_shelf_code
+            （未知 kind 允许写入——新类型由使用者定义，list 输出如实呈现）
+  count     命中次数（排产权重）
+  first_seen 首次命中日期
+  last_seen 最近命中日期（本脚本新增字段；旧条目首次追加时补写）
+  notes[]   每次命中的场景备注（增量追加，不覆盖）
+"""
+
+from __future__ import annotations
+
+import argparse
+import datetime
+import json
+import sys
+from pathlib import Path
+
+# knowledge_gaps.json 归 ops-family（族表 SKILL.md 承诺的落盘位置）；脚本放
+# ops-mechanism/scripts 是因为「路由时发现未收录」发生在机制路由阶段（Step 3
+# 查模板路径），ops-mechanism 是调用方。parents[2] = 套件根（tool-ops-tune/）。
+GAPS_PATH = (
+    Path(__file__).resolve().parents[2]
+    / "ops-family"
+    / "references"
+    / "knowledge_gaps.json"
+)
+
+KNOWN_KINDS = ("family_missing", "guide_missing", "guide_skeleton", "no_shelf_code")
+
+
+def load_gaps() -> dict:
+    if not GAPS_PATH.exists():
+        return {
+            "schema": "opstune.knowledge-gaps/1.0",
+            "description": "ops-family 知识缺口计数落盘",
+            "gaps": [],
+        }
+    return json.loads(GAPS_PATH.read_text(encoding="utf-8"))
+
+
+def save_gaps(gaps: dict) -> None:
+    GAPS_PATH.parent.mkdir(parents=True, exist_ok=True)
+    GAPS_PATH.write_text(
+        json.dumps(gaps, ensure_ascii=False, indent=2) + "\n", encoding="utf-8"
+    )
+
+
+def main() -> int:
+    ap = argparse.ArgumentParser(
+        description="知识缺口计数写入（知识缺口感知回路 27-4 的输入）"
+    )
+    ap.add_argument("--family", help="算子族名（如 fa / elementwise）")
+    ap.add_argument(
+        "--kind",
+        help="缺口类型（family_missing/guide_missing/guide_skeleton/no_shelf_code，"
+        "或自定义新类型）",
+    )
+    ap.add_argument("--note", default=None, help="本次命中场景备注（追加进 notes）")
+    ap.add_argument("--list", action="store_true", help="列出全部缺口与计数")
+    args = ap.parse_args()
+
+    gaps = load_gaps()
+
+    if args.list or (not args.family and not args.kind):
+        if not gaps.get("gaps"):
+            print("（无缺口记录）")
+            return 0
+        print(f"{'family':<14} {'kind':<16} {'count':>5}  first_seen   last_seen")
+        for g in sorted(gaps["gaps"], key=lambda g: -g.get("count", 0)):
+            print(
+                f"{g['family']:<14} {g['kind']:<16} {g.get('count', 0):>5}"
+                f"  {g.get('first_seen', '-'):<11} {g.get('last_seen', '-'):<10}"
+            )
+        return 0
+
+    if not args.family or not args.kind:
+        print("ERROR: 需要 --family 与 --kind（或 --list）", file=sys.stderr)
+        return 1
+    if args.kind not in KNOWN_KINDS:
+        # 不拒绝：新类型是正常演进，但 stderr 提示核对（防拼写错误把计数写进错误桶）
+        print(
+            f"WARN: kind {args.kind!r} 不在已知四类 {KNOWN_KINDS}——若是新类型忽略本提示；"
+            f"若是笔误请改正后重跑（写错桶会稀释排产信号）",
+            file=sys.stderr,
+        )
+
+    today = datetime.date.today().isoformat()
+    entry = next(
+        (
+            g
+            for g in gaps["gaps"]
+            if g["family"] == args.family and g["kind"] == args.kind
+        ),
+        None,
+    )
+    if entry is None:
+        entry = {
+            "family": args.family,
+            "kind": args.kind,
+            "note": "",
+            "first_seen": today,
+            "count": 0,
+            "notes": [],
+        }
+        gaps["gaps"].append(entry)
+        print(f"[new] {args.family}/{args.kind} 条目登记")
+    else:
+        print(f"[hit] {args.family}/{args.kind}")
+    entry["count"] = entry.get("count", 0) + 1
+    entry["last_seen"] = today
+    if args.note:
+        entry.setdefault("notes", []).append(f"{today} {args.note}")
+    save_gaps(gaps)
+    print(f"count={entry['count']}  → {GAPS_PATH}")
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())

package/skills/tool-ops-tune/ops-mechanism/scripts/mc2_search.py

@@ -0,0 +1,367 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+# ----------------------------------------------------------------------------------------------------------
+# Copyright (c) 2026 Huawei Technologies Co., Ltd.
+# This program is free software, you can redistribute it and/or modify it under the terms and conditions of
+# CANN Open Software License Agreement Version 2.0 (the "License").
+# Please refer to the License for details. You may not use this file except in compliance with the License.
+# THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
+# INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
+# See LICENSE in the root of the software repository for the full text of the License.
+# ----------------------------------------------------------------------------------------------------------
+"""mc2_search.py — MC² 配平系统搜索器（REVIEW2 G6-M2 / 反馈 P1-4）
+
+设计对齐：
+  - mc2-balancing.md 机制卡写明「统一搜索算法（禁止经验值交付）」：四剪枝 +
+    Top-N 输出——此前只有 markdown 描述，全仓无执行载体（REVIEW2 G6 定性）。
+  - 本脚本是执行载体：输入 TilingData + R（隔离测试产出）→ 长短块排布候选
+    搜索 → 四剪枝 → Top-N 候选 + 剪枝记录（配平方案卡必须附本输出）。
+
+搜索空间（机制卡原文语义）：
+  长块 M ∈ 候选集（baseM 合法对齐粒度的倍数序列）
+  短块 M = 长块 M / ρ（对齐修正），ρ = max(R, 1/R)（长块与短块 M 之比）
+  块数切分：长块数 + 短块数在总 M（矩阵行数）预算内排布
+
+四剪枝（可行性，机制卡编号）：
+  ① 核利用率：usedCoreNum × 块数覆盖不下降（候选的覆盖块数 ≥ 当前方案的）
+  ② L1/L0 容量：baseM/baseN/baseK 组合不超片上预算（soc-spec.yaml 容量表）
+  ③ 膨胀下界：计算膨胀 + 通信膨胀估算不超过收益（短块填充带来的冗余开销）
+  ④ 对齐约束：M/N/K 对齐要求（512B 地址对齐 → 分形对齐 16/32 粒度）
+
+输出（opstune.mc2-search/1.0）：
+  candidates[]：Top-N（--top，默认 5）——每候选带 block 布局、容量占用计算
+                过程、剪枝检查通过记录
+  pruned[]：被剪枝候选 + 剪枝原因（可审计证据：剪枝记录本身是证据）
+  meta：R / ρ / 搜索空间统计 / soc 容量参数溯源
+
+用法：
+  python3 mc2_search.py --m 4096 --base-n 256 --base-k 256 --used-core-num 16 \
+      --r-value 2.4 --soc a5 [-o candidates.json]
+  # R 来自 isolation_test.py 的 r_value；TilingData 来自 host 侧打印（机制卡 MC-1）
+"""
+
+from __future__ import annotations
+
+import argparse
+import datetime
+import json
+import sys
+from pathlib import Path
+
+SCHEMA = "opstune.mc2-search/1.0"
+SOC_SPEC_FILE = "soc-spec" + ".yaml"  # 判定引擎同源峰值标尺
+DEFAULT_SOC_SPEC = (
+    Path(__file__).resolve().parents[2] / "ops-verdict" / "references" / SOC_SPEC_FILE
+)
+
+# 对齐粒度（机制卡剪枝④：分形对齐 16 / 512B 地址对齐折算元素粒度）
+FRACTAL_ALIGN = 16  # M/N/K 分形对齐粒度（Cube MAC 16×16×16）
+ELEM_ALIGN_512B = {  # 512B 地址对齐折算的最小元素数（dtype 元素大小 → 元素数）
+    "fp16": 256,
+    "bf16": 256,
+    "fp32": 128,
+    "int8": 512,
+}
+# 512B 对齐是搬运效率推荐态而非硬约束（DataCopyPad 场景允许非 512B 对齐；
+# 分形对齐 16 才是 Cube 的硬约束）——剪枝④按此分两级：分形不齐 → 剪枝；
+# 512B 不齐 → 记录 warn 不剪枝（候选照常输出，标注 warn_align_512b）
+
+
+def load_soc_caps(soc: str, soc_spec_path: Path) -> dict:
+    """从 soc-spec.yaml 取容量约束（剪枝②的真源，不硬编码）。"""
+    import yaml
+
+    spec = yaml.safe_load(Path(soc_spec_path).read_text(encoding="utf-8"))
+    block = spec.get("socs", {}).get(soc)
+    if not block:
+        raise SystemExit(f"ERROR: soc-spec 无 {soc} 条目（{soc_spec_path}）")
+    buf = block.get("buffer", {})
+    ref = block.get("reference", {})
+    return {
+        "l1_kb": buf.get("l1_kb"),
+        "l0a_kb": buf.get("l0a_kb"),
+        "l0b_kb": buf.get("l0b_kb"),
+        "l0c_kb": buf.get("l0c_kb"),
+        "ub_kb": buf.get("ub_kb"),
+        "core_num": ref.get("ai_core_cnt", {}).get("value"),
+        "source": f"soc-spec:{soc}",
+    }
+
+
+def align_up(v: int, grain: int) -> int:
+    return ((v + grain - 1) // grain) * grain
+
+
+def long_block_candidates(
+    m: int, base_m_min: int, base_m_max: int, dtype: str, step_grain: int
+) -> list[int]:
+    """长块 M 候选集：[base_m_min, min(base_m_max, m)] 内按 step_grain 步进的
+    对齐值 + m 本身（整除免短块收尾）。"""
+    align = max(FRACTAL_ALIGN, ELEM_ALIGN_512B.get(dtype, FRACTAL_ALIGN))
+    hi = min(base_m_max, align_up(m, align))
+    cands = set()
+    v = align_up(base_m_min, align)
+    while v <= hi:
+        cands.add(v)
+        v = align_up(v + step_grain, align)
+    cands.add(hi)
+    return sorted(cands)
+
+
+def prune(candidate: dict, ctx: dict) -> tuple[list[str], list[str]]:
+    """四剪枝：返回 (剪枝原因, 警告) 两列表（原因空 = 通过全部剪枝）。"""
+    reasons, warns = [], []
+    # ④ 对齐约束：分形对齐 16 是 Cube 硬约束（剪枝）；512B 对齐是搬运推荐态（警告）
+    for k in ("long_m", "short_m"):
+        if candidate[k] % FRACTAL_ALIGN != 0:
+            reasons.append(f"④ {k}={candidate[k]} 不满足分形对齐 {FRACTAL_ALIGN}")
+    if candidate["long_m"] % ctx["elem_align_512b"] != 0:
+        warns.append(
+            f"长块 M={candidate['long_m']} 非 512B 对齐粒度"
+            f"（{ctx['elem_align_512b']} 元素）——搬运效率推荐态，DataCopyPad 场景可接受"
+        )
+    if candidate["short_m"] % ctx["elem_align_512b"] != 0:
+        warns.append(
+            f"短块 M={candidate['short_m']} 非 512B 对齐粒度——同上，记录不剪枝"
+        )
+    if ctx["base_n"] % FRACTAL_ALIGN != 0 or ctx["base_k"] % FRACTAL_ALIGN != 0:
+        reasons.append("④ baseN/baseK 未分形对齐（输入参数问题）")
+
+    # ② L1/L0 容量：A/B tile + 双 buffer 不超预算（元素 × dtype 字节 × buffer 份数）
+    elem_bytes = ctx["elem_bytes"]
+    buf_num = ctx.get("l1_buffer_num", 2)
+    a_kb = candidate["long_m"] * ctx["base_k"] * elem_bytes / 1024
+    b_kb = ctx["base_n"] * ctx["base_k"] * elem_bytes / 1024
+    l1_use_kb = (a_kb + b_kb) * buf_num
+    if l1_use_kb > ctx["l1_kb"]:
+        reasons.append(
+            f"② L1 超预算：({a_kb:.0f}+{b_kb:.0f})×{buf_num} = {l1_use_kb:.0f}KB"
+            f" > {ctx['l1_kb']}KB"
+        )
+    l0a_use_kb = candidate["long_m"] * ctx["base_k"] * elem_bytes / 1024
+    l0b_use_kb = ctx["base_n"] * ctx["base_k"] * elem_bytes / 1024
+    if l0a_use_kb > ctx["l0a_kb"]:
+        reasons.append(f"② L0A 超预算：{l0a_use_kb:.0f}KB > {ctx['l0a_kb']}KB")
+    if l0b_use_kb > ctx["l0b_kb"]:
+        reasons.append(f"② L0B 超预算：{l0b_use_kb:.0f}KB > {ctx['l0b_kb']}KB")
+    l0c_use_kb = candidate["long_m"] * ctx["base_n"] * elem_bytes / 1024
+    if l0c_use_kb > ctx["l0c_kb"]:
+        reasons.append(f"② L0C 超预算：{l0c_use_kb:.0f}KB > {ctx['l0c_kb']}KB")
+
+    # ① 核利用率：候选覆盖块数（长+短块要占满核）不下降
+    if candidate["blocks_total"] < ctx["current_blocks"]:
+        reasons.append(
+            f"① 核覆盖下降：候选块数 {candidate['blocks_total']}"
+            f" < 当前 {ctx['current_blocks']}"
+        )
+    if candidate["blocks_total"] < ctx["used_core_num"]:
+        reasons.append(
+            f"① 块数 < usedCoreNum（{candidate['blocks_total']}"
+            f" < {ctx['used_core_num']}）：有空闲核"
+        )
+
+    # ③ 膨胀下界：短块膨胀（补齐对齐引入的冗余计算）+ 排布浪费 ≤ 掩盖收益
+    # 收益上界（机制卡：配平收益 ≤ min 侧掩盖，估为 (ρ-1)/ρ × T_max 占比）
+    # 膨胀 = 冗余元素 / 总元素；排布浪费 = 尾部空隙 / 总 M
+    inflate_pct = candidate.get("inflation_pct", 0.0) + candidate.get(
+        "tail_waste_pct", 0.0
+    )
+    if inflate_pct > ctx["max_inflation_pct"]:
+        reasons.append(
+            f"③ 膨胀 {inflate_pct:.1f}% > 上限 {ctx['max_inflation_pct']}%"
+            f"（对齐冗余 {candidate.get('inflation_pct', 0):.1f}% + "
+            f"尾部浪费 {candidate.get('tail_waste_pct', 0):.1f}%）"
+        )
+    return reasons, warns
+
+
+def build_candidate(long_m: int, rho: float, m: int, ctx: dict) -> dict:
+    # 短块 M = 长块 M / ρ（对齐修正向上）
+    short_m = align_up(max(1, round(long_m / rho)), FRACTAL_ALIGN)
+    # 排布：R>1（计算慢）→ 长块主导；R<1（通信慢）→ 短块多塞间隙。
+    # 简化排布（机制卡「实际排布按 Top-N 实测修正」）：长块铺满为主，
+    # 末段用短块收尾降低尾部浪费。
+    long_blocks = m // long_m
+    remain = m - long_blocks * long_m
+    short_blocks = 0
+    if remain > 0:
+        if remain % short_m == 0 and remain // short_m > 0:
+            short_blocks = remain // short_m
+        else:
+            # 短块收不齐：短块数 = ceil，末块对齐膨胀记入 inflation
+            short_blocks = -(-remain // short_m)
+    blocks_total = long_blocks + short_blocks
+    covered = long_blocks * long_m + abs(short_blocks) * short_m
+    inflation_pct = max(0.0, (covered - m) / m * 100) if m > 0 else 0.0
+    tail_waste_pct = max(0.0, (covered - m) / m * 100) if m > 0 else 0.0
+    return {
+        "variant": f"long{long_m}_short{short_m}",
+        "long_m": long_m,
+        "short_m": short_m,
+        "rho": round(rho, 3),
+        "long_blocks": long_blocks,
+        "short_blocks": short_blocks,
+        "blocks_total": blocks_total,
+        "inflation_pct": round(inflation_pct, 2),
+        "tail_waste_pct": round(tail_waste_pct, 2),
+        "capacity_kb": {
+            "l1": round(
+                (long_m + ctx["base_n"])
+                * ctx["base_k"]
+                * ctx["elem_bytes"]
+                * ctx.get("l1_buffer_num", 2)
+                / 1024,
+                1,
+            ),
+            "l0a": round(long_m * ctx["base_k"] * ctx["elem_bytes"] / 1024, 1),
+            "l0b": round(ctx["base_n"] * ctx["base_k"] * ctx["elem_bytes"] / 1024, 1),
+            "l0c": round(long_m * ctx["base_n"] * ctx["elem_bytes"] / 1024, 1),
+        },
+    }
+
+
+def main() -> int:
+    ap = argparse.ArgumentParser(
+        description="MC² 配平系统搜索器（四剪枝 + Top-N，禁经验值交付）"
+    )
+    ap.add_argument("--m", type=int, required=True, help="矩阵 M（总行数，TilingData）")
+    ap.add_argument("--base-n", type=int, required=True, help="baseN（TilingData）")
+    ap.add_argument("--base-k", type=int, required=True, help="baseK（TilingData）")
+    ap.add_argument(
+        "--used-core-num",
+        type=int,
+        required=True,
+        help="usedCoreNum（TilingData；机制卡 MC-1 硬约束）",
+    )
+    ap.add_argument(
+        "--r-value",
+        type=float,
+        required=True,
+        help="隔离测试 R = T_compute/T_comm（isolation_test.py 的 r_value）",
+    )
+    ap.add_argument("--soc", default="a5", help="a2/a3/a5（容量约束读 soc-spec）")
+    ap.add_argument("--dtype", default="fp16", choices=["fp16", "bf16", "fp32", "int8"])
+    ap.add_argument("--base-m-min", type=int, default=64, help="长块 M 搜索下界")
+    ap.add_argument("--base-m-max", type=int, default=512, help="长块 M 搜索上界")
+    ap.add_argument("--step", type=int, default=32, help="搜索步进（粒度）")
+    ap.add_argument(
+        "--l1-buffer-num", type=int, default=2, help="L1 buffer 份数（pingpong=2）"
+    )
+    ap.add_argument(
+        "--max-inflation-pct",
+        type=float,
+        default=15.0,
+        help="剪枝③膨胀上限 %%（默认 15，🟡 工程初值）",
+    )
+    ap.add_argument(
+        "--current-blocks",
+        type=int,
+        default=None,
+        help="当前方案块数（剪枝①基准；缺省取 used-core-num）",
+    )
+    ap.add_argument("--top", type=int, default=5, help="Top-N 输出")
+    ap.add_argument("--soc-spec", default=str(DEFAULT_SOC_SPEC))
+    ap.add_argument("-o", "--output", default=None, help="candidates.json 输出")
+    args = ap.parse_args()
+
+    caps = load_soc_caps(args.soc, Path(args.soc_spec))
+    rho = max(args.r_value, 1.0 / args.r_value) if args.r_value > 0 else 1.0
+    ctx = {
+        "base_n": args.base_n,
+        "base_k": args.base_k,
+        "used_core_num": args.used_core_num,
+        "current_blocks": args.current_blocks
+        if args.current_blocks is not None
+        else args.used_core_num,
+        "elem_bytes": {"fp16": 2, "bf16": 2, "fp32": 4, "int8": 1}[args.dtype],
+        "elem_align": max(FRACTAL_ALIGN, ELEM_ALIGN_512B[args.dtype]),
+        "elem_align_512b": ELEM_ALIGN_512B[args.dtype],
+        "l1_kb": caps["l1_kb"],
+        "l0a_kb": caps["l0a_kb"],
+        "l0b_kb": caps["l0b_kb"],
+        "l0c_kb": caps["l0c_kb"],
+        "l1_buffer_num": args.l1_buffer_num,
+        "max_inflation_pct": args.max_inflation_pct,
+    }
+
+    # 搜索 + 剪枝
+    candidates, pruned = [], []
+    space = long_block_candidates(
+        args.m, args.base_m_min, args.base_m_max, args.dtype, args.step
+    )
+    for lm in space:
+        c = build_candidate(lm, rho, args.m, ctx)
+        reasons, warns = prune(c, ctx)
+        if warns:
+            c["warn_align_512b"] = warns
+        if reasons:
+            pruned.append({"variant": c["variant"], "reasons": reasons})
+        else:
+            candidates.append(c)
+
+    # 排序取 Top-N：块数多（核覆盖好）优先，其次膨胀小（机制卡：估算 ≠ 实测，
+    # Top-N 供逐个实测——排序只是实测顺序，不是最终结论）
+    candidates.sort(
+        key=lambda c: (-c["blocks_total"], c["inflation_pct"] + c["tail_waste_pct"])
+    )
+    top = candidates[: args.top]
+
+    result = {
+        "schema": SCHEMA,
+        "generated_at": datetime.datetime.now().isoformat(timespec="seconds"),
+        "soc": args.soc,
+        "soc_caps": {
+            k: caps[k] for k in ("l1_kb", "l0a_kb", "l0b_kb", "l0c_kb", "core_num")
+        },
+        "soc_caps_source": caps["source"],
+        "input": {
+            "m": args.m,
+            "base_n": args.base_n,
+            "base_k": args.base_k,
+            "used_core_num": args.used_core_num,
+            "r_value": args.r_value,
+            "dtype": args.dtype,
+            "rho": round(rho, 3),
+        },
+        "search_space": {
+            "long_m_range": [min(space), max(space)] if space else [],
+            "n_explored": len(space),
+        },
+        "candidates": top,
+        "pruned": pruned,
+        "notes": [
+            "排序 = 实测顺序（块数优先、膨胀次之），不是最终结论——每候选须逐个实测"
+            "（tileCnt 扫描、跨 rank max 口径）后择优（mc2-balancing.md）",
+            f"剪枝③膨胀上限 {args.max_inflation_pct}% 为 🟡 工程初值，实测校准后回写",
+        ],
+        "next": "逐候选实测同口径（msprof task-based、同 warmup/loops、跨 rank max），"
+        "delta_compare 出三段结论；配平方案卡必须附本输出（含剪枝记录）",
+    }
+    if args.output:
+        out = Path(args.output)
+        out.parent.mkdir(parents=True, exist_ok=True)
+        out.write_text(
+            json.dumps(result, indent=2, ensure_ascii=False) + "\n", encoding="utf-8"
+        )
+        print(f"→ {out}")
+    print(
+        f"MC² 配平搜索：R={args.r_value} ρ={rho:.2f} → 探索 {len(space)} / "
+        f"通过 {len(candidates)} / 剪枝 {len(pruned)}，Top-{len(top)}："
+    )
+    for c in top:
+        print(
+            f"  {c['variant']}: 长块×{c['long_blocks']} + 短块×{c['short_blocks']}"
+            f" = {c['blocks_total']} 块，膨胀 {c['inflation_pct'] + c['tail_waste_pct']:.1f}%，"
+            f"L1 {c['capacity_kb']['l1']}KB"
+        )
+    if not top:
+        print(
+            "  （无候选通过四剪枝——检查 TilingData/R 输入或放宽 --max-inflation-pct）",
+            file=sys.stderr,
+        )
+        return 1
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())