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

package/skills/tool-ops-tune/ops-verdict/rules/L1-single-core.yaml

@@ -0,0 +1,316 @@
+# L1 单核层判定规则包：核内 pipe 级 Bound
+#
+# 阈值来源（知识契约：每条阈值可溯源，变更须过 evals/run_evals.py 回归）：
+#   [msprof-doc]        msprof 官方文档与 CSV 字段说明（MindStudio 8.3 口径）
+#   [tuning-practice]   一线调优实践沉淀（阈值实测来源，标定于 A2/A5 真机案例）
+schema: opstune.rules/1.0
+layer: L1
+title: 单核流水判定（核内 pipe 级）
+soc: [a2, a3, a5]
+
+# ---------------------------------------------------------------------------
+# 指标词汇表：规则-指标绑定契约的解析层。
+# columns 为候选列名（按序取首个可解析者，兼容 A5 aic_/aiv_ 双列名与旧版百分数格式）。
+# normalize: ratio_to_pct — 0~1.5 视为比例 ×100（仅限 ratio 类指标，时间/计数类禁止）。
+# group: pipe — 参与主导 pipe 排序（rank 聚合）与串行指纹求和。
+# source: scalar — 由调用方经 --scalar 提供（如带宽占峰值百分比），缺失时反证不可评估。
+# ---------------------------------------------------------------------------
+metrics:
+  vec_ratio:
+    source: pipe
+    columns: [aiv_vec_ratio, aic_vec_ratio, "VectorUtilization(%)", vec_ratio, VECTOR(%)]
+    normalize: ratio_to_pct
+    group: pipe
+  cube_ratio:
+    source: pipe
+    columns: [aic_cube_ratio, aiv_cube_ratio, "CubeUtilization(%)", cube_ratio, CUBE(%)]
+    normalize: ratio_to_pct
+    group: pipe
+  mte2_ratio:
+    source: pipe
+    columns: [aic_mte2_ratio, aiv_mte2_ratio, ai_mte2_ratio, "Mte2Utilization(%)", mte2_ratio, MTE2(%)]
+    normalize: ratio_to_pct
+    group: pipe
+  mte3_ratio:
+    source: pipe
+    columns: [aic_mte3_ratio, aiv_mte3_ratio, ai_mte3_ratio, "Mte3Utilization(%)", mte3_ratio, MTE3(%)]
+    normalize: ratio_to_pct
+    group: pipe
+  mte1_ratio:
+    source: pipe
+    columns: [aic_mte1_ratio, aiv_mte1_ratio, "Mte1Utilization(%)", mte1_ratio, MTE1(%)]
+    normalize: ratio_to_pct
+    group: pipe
+  fixpipe_ratio:
+    source: pipe
+    columns: [aic_fixpipe_ratio, aiv_fixpipe_ratio, "FixpipeUtilization(%)", fixp_ratio, FIXP(%)]
+    normalize: ratio_to_pct
+    group: pipe
+  scalar_ratio:
+    source: pipe
+    columns: [aic_scalar_ratio, aiv_scalar_ratio, ai_scalar_ratio, "ScalarUtilization(%)", scalar_ratio, SCALAR(%)]
+    normalize: ratio_to_pct
+    group: pipe
+  mte2_active_bw_pct_of_peak:
+    source: scalar
+    description: MTE2 活跃带宽占理论峰值百分比（调用方按目标 SoC 峰值换算后提供）
+  vec_total_cflt_ratio:
+    source: scalar
+    description: Vector 指令 Bank conflict 比例（ResourceConflictRatio.csv ai*_vec_total_cflt_ratio，调用方换算后提供）
+  vec_fp32_ratio:
+    source: scalar
+    description: Vector 中 fp32 指令占比（ArithmeticUtilization.csv，调用方换算后提供）
+  fixpipe_align_hint:
+    source: scalar
+    description: 保留位（未使用）
+
+rules:
+  # ---------------------------------------------------------------- MTE2 搬入
+  - id: L1_MTE2_BOUND
+    name: MTE2 搬入瓶颈
+    priority: 10
+    parent: L1_MEMORY_SATURATION
+    threshold_source: "[msprof-doc] >80% 直接判 / 最大且 >70%；watch 50% 档取 [tuning-practice]"
+    metrics_required: [mte2_ratio]
+    metrics_optional: [mte2_active_bw_pct_of_peak, fixpipe_ratio]
+    condition:
+      any:
+        - {metric: mte2_ratio, agg: mean, op: gt, value: 80}
+        - all:
+            - {metric: mte2_ratio, agg: mean, op: gt, value: 70}
+            - {metric: mte2_ratio, agg: rank, op: eq, value: 1}
+    watch:
+      condition: {metric: mte2_ratio, agg: mean, op: gt, value: 50}
+      note: MTE2 占比 50~70%，暂非主导，已具备搬运优化价值
+    severity:
+      metric: mte2_ratio
+      bands:
+        - {min: 85, level: critical, note: 接近极限，单机制收益有限，需数据量/复用级优化}
+        - {min: 70, level: moderate}
+      default: light
+    counter_evidence:
+      - id: true-bandwidth-saturation
+        condition: {metric: mte2_active_bw_pct_of_peak, agg: value, op: gt, value: 80}
+        note: active_bw 已达理论 80%+：真带宽饱和。优化在数据量削减/复用，不在搬运组织
+        action: annotate
+      - id: fake-mte2-small-packet
+        condition: {metric: mte2_active_bw_pct_of_peak, agg: value, op: lt, value: 40}
+        note: busy 高但带宽 <40% 峰值：疑似小包/非对齐/频繁同步（假 MTE2 bound）
+        action: annotate
+      - id: fixpipe-misalignment
+        condition: {metric: fixpipe_ratio, agg: mean, op: gt, value: 15}
+        note: fixpipe >15%：GM 地址未 512B 对齐问题叠加 [tuning-practice]
+        action: annotate
+    confidence: high
+    speedup_range: [1.2, 2.0]
+    speedup_source: "[tuning-practice] MTE2 优化方法预期收益，🟡 首期标定值"
+    mechanism_refs:
+      - ops-mechanism:references/optimize/data-copy.md
+      - ops-mechanism:references/optimize/memory-hierarchy.md
+
+  # ---------------------------------------------------------------- CUBE 计算
+  - id: L1_CUBE_BOUND
+    name: Cube 矩阵计算瓶颈
+    priority: 20
+    parent: L1_COMPUTE_SATURATION
+    threshold_source: "[msprof-doc]"
+    metrics_required: [cube_ratio]
+    metrics_optional: []
+    condition:
+      any:
+        - {metric: cube_ratio, agg: mean, op: gt, value: 80}
+        - all:
+            - {metric: cube_ratio, agg: mean, op: gt, value: 70}
+            - {metric: cube_ratio, agg: rank, op: eq, value: 1}
+    watch:
+      condition: {metric: cube_ratio, agg: mean, op: gt, value: 40}
+      note: Cube 占比 >40%（MatMul 类正常水位 [msprof-doc]）
+    severity:
+      metric: cube_ratio
+      bands:
+        - {min: 85, level: critical, note: 接近理论极限，先核 fops 折算的真实利用率}
+        - {min: 70, level: moderate}
+      default: light
+    counter_evidence:
+      - id: check-real-flops
+        condition: {metric: cube_ratio, agg: mean, op: gt, value: 85}
+        note: busy 接近饱和时必须用 aic_cube_fops 折算理论利用率，排除空转指令 [tuning-practice]
+        action: annotate
+    confidence: high
+    speedup_range: [1.1, 1.5]
+    speedup_source: 🟡 首期标定值
+    mechanism_refs:
+      - ops-mechanism:references/optimize/ascendc-tiling.md
+      - ops-family:references/matmul/guide.md
+
+  # ---------------------------------------------------------------- VEC 向量
+  - id: L1_VEC_BOUND
+    name: Vector 向量计算瓶颈
+    priority: 30
+    parent: L1_COMPUTE_SATURATION
+    threshold_source: "[msprof-doc] 主判定表；[tuning-practice] 50-65/65-80/>80 三档"
+    metrics_required: [vec_ratio]
+    metrics_optional: []
+    condition:
+      any:
+        - {metric: vec_ratio, agg: mean, op: gt, value: 80}
+        - all:
+            - {metric: vec_ratio, agg: mean, op: gt, value: 70}
+            - {metric: vec_ratio, agg: rank, op: eq, value: 1}
+    watch:
+      condition: {metric: vec_ratio, agg: mean, op: gt, value: 50}
+      note: VEC 占比 50~70%，UB 融合/减 Cast 已有收益空间 [tuning-practice]
+    severity:
+      metric: vec_ratio
+      bands:
+        - {min: 80, level: critical, note: VEC 接近理论极限，优化空间有限，转数据量/算法侧}
+        - {min: 65, level: moderate, note: 需减少 Cast 或融合指令}
+        - {min: 50, level: light, note: DoubleBuffer + UB 融合有较大收益}
+      default: light
+    counter_evidence:
+      - id: bank-conflict-amplify
+        condition: {metric: vec_total_cflt_ratio, agg: mean, op: gt, value: 5}
+        note: Bank conflict 放大了 VEC 耗时，先解冲突再谈向量化 [tuning-practice 交叉关联诊断]
+        action: annotate
+      - id: cast-heavy
+        condition: {metric: vec_fp32_ratio, agg: mean, op: gt, value: 60}
+        note: fp32 指令占比高：存在大量 fp16→fp32 转换，减 Cast 优先 [tuning-practice]
+        action: annotate
+    confidence: high
+    speedup_range: [1.1, 1.5]
+    speedup_source: "[tuning-practice] UB 融合 20-50% / 减 Cast 10-30%（档位内取保守下界），🟡 待校准"
+    mechanism_refs:
+      - ops-mechanism:references/optimize/api-usage.md
+      - ops-mechanism:references/optimize/pipeline.md
+      - ops-mechanism:references/optimize/ascendc-regbase-vf.md
+
+  # ---------------------------------------------------------------- MTE3 搬出
+  - id: L1_MTE3_BOUND
+    name: MTE3 搬出瓶颈
+    priority: 40
+    parent: L1_MEMORY_SATURATION
+    threshold_source: "[msprof-doc]"
+    metrics_required: [mte3_ratio]
+    metrics_optional: []
+    condition:
+      any:
+        - {metric: mte3_ratio, agg: mean, op: gt, value: 80}
+        - all:
+            - {metric: mte3_ratio, agg: mean, op: gt, value: 70}
+            - {metric: mte3_ratio, agg: rank, op: eq, value: 1}
+    severity:
+      metric: mte3_ratio
+      bands:
+        - {min: 85, level: critical}
+        - {min: 70, level: moderate}
+      default: light
+    confidence: high
+    speedup_range: [1.1, 1.4]
+    speedup_source: 🟡 首期标定值
+    mechanism_refs:
+      - ops-mechanism:references/optimize/data-copy.md
+
+  # ---------------------------------------------------------------- FIXP
+  - id: L1_FIXP_BOUND
+    name: FixPipe 输出瓶颈
+    priority: 50
+    parent: L1_MEMORY_SATURATION
+    threshold_source: "[msprof-doc] >80% 判 bound；>15% 对齐预警取 [tuning-practice]"
+    metrics_required: [fixpipe_ratio]
+    metrics_optional: []
+    condition: {metric: fixpipe_ratio, agg: mean, op: gt, value: 80}
+    watch:
+      condition: {metric: fixpipe_ratio, agg: mean, op: gt, value: 15}
+      note: fixpipe >15%：GM 地址未 512B 对齐的直接信号，修复成本低收益显著 [tuning-practice FA 案例]
+    severity:
+      metric: fixpipe_ratio
+      bands:
+        - {min: 85, level: critical}
+        - {min: 80, level: moderate}
+      default: light
+    confidence: high
+    speedup_range: [1.1, 1.3]
+    speedup_source: 🟡 首期标定值
+    mechanism_refs:
+      - ops-mechanism:references/optimize/data-copy.md
+
+  # ---------------------------------------------------------------- SCALAR
+  - id: L1_SCALAR_BOUND
+    name: Scalar 标量瓶颈
+    priority: 60
+    parent: L1_SCALAR_SATURATION
+    threshold_source: "[tuning-practice] >30% 判定；[tuning-practice] >90% 且 vec/mte 低 → 标量访问/地址计算"
+    metrics_required: [scalar_ratio, vec_ratio]
+    metrics_optional: []
+    condition:
+      any:
+        - {metric: scalar_ratio, agg: mean, op: gt, value: 80}
+        - all:
+            - {metric: scalar_ratio, agg: mean, op: gt, value: 70}
+            - {metric: scalar_ratio, agg: rank, op: eq, value: 1}
+    watch:
+      condition: {metric: scalar_ratio, agg: mean, op: gt, value: 30}
+      note: scalar >30%：常见于数据量小、头开销占比大场景；检查 TilingData 体积与核数 [tuning-practice]
+    severity:
+      metric: scalar_ratio
+      bands:
+        - {min: 85, level: critical}
+        - {min: 70, level: moderate}
+      default: light
+    counter_evidence:
+      - id: address-compute-heavy
+        condition: {metric: scalar_ratio, agg: mean, op: gt, value: 90}
+        note: scalar >90% 而 vec/mte 低：标量访问/地址计算主导，查指针运算与分支 [tuning-practice]
+        action: annotate
+    confidence: high
+    speedup_range: [1.1, 1.5]
+    speedup_source: 🟡 首期标定值
+    mechanism_refs:
+      - ops-mechanism:references/optimize/api-usage.md
+
+  # ------------------------------------------------- LATENCY/SYNC 可疑档
+  - id: L1_LATENCY_SYNC_SUSPECT
+    name: 全流水低占用（LATENCY/SYNC 可疑）
+    priority: 70
+    parent: L1_LATENCY
+    threshold_source: "[tuning-practice] 全 pipe busy 均值 <50% 不是『无 bound』"
+    metrics_required: [vec_ratio, cube_ratio, mte2_ratio, mte3_ratio, scalar_ratio]
+    metrics_optional: []
+    condition:
+      all:
+        - {metric: __max_pipe_mean, op: lt, value: 50}
+    severity:
+      metric: __max_pipe_mean
+      bands:
+        - {max: 30, level: critical, note: 全 pipe <30%，重度等待}
+        - {max: 50, level: moderate}
+      default: light
+    counter_evidence:
+      - id: serial-fingerprint
+        condition: {metric: __pipe_sum_mean, op: between, value: [85, 115]}
+        note: 多 pipe 互补求和≈100%：串行无重叠指纹，Double Buffer / 流水掩盖是首选机制 [tuning-practice]
+        action: annotate
+    confidence: medium
+    speedup_range: [1.2, 2.5]
+    speedup_source: "[tuning-practice] Double Buffer 类机制实测区间，🟡 待校准"
+    mechanism_refs:
+      - ops-mechanism:references/optimize/pipeline.md
+
+  # ------------------------------------------------- 串行无重叠指纹（独立发现）
+  - id: L1_SERIAL_NO_OVERLAP
+    name: 串行无重叠指纹
+    priority: 80
+    parent: L1_LATENCY
+    threshold_source: "[tuning-practice] cube 0.44+mte2 0.42+mte1 0.18≈1 案例"
+    metrics_required: [vec_ratio, cube_ratio, mte2_ratio, mte3_ratio, mte1_ratio, scalar_ratio, fixpipe_ratio]
+    metrics_optional: []
+    condition:
+      all:
+        - {metric: __max_pipe_mean, op: lt, value: 70}
+        - {metric: __pipe_sum_mean, op: between, value: [85, 115]}
+    severity: {default: moderate}
+    confidence: medium
+    speedup_range: [1.3, 2.0]
+    speedup_source: "单缓冲→双缓冲典型收益，🟡 待校准"
+    mechanism_refs:
+      - ops-mechanism:references/optimize/pipeline.md

package/skills/tool-ops-tune/ops-verdict/rules/L2-inter-core.yaml

@@ -0,0 +1,173 @@
+# L2 核间层判定规则包：多核负载与占用
+#
+# 阈值来源：
+#   [tuning-practice] 一线调优实践沉淀：长尾 1.25x 档 / A5 典型 28 AIC·56 AIV（soc-spec.yaml 实测勘误）
+#   [tuning-practice] 逐核耗时离散度三档（<10 达标 / 10~30 警告 / >30 严重，A2/A5 实测标定）
+#   [tuning-practice] 头开销参考值（Atlas A2 满核 ~20-21us，实测；A5 未标定，见 soc-spec.yaml 🟡）
+schema: opstune.rules/1.0
+layer: L2
+title: 核间负载判定
+soc: [a2, a3, a5]
+
+metrics:
+  row_count:
+    source: derived
+    description: PipeUtilization 有效 block 行数（活跃核数）
+  block_dim:
+    source: scalar
+    description: OpBasicInfo Block Dim（逻辑核数）
+  physical_cores:
+    source: scalar
+    description: 目标 SoC 物理核数（调用方提供：优先按 soc-spec.yaml skus 解析实际 SKU；a5 典型 28 AIC/56 AIV、a2 典型 24 AIC/48 AIV——旧注释「A5 约 48 AIV/28 AIC」已按 ini 实测勘误）
+  task_duration_us:
+    source: scalar
+    description: OpBasicInfo Task Duration(us)
+  dominant_time_us:
+    source: derived
+    description: max(aic_time, aiv_time) 逐核均值（核执行主体时间；引擎按 CSV 逐核耗时均值计算）
+  time_mean_us:
+    source: derived
+    description: 逐核耗时均值（活跃核）
+  time_max_us:
+    source: derived
+    description: 逐核耗时最大值（最慢核）
+  time_min_us:
+    source: derived
+    description: 逐核耗时最小值
+  host_dispatch_us:
+    source: scalar
+    description: host 下发间隙（launch → task start，µs；scene=cluster 由 host 侧时间轴产物换算后提供——P2-4 规划态，暂无生产者）
+
+rules:
+  - id: L2_CORE_TAIL
+    name: 核间长尾
+    priority: 10
+    threshold_source: "[tuning-practice] max > mean×1.25；[tuning-practice] 离散度三档"
+    metrics_required: [time_mean_us, time_max_us, time_min_us]
+    metrics_optional: []
+    condition:
+      all:
+        - {metric: time_mean_us, op: gt, value: 0}
+        - {metric: time_max_us, op: gt_ratio, value: {metric: time_mean_us, factor: 1.25}}
+    severity:
+      metric: __tail_ratio
+      bands:
+        - {min: 1.6, level: critical}
+        - {min: 1.25, level: moderate}
+      default: light
+    counter_evidence:
+      - id: numa-two-cluster
+        condition: {metric: __two_cluster_gap_pct, op: gt, value: 2}
+        note: 前后半段核均值差距 ≥2%：疑似 NUMA/L2 slice 负载偏斜，block swat / 尾轮均衡优先 [tuning-practice]
+        action: annotate
+    confidence: high
+    speedup_range: [1.05, 1.3]
+    speedup_source: 🟡 首期标定值
+    mechanism_refs:
+      - ops-mechanism:references/optimize/ascendc-tiling.md
+
+  - id: L2_IMBALANCE_MILD
+    name: 核间负载不均衡（离散度警告档）
+    priority: 15
+    threshold_source: "[tuning-practice] (max-min)/max：10~30% 警告"
+    metrics_required: [time_mean_us, time_max_us, time_min_us]
+    metrics_optional: []
+    condition:
+      all:
+        - {metric: __imbalance_pct, op: gt, value: 10}
+        - {metric: __imbalance_pct, op: lte, value: 30}
+    severity: {default: light}
+    confidence: medium
+    speedup_range: [1.0, 1.1]
+    speedup_source: 🟡 首期标定值
+    mechanism_refs:
+      - ops-mechanism:references/optimize/ascendc-tiling.md
+
+  - id: L2_IMBALANCE_SEVERE
+    name: 核间负载不均衡（离散度严重档）
+    priority: 16
+    threshold_source: "[tuning-practice] (max-min)/max >30% 严重问题"
+    metrics_required: [time_mean_us, time_max_us, time_min_us]
+    metrics_optional: []
+    condition: {metric: __imbalance_pct, op: gt, value: 30}
+    severity: {default: critical}
+    confidence: high
+    speedup_range: [1.1, 1.5]
+    speedup_source: 🟡 首期标定值
+    mechanism_refs:
+      - ops-mechanism:references/optimize/ascendc-tiling.md
+
+  - id: L2_UNDERUTILIZATION
+    name: 切核不足（occupancy 可疑）
+    priority: 20
+    threshold_source: "[tuning-practice] block 行数 ≪ 物理核数；[tuning-practice] Block Dim 不足是最常见低垂果实"
+    metrics_required: [row_count, physical_cores]
+    metrics_optional: [block_dim]
+    condition:
+      all:
+        - {metric: row_count, op: gt, value: 0}
+        - {metric: row_count, op: lt_ratio, value: {metric: physical_cores, factor: 0.33}}
+    severity:
+      metric: __occupancy_pct
+      bands:
+        - {max: 10, level: critical, note: 活跃核不足物理核 10%，切分严重不足}
+        - {max: 33, level: moderate}
+      default: light
+    counter_evidence:
+      - id: small-case-fewer-cores
+        condition: {metric: task_duration_us, op: lt, value: 25}
+        note: Task 极短（<25us）：数据量小，增核可能放大头开销（A2 满核头开销 ~20-21us），需权衡 [tuning-practice]
+        action: annotate
+    confidence: medium
+    speedup_range: [1.2, 3.0]
+    speedup_source: "[tuning-practice] Block Dim 4→20 核案例 4.75x（区间下界保守化），🟡 待校准"
+    mechanism_refs:
+      - ops-mechanism:references/optimize/ascendc-tiling.md
+
+  - id: L2_HEAD_OVERHEAD
+    name: 头开销占比过高
+    priority: 30
+    threshold_source: "[tuning-practice] overhead = TaskDuration − max(aic,aiv)；[tuning-practice] 头开销占比 >30% 说明数据量太小或核数太多"
+    metrics_required: [task_duration_us, dominant_time_us]
+    metrics_optional: []
+    condition:
+      all:
+        - {metric: task_duration_us, op: gt, value: 0}
+        # 木桶口径：头开销以最慢核为基准（Task 总时长 − 最慢核主体时间）/ Task 总时长
+        - {metric: __head_overhead_pct, op: gt, value: 30}
+    severity:
+      metric: __head_overhead_pct
+      bands:
+        - {min: 50, level: critical}
+        - {min: 30, level: moderate}
+      default: light
+    confidence: medium
+    speedup_range: [1.05, 1.2]
+    speedup_source: 🟡 首期标定值
+    mechanism_refs:
+      - ops-mechanism:references/optimize/api-usage.md
+
+  # ------------------------------------------------ host 下发档（P2-4/G17，scene=cluster）
+  # 集群视角的 host 侧证据：下发阻塞（launch → task start 的 host 间隙）主导时，
+  # 算子侧优化收益有限，先解 host 下发（流组织 / launch 合并 / host 占用）。
+  - id: L2_HOST_DISPATCH_BLOCK
+    name: host 下发阻塞（scene=cluster 专用档）
+    priority: 35
+    threshold_source: "🟡 host_dispatch_us > 50 且占 task_duration >30% 为初值占位（首个集群场景实战后标定）"
+    metrics_required: [host_dispatch_us, task_duration_us]
+    metrics_optional: []
+    condition:
+      all:
+        - {metric: host_dispatch_us, op: gt, value: 50}
+        - {metric: __host_dispatch_pct, op: gt, value: 30}
+    severity:
+      metric: __host_dispatch_pct
+      bands:
+        - {min: 50, level: critical, note: 下发间隙超过 task 半程：host 侧完全主导}
+        - {min: 30, level: moderate}
+      default: light
+    confidence: low
+    speedup_range: [1.0, 1.2]
+    speedup_source: 🟡 待标定（P2-4 规划态规则，缺真实集群数据）
+    mechanism_refs:
+      - ops-mechanism:references/optimize/api-usage.md

package/skills/tool-ops-tune/ops-verdict/rules/L3-inter-card.yaml

@@ -0,0 +1,156 @@
+# L3 卡间层判定规则包：MC² 通算融合通信-计算比
+#
+# 阈值来源：
+#   [mc2-practice] MC² 通算融合一线调优实践：R 值五档（真机实测 R 从 0.3 到 9.8）
+#   [mc2-practice] 掩盖率公式 / 跨 rank max 口径（多卡实测沉淀）
+#
+# 前置要求（由 ops-collect 隔离测试产出，缺失时本规则包整体降级为 UNRESOLVED）：
+#   - T_comm：tileCnt=1 注释计算 process，多 rank 并行采集，取各 rank 稳态均值最大值
+#   - T_compute：tileCnt=1 注释通信 process，同上口径
+schema: opstune.rules/1.0
+layer: L3
+title: 卡间通算比判定（MC²）
+soc: [a5]
+requires_experiment:
+  kind: isolation_test
+  outputs: [t_comm_us, t_compute_us, m_rows]
+  note: 隔离测试数据缺失时本包所有规则不触发，输出 UNRESOLVED + 补采指令
+
+metrics:
+  t_comm_us:
+    source: scalar
+    description: 纯通信耗时（隔离测试，跨 rank max 稳态均值）
+  t_compute_us:
+    source: scalar
+    description: 纯计算耗时（隔离测试，跨 rank max 稳态均值）
+  r_value:
+    source: derived
+    description: R = T_compute / T_comm
+  mask_ratio:
+    source: scalar
+    description: 掩盖率 = (融合前耗时−融合后耗时)/min(T_comm,T_compute)
+  rank_imbalance:
+    source: scalar
+    description: max(rank 稳态均值)/mean(rank 稳态均值)
+  mte2_ratio_full:
+    source: scalar
+    description: 完整算子 aic_mte2_ratio（含通信等待，跨 rank max）
+  mte2_ratio_pure:
+    source: scalar
+    description: 隔离测试纯计算路径 aic_mte2_ratio（无通信等待）
+
+rules:
+  - id: L3_MC2_BALANCED
+    name: 通算完美平衡
+    priority: 10
+    threshold_source: "[mc2-practice] R≈1 (0.9~1.1)"
+    metrics_required: [t_comm_us, t_compute_us]
+    metrics_optional: []
+    condition:
+      all:
+        - {metric: r_value, op: between, value: [0.9, 1.1]}
+    severity: {default: light}
+    confidence: high
+    speedup_range: [1.0, 1.1]
+    speedup_source: "[mc2-practice] 等分切分即最优"
+    mechanism_refs:
+      - ops-mechanism:references/optimize/mc2.md
+
+  - id: L3_MC2_COMPUTE_BOUND
+    name: 通算计算 Bound
+    priority: 20
+    threshold_source: "[mc2-practice] 1.1 < R ≤ 2"
+    metrics_required: [t_comm_us, t_compute_us]
+    metrics_optional: [mask_ratio]
+    condition:
+      all:
+        - {metric: r_value, op: gt, value: 1.1}
+        - {metric: r_value, op: lte, value: 2}
+    severity: {default: moderate}
+    confidence: high
+    speedup_range: [1.1, 1.8]
+    speedup_source: "[mc2-practice] 用通信掩盖计算，增大 tileCnt；🟡 待校准"
+    mechanism_refs:
+      - ops-mechanism:references/optimize/mc2.md
+      - ops-mechanism:references/optimize/mc2-balancing.md
+
+  - id: L3_MC2_STRONG_COMPUTE
+    name: 通算强计算 Bound
+    priority: 21
+    threshold_source: "[mc2-practice] R > 2"
+    metrics_required: [t_comm_us, t_compute_us]
+    metrics_optional: []
+    condition: {metric: r_value, op: gt, value: 2}
+    severity: {default: critical}
+    confidence: high
+    speedup_range: [1.0, 1.2]
+    speedup_source: "[mc2-practice] 增大 tileCnt 收益极小，倾向少切分；须搜索确认"
+    mechanism_refs:
+      - ops-mechanism:references/optimize/mc2.md
+      - ops-family:references/matmul/guide.md
+
+  - id: L3_MC2_COMM_BOUND
+    name: 通算通信 Bound
+    priority: 30
+    threshold_source: "[mc2-practice] R < 0.9（上界 0.9 与 BALANCED 下界 0.9 对齐，消除区间重叠）"
+    metrics_required: [t_comm_us, t_compute_us]
+    metrics_optional: [mask_ratio]
+    condition:
+      all:
+        - {metric: r_value, op: lt, value: 0.9}
+        - {metric: r_value, op: gte, value: 0.5}
+    severity: {default: moderate}
+    confidence: high
+    speedup_range: [1.1, 1.8]
+    speedup_source: "[mc2-practice] 用计算掩盖通信，增大 tileCnt；🟡 待校准"
+    mechanism_refs:
+      - ops-mechanism:references/optimize/mc2.md
+      - ops-mechanism:references/optimize/mc2-balancing.md
+
+  - id: L3_MC2_STRONG_COMM
+    name: 通算强通信 Bound
+    priority: 31
+    threshold_source: "[mc2-practice] R < 0.5"
+    metrics_required: [t_comm_us, t_compute_us]
+    metrics_optional: []
+    condition: {metric: r_value, op: lt, value: 0.5}
+    severity: {default: critical}
+    confidence: high
+    speedup_range: [1.2, 2.0]
+    speedup_source: "[mc2-practice] 从短块出发搜索，最大化 tileCnt；🟡 待校准"
+    mechanism_refs:
+      - ops-mechanism:references/optimize/mc2.md
+      - ops-mechanism:references/optimize/mc2-balancing.md
+
+  - id: L3_MTE2_POLLUTED
+    name: MTE2 指标被通信等待污染（路由规则）
+    priority: 5
+    threshold_source: "[mc2-practice] ratio_full−ratio_pure > 0.1 或 ratio_pure/ratio_full < 0.8"
+    metrics_required: [mte2_ratio_full, mte2_ratio_pure]
+    metrics_optional: []
+    condition:
+      any:
+        - {metric: __mte2_pollution_gap, op: gt, value: 10}
+        - {metric: __mte2_pure_ratio_pct, op: lt, value: 80}
+    severity: {default: moderate}
+    confidence: high
+    routing:
+      on_hit: [disable_rule, L1_MTE2_BOUND]
+      reroute_to: ops-mechanism:references/optimize/mc2.md
+      note: MTE2 含通信等待空转，单核访存优化无效，路由到通信掩盖策略
+    mechanism_refs:
+      - ops-mechanism:references/optimize/mc2.md
+
+  - id: L3_RANK_IMBALANCE
+    name: 跨 rank 木桶不均衡
+    priority: 40
+    threshold_source: 🟡 M1 实测标定（初值 1.1 为工程经验占位）
+    metrics_required: [rank_imbalance]
+    metrics_optional: []
+    condition: {metric: rank_imbalance, op: gt, value: 1.1}
+    severity: {default: moderate}
+    confidence: low
+    speedup_range: [1.0, 1.15]
+    speedup_source: 🟡 待标定
+    mechanism_refs:
+      - ops-mechanism:references/optimize/mc2.md