@cannbot-plugin/cannbot-harness 0.0.0-stage → 0.1.5

package/assets/ops-direct-invoke/skills/ops-spec-gen/registries/category_paradigm_map.yaml

@@ -0,0 +1,145 @@
+# 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.
+
+# 算子 spec.yaml — category ↔ paradigm 一致性 + paradigm 内部约束
+schema_version: 1
+
+# 每个 category 必含的 paradigm（stage 2 强校验）
+category_requires_paradigms:
+  Reduction:            [Reduction]
+  ReductionComposite:  [Reduction, FusedComposite]
+  Contraction:          [Contraction]
+  LayoutTransform:     [LayoutTransform]
+  IndexGather:         [IndexGather]
+  ScatterUpdate:       [ScatterUpdate]
+  AtomicUpdate:        [AtomicUpdate]
+  Broadcast:            [Broadcast]
+  ArgReduce:           [ArgReduce]
+  SortSelect:          [SortSelect]
+  MaskPredicate:       [MaskPredicate]
+  Padding:              [Padding]
+  Interpolation:        [Interpolation]
+  Histogram:            [Histogram]
+  SlidingWindow:       [SlidingWindow]
+  ControlFlow:         [ControlFlow]
+  Recurrence:           [Recurrence]
+  RandomSampling:      [RandomSampling]
+  Stateful:             [Stateful]
+  Sparse:               [Sparse]
+  Quantization:         [Quantization]
+  Spectral:             [Spectral]
+  VariableOutput:      [VariableOutput]
+  Collective:           [CollectiveCommunication]
+  FusedComposite:      [FusedComposite]   # 额外要求 ≥ 2 条基础 paradigm（Elementwise 被吸收，不计入；脚本特判）
+
+# 互斥对（同时出现 ⇒ stage 2 FAIL）
+mutually_exclusive_paradigms:
+  - [ScatterUpdate, AtomicUpdate]
+
+# 条件范式路由（paradigm_routing）—— 按属性值动态激活不同 paradigm 子集
+#
+# 适用场景：同一算子在不同属性值下呈现不同计算范式（如 mse_loss_v2 的 reduction 属性：
+#   "none" → Elementwise，"sum"/"mean" → Reduction）。
+#
+# spec.yaml 格式：
+#   op:
+#     paradigms: [Reduction, Elementwise]
+#     paradigm_routing:
+#       attribute: reduction          # 引用 attributes[].name，须为 enum/string 类型
+#       cases:
+#         - value: "none"
+#           active_paradigms: [Elementwise]
+#         - values: ["sum", "mean"]
+#           active_paradigms: [Reduction]
+#
+# stage 2 校验规则（_check_paradigm_routing in validate_spec.py）：
+#   1. attribute 必须在 attributes[] 中存在且为 enum/string 类型
+#   2. 每个 case 须有 value 或 values（不可同时有）
+#   3. 每个 case 的 active_paradigms 必须是 op.paradigms 的子集
+#   4. enum 类型属性的所有 values 必须被 cases 完全覆盖
+#
+# 下游影响：
+#   - slice_design_inputs.py: 按分区注入不同的 patterns.md
+#   - validate_completeness.py: 不剥离 Elementwise，按分区 active_paradigms 校验
+#   - DESIGN.md §3.2: 模板划分表增加「范式分区」列
+
+# paradigm 内部约束（stage 2 校验）—— 真值在 scripts/validate_spec.py 内的硬编码逻辑
+#
+# 历史遗留：本文件曾经计划用声明式 yaml 描述这些约束让 stage 2 自动加载，但
+# 那种通用 mini-DSL 表达力不够（require_role_or_inplace / output_dtype_rule_pattern
+# 都是特化逻辑），最终改为硬编码。这里只列**清单**作维护参考，**改本段不会**
+# **影响校验行为**——要改逻辑请改 scripts/validate_spec.py 的 stage_2 函数。
+#
+# 当前 stage 2 实际校验的 paradigm 内部约束：
+#   - NumericalStable          ⇒ numerical_stability.required == true
+#   - FusedComposite           ⇒ math_semantics.composition 必填、primitives ≥ 2、
+#                                  op 在 PRIMITIVE_WHITELIST、中间不泄漏、dataflow 闭合
+#   - Reduction (无 Recurrence) ⇒ inputs 含 axis/dim/axes 之一 / attributes 含 axis/dim/axes /
+#                                  reduction.axis_source: fixed / reduction.axis_source: implicit_all
+#   - ArgReduce                ⇒ outputs[].dtype_rule 返回固定 int32/int64
+#                                  (新格式：dtype_rule 是 numpy_expr，如 `y.dtype = np.int64`)
+#   - Stateful                 ⇒ inputs.role=state 或 outputs.aliasing=inplace_with(...)
+#   - Quantization             ⇒ 以下任一（attribute 或 tensor input 均可）：
+#                                  {scale, zero_point} / {scale} / {x1Scale, x2Scale}
+#                                  / {dequant_scale} / {deqScale}
+#   - VariableOutput           ⇒ outputs[].data_dependent_shape: true
+#                                  + outputs[].shape_rule_kind: data_dependent
+#                                  + outputs[].shape_bounds.max_elements 非空
+#                                  （校验点：validate_spec.py + stages.py stage 3）
+#   - RandomSampling           ⇒ attributes 含 seed
+#   - CollectiveCommunication  ⇒ op.platform_constraints.requires_hccl: true
+
+# 范式自动注入提示（生成器使用，非校验器 - WARN 而非 FAIL）
+paradigm_inject_hints:
+  Reduction:
+    notes:
+      - "fp16/bf16 输入时 dtype_policy.accumulator_dtype 应 ≥ fp32"
+      - "boundary_conditions 必含'reduce 轴长度为 1' / 'rank=0 标量' / '空 Tensor'"
+  Contraction:
+    notes:
+      - "dtype_policy.accumulator_dtype 必填且 ≥ fp32"
+      - "建议在 invariants 加 equals_when_input_is_zero 或 identity_right"
+  NumericalStable:
+    notes:
+      - "numerical_stability.techniques 至少 1 条；anti_pattern_id 关联 Developer Agent"
+      - "extreme_inputs 必含 fp16 上溢边界 case"
+  FusedComposite:
+    notes:
+      - "math_semantics.composition.primitives 至少 2 条，op 在 PRIMITIVE_WHITELIST 内"
+      - "dataflow.no_leak: true 时中间 tensor 不得出现在 outputs"
+  SlidingWindow:
+    notes:
+      - "attributes 必含 kernel_size / stride / dilation / padding"
+      - "outputs.shape_rule 用 numpy_expr 写卷积/池化输出公式（如 (H + 2*pad - dilation*(kernel-1) - 1) // stride + 1）"
+      - "boundary 必含 'stride > kernel'"
+  Padding:
+    notes:
+      - "attributes 必含 pad_value 与 pad_mode (constant|reflect|replicate)"
+  Quantization:
+    notes:
+      - "numerical_tolerance.per_dtype 必含相应 int8/uint8/int4 容差"
+  Spectral:
+    notes:
+      - "numerical_tolerance.per_dtype 必须含 complex64 / complex128"
+      - "boundary 必含'长度=2 幂次' 与'非 2 幂次'"
+  LayoutTransform:
+    notes:
+      - "numerical_tolerance.metric 推荐 bitwise_equal"
+  AtomicUpdate:
+    notes:
+      - "determinism.accumulation_order 必填（写冲突顺序敏感）"
+  Histogram:
+    notes:
+      - "determinism.required: true + 原子加策略"
+  Recurrence:
+    notes:
+      - "determinism.accumulation_order 必为 sequential"
+  DynamicShape:
+    notes:
+      - "shape_rule_kind 选择：attr-driven（reshape/stridedslice，shape 由属性推出）用 numpy_expr 引用 attribute；value-driven（splitv size_splits 张量输入决定 shape）用 data_dependent + shape_rule_description + shape_bounds"
+      - "shape_set 至少 3 个（由 ascendc-st-design skill 管理）"

package/assets/ops-direct-invoke/skills/ops-spec-gen/registries/chip_registry.yaml

@@ -0,0 +1,82 @@
+# 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.
+
+# Chip registry —— 受控 Atlas 芯片型号 + dtype 矩阵 + DAV 编译宏映射
+#
+# 作用：
+#   * spec.op.platform_constraints.supported_chips 字段的合法取值集合
+#   * spec.op.platform_constraints.dav_macros 字段的合法取值集合
+#   * stage 2 校验 dtype_set 与 supported_chips 兼容性——若 spec 声明 dtype_set 含
+#     fp8/fp4/hf8/int4/bf16 等芯片敏感类型，但 supported_chips 包含不支持该 dtype
+#     的型号，触发 paradigm_constraint.dtype_chip_mismatch
+#
+# 维护：
+#   * 新增芯片：登记到本文件 + schemas/op-spec.json 的 supported_chips.enum
+#   * 新增 dtype：同步更新各芯片的 supported_dtypes 列表
+#   * 新增 DAV 宏：登记到本文件 dav_macros + schemas 的 dav_macros.enum
+#   * registry/schema 同步由 scripts/check_registry_schema_sync.py 守门
+
+schema_version: 1
+
+# 各芯片支持的数值 dtype（仅算子层面可用的；layout 类型不在此处）+ 对应 DAV 编译宏
+chips:
+  - id: Ascend310
+    name: "Atlas 310 推理（昇腾 310）"
+    dav_macros: [DAV_1001]
+    supported_dtypes: [float16, float32, int8, int16, int32, uint8, bool]
+
+  - id: Ascend310P
+    name: "Atlas 310P 推理"
+    dav_macros: [DAV_2002]
+    supported_dtypes: [float16, float32, int8, int16, int32, uint8, bool]
+
+  - id: Ascend910
+    name: "Atlas 910 训练（昇腾 910 原版）"
+    dav_macros: [DAV_2002]
+    supported_dtypes: [float16, float32, float64, int8, int16, int32, int64, uint8, uint16, uint32, bool, complex64]
+
+  - id: Ascend310B
+    name: "Atlas 200I/500 A2 推理"
+    dav_macros: [DAV_2002]
+    supported_dtypes: [float16, float32, int8, int16, int32, uint8, bool]
+
+  - id: Ascend910B
+    name: "Atlas A2 训练 / 推理（昇腾 910B/910C）"
+    dav_macros: [DAV_2201]
+    supported_dtypes:
+      [float16, bfloat16, float32, float64,
+       int4, int8, int16, int32, int64,
+       uint4, uint8, uint16, uint32, uint64,
+       bool, complex64, complex128]
+
+  - id: Ascend950PR
+    name: "Atlas 950PR（昇腾 950 推理）"
+    dav_macros: [DAV_3510]
+    supported_dtypes:
+      [float4_e2m1, float8_e4m3fn, float8_e5m2, hifloat8,
+       float16, bfloat16, float32, float64,
+       int4, int8, int16, int32, int64,
+       uint4, uint8, uint16, uint32, uint64,
+       bool, complex64, complex128]
+
+  - id: Ascend950DT
+    name: "Atlas 950DT（昇腾 950 训练）"
+    dav_macros: [DAV_3510]
+    supported_dtypes:
+      [float4_e2m1, float8_e4m3fn, float8_e5m2, hifloat8,
+       float16, bfloat16, float32, float64,
+       int4, int8, int16, int32, int64,
+       uint4, uint8, uint16, uint32, uint64,
+       bool, complex64, complex128]
+
+# DAV 宏 → 芯片反向索引（生成器 / 校验器使用；冗余但避免遍历芯片列表）
+dav_macro_to_chips:
+  DAV_1001: [Ascend310]
+  DAV_2002: [Ascend310P, Ascend910, Ascend310B]
+  DAV_2201: [Ascend910B]
+  DAV_3510: [Ascend950PR, Ascend950DT]

package/assets/ops-direct-invoke/skills/ops-spec-gen/registries/error_code_enum.yaml

@@ -0,0 +1,25 @@
+# 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.
+
+# 算子 spec.yaml — boundary_conditions[].machine_check (kind=raises_error).error_type 枚举
+#
+# 语义类别（runtime-agnostic）：spec 层不引用具体 ACLNN 错误码常量；由各 runtime
+# adapter（ACLNN / PyTorch / MindSpore）在 ST/UT 阶段把语义类别映射到具体错误类型。
+#
+# 取值原则：
+#   * 类别名能让 agent 直接理解 case 期望失败原因
+#   * 仅描述算子能主动声明的"输入/属性非法"类失败；纯运行时异常（OOM、internal）
+#     仅作兜底，spec 用例通常不应主动断言
+schema_version: 1
+error_codes:
+  - shape_mismatch                  # 输入 shape 不满足约束（rank 不符、广播失败、维度冲突）
+  - dtype_not_supported             # 输入 dtype 不在支持集 / 多输入 dtype 组合非法
+  - attribute_value_out_of_range    # 算子属性（axis/keepdim/概率…）取值非法或越界
+  - null_input                      # 必填输入为空指针/空 tensor 但算子不允许
+  - out_of_memory                   # workspace/tiling 资源不足
+  - internal_error                  # 兜底：spec 用例一般不主动声明

package/assets/ops-direct-invoke/skills/ops-spec-gen/registries/framework_oracle_registry.yaml

@@ -0,0 +1,22 @@
+# 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.
+
+# Framework registry —— stage 9 oracle 可识别的框架集合
+#
+# 作用：作为 stage 9 判别 `reference_oracle.framework` 是否是"已知框架"的参考。
+#       未知框架 ⇒ stage 9 emit `unknown_framework` WARN（仍会尝试真 import）。
+#
+# 使用方法：新增框架只需在本文件 frameworks: 下加一项。代码侧不需修改。
+schema_version: 1
+
+frameworks:
+  - torch        # PyTorch
+  - numpy        # NumPy
+  - scipy        # SciPy
+  - tensorflow   # TensorFlow
+  - jax          # JAX

package/assets/ops-direct-invoke/skills/ops-spec-gen/registries/invariant_kind_registry.yaml

@@ -0,0 +1,130 @@
+# 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.
+
+# Invariant（算子不变量）—— 算子输出必须满足的、可机器断言的数学性质
+#
+# 作用：
+#   算子开发时，光靠"和参考实现对拍"不够。比如 softmax 输出是否真的归一化（sum=1）、
+#   relu 是否真的非负、matmul 输出 shape 是否符合 outputs[].shape_rule 推导 —— 这些"算子应
+#   有的数学性质"用 invariant 显式声明出来，校验器/测试器就能机器断言，而不是依赖人审。
+#
+# 使用方法：
+#   spec.yaml 的 math_semantics.invariants[] 数组里每条形如：
+#     - name: sum_to_one        # 自取的人读名
+#       kind: reduce_equals     # 必须落在本文件的某条 kind
+#       reducer: sum            # 由 kind 的 required_fields 决定
+#       axis: "${attr.dim}"
+#       value: 1.0
+#       tolerance_inherit: true
+#   校验器（stage 2）逐条对照本文件检查：kind 是否合法、required_fields 是否齐全、
+#   tolerance_inherit 是否符合该 kind 的约束。
+#
+# 三组分类：
+#   value      —— 直接约束输出的值（元素级阈值 / 归约后等于常量 / 值域 / 离散集合）
+#   algebraic  —— 算子的代数性质（交换律 / 结合律 / 单调性 / 幺元 / 零元）—— 不依赖参考
+#                 实现，靠算子自身的两种调用方式互相对拍即可断言
+#   structural —— 结构性约束（输出 shape / 中间 tensor 不外漏），与 dtype 解耦，
+#                 因此 tolerance_inherit 必为 false
+#
+# 维护：新增 kind 必须先登记到本文件，并在 schemas/op-spec.json 的 invariants.kind.enum
+#       同步增加；validator 会自动从本文件加载白名单，无需改 Python 代码。
+#
+# 何时用、何时不用：日常算子若 reference_oracle 可达，machine_check.kind=matches_oracle
+#       已是主力检测手段，invariants 可留空；下列范式仍建议显式声明 1+ 条 invariant：
+#         - NumericalStable    (e.g. softmax 的 sum_to_one)
+#         - FusedComposite     (e.g. no_leak_intermediates)
+#         - Reduction          (e.g. reduce_equals 锁归一化语义)
+#         - Contraction        (e.g. equals_when_input_is_zero)
+#       oracle 不可达（自研 / collective / quantization）时 invariant 是唯一机器断言手段。
+
+schema_version: 1
+
+# A 组：value（输出值约束）—— 允许 tolerance_inherit
+value:
+  # 元素级 ≥ 阈值；e.g. relu 输出 ≥ 0；softmax 输出 ≥ 0
+  - kind: elementwise_ge
+    required_fields: [value]
+
+  # 元素级 ≤ 阈值；e.g. sigmoid / softmax 输出 ≤ 1
+  - kind: elementwise_le
+    required_fields: [value]
+
+  # 元素级 == 标量；e.g. zeros_like 输出 ≡ 0；常量算子
+  - kind: elementwise_eq
+    required_fields: [value]
+
+  # 沿轴归约后等于常量；e.g. softmax 沿 reduce_axis 求和 == 1.0；概率分布算子
+  - kind: reduce_equals
+    required_fields: [reducer, axis, value]
+    reducer_enum: [sum, mean, max, min]
+
+  # 输出值落在闭区间 [lo, hi]；e.g. sigmoid ∈ [0,1]；tanh ∈ [-1,1]；clamp(x, lo, hi)
+  - kind: range_in
+    required_fields: [range]
+
+  # 输出取值在有限集合内；e.g. one_hot ∈ {0,1}；sign ∈ {-1,0,1}；bool 算子 ∈ {True,False}
+  - kind: produces_in_set
+    required_fields: [set]
+    forbid_tolerance_inherit: true   # 离散集合无需容差
+
+# B 组：algebraic（代数性质）—— 允许 tolerance_inherit
+algebraic:
+  # f(f(x)) == f(x)；e.g. relu(relu(x)) == relu(x)；abs(abs(x)) == abs(x)；contiguous
+  - kind: idempotent
+    required_fields: []
+
+  # 输入 swap 后输出不变；e.g. add 的 x+y == y+x；mul 的 x*y == y*x；max(x,y) == max(y,x)
+  - kind: equals_under_swap
+    required_fields: [swap]
+
+  # 一侧为零时输出 == 另一侧；e.g. add(x, 0) == x；sub(x, 0) == x（zero_input 是零的那一侧）
+  - kind: equals_input_when_other_is_zero
+    required_fields: [identity_input, zero_input]
+
+  # 一侧为单位元时输出 == 另一侧；e.g. matmul(A, I) == A；mul(x, 1) == x（仅 composition 内可用）
+  - kind: equals_input_when_other_is_identity
+    required_fields: [identity_input, other_input, condition]
+    only_in_composition: true
+
+  # 任一输入为零时输出为零；e.g. matmul(0, B) == 0；mul(0, x) == 0；contraction 类常用
+  - kind: equals_when_input_is_zero
+    required_fields: [value]
+
+  # 沿 batch 维结合律 (A·B)·C ≈ A·(B·C)；e.g. batch_matmul、reduce_sum 累加（浮点容差内）
+  - kind: associative_along_batch
+    required_fields: []
+    require_tolerance_inherit: true
+
+  # 沿某轴单调；e.g. cumsum(非负输入) 沿累加轴 increasing；sort 沿排序轴单调
+  - kind: monotone_along
+    required_fields: [axis, direction]
+    direction_enum: [increasing, decreasing]
+
+# C 组：structural（结构性约束）—— tolerance_inherit 必须为 false
+structural:
+  # 输出是输入的某种 permutation（值集合相同，位置可变）；e.g. sort、shuffle
+  - kind: permutation_of_input
+    required_fields: []
+    forbid_tolerance_inherit: true
+
+  # 对输出施以某外部 op 后等于某输入；e.g. complex 满足 torch.real(out)==real / torch.imag(out)==imag；fft 与 ifft 互逆
+  - kind: equals_after_op
+    required_fields: [op, target_input]
+    forbid_tolerance_inherit: false   # algebraic 可有
+
+  # 输出 shape 等于某既定模式；macro 字段是该模式的标签（保留作历史兼容）。
+  # 真值约束由 stage 3 的 outputs[].shape_rule (numpy_expr) 求解；本 invariant 主要
+  # 给运行期测试 / 文档引用使用，不是 stage 3 的真值校验入口。
+  - kind: shape_equals_macro
+    required_fields: [macro]
+    forbid_tolerance_inherit: true
+
+  # 中间 tensor 不出现在 outputs；e.g. fused softmax 的 max-shifted exp 不外露；任何 FusedComposite 的 dataflow.no_leak
+  - kind: no_leak_intermediates
+    required_fields: []
+    forbid_tolerance_inherit: true

package/assets/ops-direct-invoke/skills/ops-spec-gen/registries/machine_check_kind_registry.yaml

@@ -0,0 +1,84 @@
+# 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.
+
+# Machine check（边界 case 的机器断言方式）—— 给定一个测试 case，机器如何判定算子结果是否合规
+#
+# 作用：
+#   spec.yaml 在 boundary_conditions[] 与 extreme_inputs[] 中列举了大量"特殊 case"
+#   （rank=0 标量、空 Tensor、参数越界、含 NaN、fp16 上溢 ...）。每条 case 必须配
+#   一个 machine_check —— 告诉机器"这个 case 跑出来要满足什么条件才算合规"。
+#   这样测试就能完全自动化，不需要人去看每个 case 的输出截图。
+#
+# 使用方法：
+#   每条 boundary/extreme case 形如：
+#     - case: "reduce 轴长度为 1"
+#       synthesize: {x.shape: "[1]", attr.dim: 0}
+#       machine_check:
+#         kind: equals       # 必须落在本文件登记的某条 kind
+#         value: 1.0         # 由 kind 的 required_fields 决定
+#   校验器（stage 2）逐条检查：kind 是否合法、required_fields 是否齐全、
+#   raises_error 的 error_type 是否在 error_code_enum 内。
+#
+# 何时选哪种 kind（按场景速查）：
+#   合法但退化的 case  —— equals / equals_tensor / returns_empty / equals_input / matches_oracle
+#   异常输入传播       —— produces_nan / nan_propagates / inf_clamps_to
+#   参数越界 / 不合法  —— raises_error
+#   语义不变量保持     —— matches_invariant（引用 math_semantics.invariants[].name）
+#   一般情况下能用 oracle 就用 matches_oracle —— 维护成本最低
+#
+# 维护：新增 kind 必须先登记到本文件，并在 schemas/op-spec.json 的
+#       machine_check.kind.enum 同步增加。
+
+schema_version: 1
+
+machine_check_kinds:
+  # 输出每个元素 == 给定标量；e.g. softmax([x]) → [1.0]；reduce_sum 全零输入 → 0
+  # value 类型：实数 / 复数（"a+bj" 字符串）/ 布尔 / 字符串枚举（"NaN" / "+inf"）
+  - kind: equals
+    required_fields: [value]   # 标量字面量；复数用 "<a>+<b>j" 字符串
+
+  # 输出整体 == 给定 tensor 字面量；e.g. one_hot(2, 5) → [0,0,1,0,0]
+  - kind: equals_tensor
+    required_fields: [value]   # tensor 字面量
+
+  # 输出 shape 含 0 维（空张量）；e.g. softmax(empty[0,4]) → empty 同 shape
+  - kind: returns_empty
+    required_fields: []
+
+  # 在指定范围内全是 NaN；e.g. softmax 对全 -inf 行 → 该行全 NaN（exp 全 0 → 0/0）
+  - kind: produces_nan
+    required_fields: [scope]
+    scope_examples: ["整张量", "整行", "整列", "<axis_index:int>"]
+
+  # NaN 沿指定 axis 传播；e.g. add 任一侧含 NaN → 该位置整行/整列出现 NaN
+  - kind: nan_propagates
+    required_fields: [along_axis]
+
+  # inf 输入被钳制到指定值；e.g. fp16 上溢 → +inf；某些算子内部钳制到 finite_max
+  - kind: inf_clamps_to
+    required_fields: [value]
+    value_enum: ["+inf", "-inf", finite_max, finite_min]
+
+  # 委托给 reference_oracle 计算真值，按 dtype 容差比对；最常用，oracle 可达时优先选
+  # e.g. matmul 标准 case → 用 torch.matmul 算参考输出比对
+  - kind: matches_oracle
+    required_fields: []
+
+  # 满足某条已声明的 invariant；用 name 引用 math_semantics.invariants[].name
+  # e.g. softmax fp16 上溢 case → matches_invariant: sum_to_one（验证语义未崩）
+  - kind: matches_invariant
+    required_fields: [name]
+
+  # 报错且错误语义类别匹配；e.g. matmul K 维不匹配 → shape_mismatch
+  # error_type 必须在 registries/error_code_enum.yaml 的语义类别枚举内
+  - kind: raises_error
+    required_fields: [error_type]
+
+  # 输出 == 输入；e.g. reduce_sum(rank=0 标量) → 标量自己；任何"对该 case 退化为恒等"
+  - kind: equals_input
+    required_fields: []

package/assets/ops-direct-invoke/skills/ops-spec-gen/registries/paradigm_enum.yaml

@@ -0,0 +1,48 @@
+# 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.
+
+# 算子 spec.yaml — op.paradigms 受控集合（27 项 PascalCase）
+schema_version: 1
+paradigms:
+  # 基础（6）
+  - {name: Elementwise,             group: basic,        priority: P0}
+  - {name: Broadcast,                group: access_mod,   priority: P0}
+  - {name: Reduction,                group: basic,        priority: P0}
+  - {name: Contraction,              group: basic,        priority: P0}
+  - {name: ArgReduce,                group: basic,        priority: P0}
+  - {name: LayoutTransform,          group: basic,        priority: P0}
+  # 访存修饰（5）
+  - {name: Padding,                  group: access_mod,   priority: P1}
+  - {name: IndexGather,              group: access_mod,   priority: P0}
+  - {name: ScatterUpdate,            group: access_mod,   priority: P0}
+  - {name: AtomicUpdate,             group: access_mod,   priority: P1}
+  - {name: Interpolation,            group: access_mod,   priority: P0}
+  # 控制类（3）
+  - {name: MaskPredicate,            group: control,      priority: P0}
+  - {name: SortSelect,               group: control,      priority: P0}
+  - {name: Histogram,                group: control,      priority: P1}
+  # 结构修饰（3）
+  - {name: SlidingWindow,            group: structural,   priority: P0}
+  - {name: ControlFlow,              group: structural,   priority: P0}
+  - {name: Recurrence,               group: structural,   priority: P0}
+  # 算法修饰（2）
+  - {name: NumericalStable,          group: algorithmic,  priority: P0}
+  - {name: Spectral,                 group: algorithmic,  priority: P1}
+  # 形状状态修饰（4）
+  - {name: DynamicShape,             group: shape_state,  priority: P0}
+  - {name: VariableOutput,           group: shape_state,  priority: P1}
+  - {name: RandomSampling,           group: shape_state,  priority: P0}
+  - {name: Stateful,                 group: shape_state,  priority: P0}
+  # 数据结构修饰（1）
+  - {name: Sparse,                   group: data_struct,  priority: P0}
+  # 转换语义（1）
+  - {name: Quantization,             group: conversion,   priority: P0}
+  # 通信原语（1）
+  - {name: CollectiveCommunication,  group: communication, priority: P0}
+  # 复合形态（1）
+  - {name: FusedComposite,           group: composite,    priority: P0}

package/assets/ops-direct-invoke/skills/ops-spec-gen/registries/primitive_whitelist.yaml

@@ -0,0 +1,67 @@
+# 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.
+
+# Primitive（白盒原语）—— FusedComposite 类算子分解后允许出现的"基础计算单元"
+#
+# 作用：
+#   有些算子（softmax、attention、layer_norm、复数构造……）不是单一计算，而是
+#   多个基础步骤拼出来的"复合算子"。spec.yaml 的 math_semantics.composition.primitives[]
+#   把这些步骤白盒地写成一串原语序列，下游的 Developer Agent 据此判定：
+#     - 哪些原语可以融合到同一个 kernel
+#     - 中间 tensor 是否真的不外漏（GM 多余往返审计 / AP-006）
+#     - 数据流是否闭合（没有未定义的中间引用）
+#   本文件登记所有合法的原语 op 名，校验器（stage 2）逐条比对。
+#
+# 使用方法：
+#   spec.yaml 中 paradigms 含 FusedComposite 时必填 composition，例如 softmax：
+#     math_semantics:
+#       composition:
+#         primitives:
+#           - id: m
+#             op: reduce               # 必须落在本文件
+#             inputs: [x]
+#             outputs: [x_max]
+#           - id: sub
+#             op: elementwise_binary
+#             inputs: [x, x_max]
+#             outputs: [x_shift]
+#           - id: e
+#             op: elementwise_unary
+#             inputs: [x_shift]
+#             outputs: [x_exp]
+#           - id: s
+#             op: reduce
+#             inputs: [x_exp]
+#             outputs: [denom]
+#           - id: d
+#             op: elementwise_binary
+#             inputs: [x_exp, denom]
+#             outputs: [y]
+#         dataflow:
+#           intermediates: [x_max, x_shift, x_exp, denom]
+#           no_leak: true
+#           fusable_groups: [[m, sub, e, s, d]]
+#
+# 维护：新增原语必须先登记到本文件，并在 schemas/op-spec.json 的
+#       composition.primitives.op.enum 同步增加。原则上保持小集合 —— 新增原语
+#       说明缺乏可复用模式，应先看是否能用现有原语组合表达。
+
+schema_version: 1
+
+primitives:
+  - broadcast            # numpy 风格广播；e.g. add 的 [1,N]+[M,N] 对齐到 [M,N]
+  - reduce               # 沿轴聚合；e.g. softmax 内部 sum(exp); reduce_sum / reduce_max
+  - elementwise_unary    # 一元逐元素；e.g. exp / log / sqrt / relu / sin
+  - elementwise_binary   # 二元逐元素；e.g. add / mul / sub / max(x,y)
+  - cast                 # dtype 转换；e.g. fp16→fp32 累加; fp32→bf16 输出
+  - permute              # 维度重排；e.g. transpose, NCHW→NHWC
+  - concat               # 沿轴拼接；e.g. cat([a,b], axis=-1)
+  - slice                # 切片；e.g. attention 中 q/k/v 拆分; 取子区间
+  - stack                # 沿新轴堆叠；e.g. where 三参数版的虚拟拼接
+  - gemm                 # 矩阵乘加（含 batch）；Cube 单元的核心；e.g. linear / matmul
+  - conv                 # 卷积（含转置卷积）；Cube + Vector 协同；e.g. conv2d