math-skill 3.0.1 → 3.1.1

package/references/books/matrix-analysis.md

@@ -69,39 +69,39 @@
 
 > 每条标注**八维落点**（对应 `../gpu-friendly-math.md` 维度编号），方便直接进 GPU 验收门。
 
-1. **随机化数值线代（randomized NLA）**：用随机投影 + QR（§2.1）做随机 SVD，把 O(n³) 完整分解降到亚二次，对超大权重/激活做低秩草图（sketching）。*落点：维度 2/3——全 GEMM、复杂度可控。*
-2. **低秩注意力 / KV 压缩**：以 Eckart–Young（§7.4）保证截断 SVD 是最优低秩逼近；用**核范数（谱范数对偶，§5.5）**作低秩正则，把 KV-Cache 投影到低维子空间。*落点：维度 2/4——GEMM 链 + 显存压缩。*
-3. **谱归一化（spectral normalization）**：power iteration 估最大奇异值（算子范数 §5.6），约束每层 Lipschitz——GAN/扩散/稳定训练。*落点：维度 1/6——matvec，但迭代串行需块化。*
-4. **Newton–Schulz 正交化（Muon 式）**：极分解（§7.3）把梯度矩阵投到最近正交矩阵，迭代只含矩阵乘——当前最 GPU 友好的"二阶味"更新。*落点：维度 2/6/8——纯 GEMM、可融合、bf16 稳。*
-5. **预条件 / Shampoo / K-FAC**：条件数（§5.8）诊断病态，用 PSD Kronecker 因子（Ch 7）近似 Hessian 做预条件，把病态损失面拉圆。*落点：维度 2/5——小矩阵 GEMM，注意逆运算的精度。*
-6. **Geršgorin 廉价谱半径门（§6.1）**：训练循环里用圆盘界 O(n²) 快速估谱半径，做低成本稳定性 gate，不必跑完整 EVD。*落点：维度 1/3——逐行求和，极廉价。*
-7. **PSD 核工程（Schur 积定理 §7.5）**：用 Hadamard 积组合多个 PSD 核，保证可学习相似度矩阵始终半正定。*落点：维度 1——逐元素张量积，天然友好。*
-8. **Perron–Frobenius 诊断（§8.2–8.5）**：把行随机 attention 当 Markov 算子，用谱 gap 量化 over-smoothing / rank collapse，指导残差与温度设计。*落点：维度 1/3——谱估计廉价，避免深层坍缩。*
-9. **块化 / communication-avoiding 分解**：把 QR、Cholesky（§3.5）写成分块版本，用 GEMM 替代逐列消元，跨设备减少通信轮次。*落点：维度 2/6——串行递推改造成并行 + overlap。*
+1. **随机化数值线代（randomized NLA）**：用随机投影 + QR（§2.1）做随机 SVD，把 O(n³) 完整分解降到亚二次，对超大权重/激活做低秩草图（sketching）。*落点：D2/D3——全 GEMM、复杂度可控。*
+2. **低秩注意力 / KV 压缩**：以 Eckart–Young（§7.4）保证截断 SVD 是最优低秩逼近；用**核范数（谱范数对偶，§5.5）**作低秩正则，把 KV-Cache 投影到低维子空间。*落点：D2/D4——GEMM 链 + 显存压缩。*
+3. **谱归一化（spectral normalization）**：power iteration 估最大奇异值（算子范数 §5.6），约束每层 Lipschitz——GAN/扩散/稳定训练。*落点：D1/D6——matvec，但迭代串行需块化。*
+4. **Newton–Schulz 正交化（Muon 式）**：极分解（§7.3）把梯度矩阵投到最近正交矩阵，迭代只含矩阵乘——当前最 GPU 友好的"二阶味"更新。*落点：D2/D6/D8——纯 GEMM、可融合、bf16 稳。*
+5. **预条件 / Shampoo / K-FAC**：条件数（§5.8）诊断病态，用 PSD Kronecker 因子（Ch 7）近似 Hessian 做预条件，把病态损失面拉圆。*落点：D2/D5——小矩阵 GEMM，注意逆运算的精度。*
+6. **Geršgorin 廉价谱半径门（§6.1）**：训练循环里用圆盘界 O(n²) 快速估谱半径，做低成本稳定性 gate，不必跑完整 EVD。*落点：D1/D3——逐行求和，极廉价。*
+7. **PSD 核工程（Schur 积定理 §7.5）**：用 Hadamard 积组合多个 PSD 核，保证可学习相似度矩阵始终半正定。*落点：D1——逐元素张量积，天然友好。*
+8. **Perron–Frobenius 诊断（§8.2–8.5）**：把行随机 attention 当 Markov 算子，用谱 gap 量化 over-smoothing / rank collapse，指导残差与温度设计。*落点：D1/D3——谱估计廉价，避免深层坍缩。*
+9. **块化 / communication-avoiding 分解**：把 QR、Cholesky（§3.5）写成分块版本，用 GEMM 替代逐列消元，跨设备减少通信轮次。*落点：D2/D6——串行递推改造成并行 + overlap。*
 
 ## GPU 友好性警告
 
 > 评分维度引用 `../gpu-friendly-math.md` 的**八维检查**（张量化 / GEMM 可映射 / 复杂度 / 显存 / 低精度 / 并行 / 稀疏 / 算子融合），此处不重复定义。
 
 **天然友好（math beautiful × GPU friendly）：**
-- **SVD 截断 / 低秩**：写成 GEMM 链（维度 2），压缩 KV-Cache/权重（维度 4）。
-- **Frobenius / 谱范数、Gram 矩阵、Hadamard 积**：批量张量代数（维度 1、2）。
-- **极分解 Newton–Schulz**：纯矩阵乘迭代（维度 2、6、8 可融合），bf16 下稳健。
-- **Geršgorin 圆盘**：O(n²) 逐行求和（维度 1），廉价稳定性估计。
-- **良态 PSD 的 blocked Cholesky / Gram 构造**：分块后是 GEMM 链（维度 2），核方法与协方差预条件常用。
+- **SVD 截断 / 低秩**：写成 GEMM 链（D2），压缩 KV-Cache/权重（D4）。
+- **Frobenius / 谱范数、Gram 矩阵、Hadamard 积**：批量张量代数（D1/D2）。
+- **极分解 Newton–Schulz**：纯矩阵乘迭代（D2/D6/D8 可融合），bf16 下稳健。
+- **Geršgorin 圆盘**：O(n²) 逐行求和（D1），廉价稳定性估计。
+- **良态 PSD 的 blocked Cholesky / Gram 构造**：分块后是 GEMM 链（D2），核方法与协方差预条件常用。
 
 **美但不可算（beautiful, not computable）：**
-- **Jordan 标准型（§3.1）**——经典反例：特征值重数对扰动极度敏感，浮点下根本无法可靠计算，**永远不要当数值工具**（违反维度 5 低精度稳定）。Weyr 型同理。
-- **完整 EVD / SVD 的 O(n³)**——大矩阵直接爆（违反维度 3），必须换随机化/迭代法。
-- **非正规矩阵（non-normal，§2.5 之外）**——特征值不能反映真实行为，需 pseudospectra；低精度下谱失真（维度 5）。
-- **病态 / 高条件数（§5.8）**——会出现灾难性抵消，要 fp64 才对，与 bf16/fp8 训练冲突（维度 5）。
-- **QR / Cholesky 的串行依赖（§2.1, §3.5）**——朴素实现是长串行递推（违反维度 6），需 blocked / communication-avoiding 变体。
-- **power iteration 的串行迭代**——单向量迭代并行度低；需块化（block / subspace iteration）才吃满 SM（维度 6）。
+- **Jordan 标准型（§3.1）**——经典反例：特征值重数对扰动极度敏感，浮点下根本无法可靠计算，**永远不要当数值工具**（违反 D5）。Weyr 型同理。
+- **完整 EVD / SVD 的 O(n³)**——大矩阵直接爆（违反 D3），必须换随机化/迭代法。
+- **非正规矩阵（non-normal，§2.5 之外）**——特征值不能反映真实行为，需 pseudospectra；低精度下谱失真（D5）。
+- **病态 / 高条件数（§5.8）**——会出现灾难性抵消，要 fp64 才对，与 bf16/fp8 训练冲突（D5）。
+- **QR / Cholesky 的串行依赖（§2.1, §3.5）**——朴素实现是长串行递推（违反 D6），需 blocked / communication-avoiding 变体。
+- **power iteration 的串行迭代**——单向量迭代并行度低；需块化（block / subspace iteration）才吃满 SM（D6）。
 
 **改造手法（呼应 `../gpu-friendly-math.md` 的 Make-It-Computable Toolkit）：**
-- 完整 EVD/SVD → **随机化 + 截断**降复杂度（维度 3）；
-- 精确分解 → **块化 / GEMM 化**消除串行（维度 2/6）；
-- 病态/非正规 → **重参数化 + 谱归一化**稳低精度（维度 5）；
+- 完整 EVD/SVD → **随机化 + 截断**降复杂度（D3）；
+- 精确分解 → **块化 / GEMM 化**消除串行（D2/D6）；
+- 病态/非正规 → **重参数化 + 谱归一化**稳低精度（D5）；
 - Jordan 等不可算标准型 → 只留作**理论证明**，数值上换 Schur/SVD。
 
 ## 该调用哪个思想透镜
@@ -125,7 +125,7 @@
 - **数值上假设精确正定**：浮点下 Gram/协方差可能丢正定性，要加 jitter（对角抖动）或用 Cholesky 带 pivot。
 - **用 Frobenius 范数当低秩正则**：Frobenius / 权重衰减压的是"能量"不是"秩"；要低秩得用核范数或显式低秩参数化（如 LoRA）。
 - **对非对称矩阵谈"特征值大小"**：度量能量/范数/稳定裕度该看**奇异值**；非正规矩阵的特征值模长会严重误导（瞬态增长远超谱半径预测）。
-- **把 O(n²) Gram 矩阵全量物化**：attention/核矩阵不分块就直接撑爆显存；应走 FlashAttention 式融合 + 分块（呼应 GPU 八维维度 4/8）。
+- **把 O(n²) Gram 矩阵全量物化**：attention/核矩阵不分块就直接撑爆显存；应走 FlashAttention 式融合 + 分块（呼应 GPU 八维 D4/D8）。
 - **堆定理而不诊断瓶颈**：先问"算法瓶颈是谱、低秩还是稳定性"，再选结构，别一上来灌矩阵理论。
 
 ## 深挖入口

package/references/books/micro-lie-theory.en.md

@@ -62,7 +62,7 @@ Two additional directly usable "components":
 
 ## GPU Friendliness Warning
 
-> The sole authority for the acceptance gate: [`../gpu-friendly-math.md`](../gpu-friendly-math.md) (eight dimensions: D1 Tensorization, D2 GEMM-mappability, D3 Complexity, D4 Memory/KV, D5 Low-precision stability, D6 Parallelism & communication, D7 Sparse structure, D8 Operator fusion). Dimension-by-dimension scoring follows.
+> The sole authority for the acceptance gate: [`../gpu-friendly-math.en.md`](../gpu-friendly-math.en.md) (eight dimensions: D1 Tensorization, D2 GEMM-mappability, D3 Complexity, D4 Memory/KV, D5 Low-precision stability, D6 Parallelism & communication, D7 Sparse structure, D8 Operator fusion). Dimension-by-dimension scoring follows.
 
 | # | Dimension | Rating | Notes |
 |---|---|---|---|

package/references/books/optimization-ml.en.md

@@ -102,32 +102,32 @@ Actual chapter map:
 
 ## GPU Friendliness Warning
 
-> The following item-by-item evaluation uses the **eight-dimension scorecard** from `../gpu-friendly-math.md` (read that first).
+> The following item-by-item evaluation uses the **eight-dimension scorecard** from `../gpu-friendly-math.en.md` (read that first).
 > Conclusion: The book's first-order / stochastic / low-rank content is naturally GPU-friendly; second-order methods and search-based methods require heavy adaptation or elimination.
 
 - **Full Hessian in second-order methods (Ch 9, 11)**
-  - Violates **Dimension 4 (Memory)**: Dense Hessian is O(d^2), impossible to materialize for hundreds of millions of parameters.
-  - Violates **Dimension 2 (GEMM-mappability)**: Inversion / factorization is not batched-GEMM friendly.
-  - Violates **Dimension 5 (Low-precision stability)**: Ill-conditioned inversion causes catastrophic cancellation under bf16/fp16.
-  - Adaptation: Matrix-free HVP + CG, L-BFGS low-rank, K-FAC block-diagonal -> restores **Dimension 2/4** friendliness.
+  - Violates **D4**: Dense Hessian is O(d^2), impossible to materialize for hundreds of millions of parameters.
+  - Violates **D2**: Inversion / factorization is not batched-GEMM friendly.
+  - Violates **D5**: Ill-conditioned inversion causes catastrophic cancellation under bf16/fp16.
+  - Adaptation: Matrix-free HVP + CG, L-BFGS low-rank, K-FAC block-diagonal -> restores **D2/D4** friendliness.
 - **Line search (Ch 7 golden section / Armijo S24.4.4)**
-  - Violates **Dimension 1 (Tensorization)**: Essentially a scalar loop + data-dependent branching (step-by-step comparison of function values).
-  - Violates **Dimension 8 (Operator fusion)**: Frequent small kernels, control-flow divergence.
+  - Violates **D1**: Essentially a scalar loop + data-dependent branching (step-by-step comparison of function values).
+  - Violates **D8**: Frequent small kernels, control-flow divergence.
   - Adaptation: Large models almost universally use scheduled step sizes (warmup + cosine) instead of step-by-step line search.
 - **First-order / SGD (Ch 8, 27)**
-  - **Dimension 1/2 friendly**: Gradients = tensor algebra, expressible as GEMM.
-  - **Dimension 6 (Parallelism and communication)** is good: Data-parallel all-reduce can overlap with backprop -- the fundamental reason this is the default approach for large models.
+  - **D1/D2 friendly**: Gradients = tensor algebra, expressible as GEMM.
+  - **D6** is good: Data-parallel all-reduce can overlap with backprop -- the fundamental reason this is the default approach for large models.
 - **Global search (Ch 14: GA/SA/PSO)**
-  - Violates **Dimension 1**: Non-differentiable, discrete stochastic search, blocks end-to-end gradient training.
-  - **Dimension 6** is also poor: Inter-population dependencies, hard to parallelize. Only suitable for black-box hyperparameter search, not for the training inner loop.
+  - Violates **D1**: Non-differentiable, discrete stochastic search, blocks end-to-end gradient training.
+  - **D6** is also poor: Inter-population dependencies, hard to parallelize. Only suitable for black-box hyperparameter search, not for the training inner loop.
 - **LP simplex / interior-point methods (Ch 16, 18)**
-  - Simplex row operations are acceptable; interior-point methods require solving a linear system at each step, **Dimension 2/6** limited by factorization and communication.
+  - Simplex row operations are acceptable; interior-point methods require solving a linear system at each step, **D2/D6** limited by factorization and communication.
   - Suitable for medium-scale constrained subproblems, not for placement inside every training step.
 - **Projections (S24.2)**
-  - Norm-ball / box projections are cheap elementwise operations (**Dimension 1/8** friendly).
-  - General polyhedral projections require solving a QP (**Dimension 2** degrades).
+  - Norm-ball / box projections are cheap elementwise operations (**D1/D8** friendly).
+  - General polyhedral projections require solving a QP (**D2** degrades).
 - **Convergence depends on condition number kappa**
-  - Large kappa -> **Dimension 5** is further amplified at low precision (ill-conditioned gradients).
+  - Large kappa -> **D5** is further amplified at low precision (ill-conditioned gradients).
   - Must pair with normalization / preconditioning; otherwise bf16 training diverges.
 
 ## Which Thinking Lens to Invoke

package/references/books/optimization-ml.md

@@ -106,28 +106,28 @@
 > 结论：本书的一阶 / 随机 / 低秩部分天然友好；二阶法与搜索类方法需重度改造或淘汰。
 
 - **二阶法的全 Hessian（Ch 9, 11）**
-  - 违反**维度 4（显存）**：稠密 Hessian 是 O(d²)，亿级参数无法物化。
-  - 违反**维度 2（GEMM 可映射）**：求逆 / 分解非 batched-GEMM 友好。
-  - 违反**维度 5（低精度稳定）**：病态求逆在 bf16/fp16 灾难性抵消。
-  - 改造：matrix-free HVP + CG、L-BFGS 低秩、K-FAC 块对角 → 重回**维度 2/4** 友好。
+  - 违反**D4**：稠密 Hessian 是 O(d²)，亿级参数无法物化。
+  - 违反**D2**：求逆 / 分解非 batched-GEMM 友好。
+  - 违反**D5**：病态求逆在 bf16/fp16 灾难性抵消。
+  - 改造：matrix-free HVP + CG、L-BFGS 低秩、K-FAC 块对角 → 重回**D2/D4** 友好。
 - **线搜索（Ch 7 黄金分割 / Armijo §24.4.4）**
-  - 违反**维度 1（张量化）**：本质是标量循环 + 数据相关分支（逐步比较函数值）。
-  - 违反**维度 8（算子融合）**：频繁小 kernel、控制流发散。
+  - 违反**D1**：本质是标量循环 + 数据相关分支（逐步比较函数值）。
+  - 违反**D8**：频繁小 kernel、控制流发散。
   - 改造：大模型几乎都用调度式步长（warmup + cosine）替代逐步线搜索。
 - **一阶 / SGD（Ch 8, 27）**
-  - **维度 1/2 友好**：梯度 = 张量代数、可写成 GEMM。
-  - **维度 6（并行与通信）**良好：数据并行 all-reduce 可与反向 overlap——大模型默认路线的根本原因。
+  - **D1/D2 友好**：梯度 = 张量代数、可写成 GEMM。
+  - **D6**良好：数据并行 all-reduce 可与反向 overlap——大模型默认路线的根本原因。
 - **全局搜索（Ch 14：GA/SA/PSO）**
-  - 违反**维度 1**：不可微、离散随机搜索，阻断端到端梯度训练。
-  - **维度 6** 也差：种群间依赖、难并行映射。仅适合黑盒超参搜索，不进训练内环。
+  - 违反**D1**：不可微、离散随机搜索，阻断端到端梯度训练。
+  - **D6** 也差：种群间依赖、难并行映射。仅适合黑盒超参搜索，不进训练内环。
 - **LP 单纯形 / 内点法（Ch 16, 18）**
-  - 单纯形的行运算尚可；内点法每步要解线性系统，**维度 2/6** 受限于分解与通信。
+  - 单纯形的行运算尚可；内点法每步要解线性系统，**D2/D6** 受限于分解与通信。
   - 适合中等规模约束子问题，不适合放进每个训练 step。
 - **投影（§24.2）**
-  - norm-ball / box 投影是廉价 elementwise（**维度 1/8** 友好）。
-  - 一般多面体投影要解 QP（**维度 2** 退化）。
+  - norm-ball / box 投影是廉价 elementwise（**D1/D8** 友好）。
+  - 一般多面体投影要解 QP（**D2** 退化）。
 - **收敛依赖条件数 κ**
-  - κ 大 → **维度 5** 在低精度下被进一步放大（病态梯度）。
+  - κ 大 → **D5** 在低精度下被进一步放大（病态梯度）。
   - 必须配归一化 / 预条件，否则 bf16 训练发散。
 
 ## 该调用哪个思想透镜

package/references/books/smooth-manifolds.en.md

@@ -25,7 +25,7 @@ Actual chapter map (2nd ed., chapter numbers match the book):
 
 ## Core Structures Transferable to AI/Infra
 
-- **Tangent space = local linearization of parameter/latent space**. `df_p: T_pM -> T_{f(p)}N` is the Jacobian / pushforward; backpropagation = pushforward along composed maps (the geometric version of the chain rule). All first-order methods live in the tangent space.
+- **Tangent space = local linearization of parameter/latent space**. `df_p: T_pM -> T_{f(p)}N` is the Jacobian / pushforward (corresponding to JVP / forward-mode AD); backpropagation = pullback on the cotangent bundle (VJP = vector-Jacobian product = pullback of covectors), i.e., pullback along composed maps (the geometric version of the chain rule). All first-order methods live in the tangent space.
 - **Gradients are covectors, not vectors**. Autodiff produces 1-forms (elements of the cotangent space); to obtain a descent direction (tangent vector) one must use the **metric to raise indices** (sharp). Euclidean metric -> ordinary gradient; Fisher metric -> natural gradient. **This is the manifold-level root cause of natural gradients / mirror descent**.
 - **Constraint sets = submanifolds**. Regular level set theorem: when g is a submersion, the solution set `g(x)=c` is a smooth submanifold; constrained optimization = unconstrained optimization on a submanifold.
 - **Lie groups = differentiable symmetry groups**. SO(n)/U(n)/Stiefel/Grassmannian are all manifolds; their Lie algebras (e.g., skew-symmetric matrices so(n)) are linear spaces, mapped back to the group via `exp` -> **reparameterize "constrained weights" as "unconstrained Lie algebra + exp"**.
@@ -53,11 +53,11 @@ Actual chapter map (2nd ed., chapter numbers match the book):
 
 ## GPU Friendliness Warning
 
-> The sole authority for the acceptance gate: the **eight dimensions** in `../gpu-friendly-math.md`. The success or failure of manifold methods almost entirely hinges on one point: **whether the retraction / exponential map can be tensorized and expressed as GEMM, or must be solved iteratively.**
+> The sole authority for the acceptance gate: the **eight dimensions** in `../gpu-friendly-math.en.md`. The success or failure of manifold methods almost entirely hinges on one point: **whether the retraction / exponential map can be tensorized and expressed as GEMM, or must be solved iteratively.**
 
 Dimension-by-dimension comparison:
 
-- **D1-D2 Tensorization / GEMM**: tangent-space operations (pushforward/pullback, Jacobian-vector products, projecting gradients to the tangent space) are **naturally batched GEMM** [v] -- backpropagation itself is pushforward, which is extremely GPU-friendly. **However**, retractions/exp typically require QR, eigendecomposition, matrix exponentials, or small-matrix inverses: QR/eig are **not clean GEMM**, they are decompositions with serial dependencies (cuSOLVER batched small matrices are acceptable, but large matrices are O(n^3) with poor parallelism) -> **reformable** rather than natively friendly.
+- **D1-D2 Tensorization / GEMM**: tangent-space operations (pushforward/pullback, Jacobian-vector products, projecting gradients to the tangent space) are **naturally batched GEMM** [v] -- backpropagation itself is pullback (VJP), which is extremely GPU-friendly. **However**, retractions/exp typically require QR, eigendecomposition, matrix exponentials, or small-matrix inverses: QR/eig are **not clean GEMM**, they are decompositions with serial dependencies (cuSOLVER batched small matrices are acceptable, but large matrices are O(n^3) with poor parallelism) -> **reformable** rather than natively friendly.
 - **D3 Complexity**: geodesic distances, parallel transport, general `log|det J|` are all O(n^3) or worse. **Fix**: restrict to manifolds with closed-form geodesics (spheres/hyperbolic/SO(3)); use triangular/coupling layers in normalizing flows so that log-det reduces to a diagonal sum (O(n)).
 - **D5 Low-precision**: [~] **biggest pitfall**. Matrix `exp / log / sqrt`, eigendecompositions, and the affine-invariant metric for SPD are **catastrophically unstable** under bf16/fp16, often silently requiring fp32/fp64. Manifold primitives frequently "appear to run but have numerically diverged long ago."
 - **D6 Parallelism & communication**: the squaring chain in scaling-and-squaring, ODE integration steps, Householder/QR all have **serial recurrences**, making cross-SM/device overlap difficult. Positive counter-example: explicit symplectic integrators (leapfrog) are highly parallel [v].

package/references/books/smooth-manifolds.md

@@ -25,7 +25,7 @@
 
 ## 可迁移到 AI/Infra 的核心结构
 
-- **切空间 = 参数/隐空间的局部线性化（local linearization）**。`df_p: T_pM → T_{f(p)}N` 就是 Jacobian / pushforward；反向传播 = 沿复合映射做 pushforward（链式法则的几何版）。一切一阶方法都活在切空间里。
+- **切空间 = 参数/隐空间的局部线性化（local linearization）**。`df_p: T_pM → T_{f(p)}N` 就是 Jacobian / pushforward（前推，对应 JVP / 前向模式 AD）；反向传播 = 余切丛上的拉回（pullback on cotangent bundle，VJP = 向量-Jacobian 积 = 余向量的拉回），即沿复合映射做 pullback（链式法则的几何版）。一切一阶方法都活在切空间里。
 - **梯度是余向量（covector），不是向量**。autodiff 给出的是 1-form（余切空间元素）；要变成可下降的方向（切向量）必须用 **度量升指标**（(sharp)）。欧氏度量 → 普通梯度；Fisher 度量 → 自然梯度（natural gradient）。**这是自然梯度 / 镜像下降的流形根因**。
 - **约束集 = 子流形（submanifold）**。正则水平集定理：当 g 是 submersion 时 `g(x)=c` 的解集是光滑子流形；约束优化 = 在子流形上做无约束优化。
 - **李群 = 可微的对称群**。SO(n)/U(n)/Stiefel/Grassmann 都是流形；其李代数（如反对称矩阵 so(n)）是线性空间，用 `exp` 映射回群 → **把"约束权重"重参数化为"无约束李代数 + exp"**。
@@ -57,11 +57,11 @@
 
 逐维对照：
 
-- **维度 1–2 张量化 / GEMM**：切空间运算（pushforward/pullback、Jacobian-向量积、把梯度投影到切空间）**天然是 batched GEMM** [v]——反向传播本就是 pushforward，这部分对 GPU 极友好。**但** retraction/exp 多半要 QR、特征分解、矩阵指数或小矩阵求逆：QR/eig **不是干净的 GEMM**，是带串行依赖的分解（cuSOLVER 批量小矩阵尚可，大矩阵 O(n³) 且并行差）→ **可改造** 而非天然友好。
-- **维度 3 复杂度**：测地线距离、平行移动、一般 `log|det J|` 都是 O(n³) 起。**改造**：限定有闭式测地线的流形（球面/双曲/SO(3)）；归一化流用三角/耦合层让 log-det 退化成对角和（O(n)）。
-- **维度 5 低精度**：[~] **最大坑**。矩阵 `exp / log / sqrt`、特征分解、SPD 的仿射不变度量在 bf16/fp16 下 **灾难性不稳定**，常静默地需要 fp32/fp64。流形原语经常"表面能跑、数值早已发散"。
-- **维度 6 并行与通信**：scaling-and-squaring 的平方链、ODE 积分步、Householder/QR 都有 **串行递推**，难跨 SM/设备 overlap。反例向好：显式辛积分器（leapfrog）高并行 [v]。
-- **维度 4/7/8 显存 / 稀疏 / 融合**：李代数/旋转参数化若限制在 **小矩阵或块对角**（如逐头旋转、SO(3) 的 Rodrigues 闭式），可融进 kernel、走 Tensor Core；大稠密流形算子则要物化大中间张量。
+- **D1–D2** 切空间运算（pushforward/pullback、Jacobian-向量积、把梯度投影到切空间）**天然是 batched GEMM** [v]——反向传播本就是 pullback（VJP），这部分对 GPU 极友好。**但** retraction/exp 多半要 QR、特征分解、矩阵指数或小矩阵求逆：QR/eig **不是干净的 GEMM**，是带串行依赖的分解（cuSOLVER 批量小矩阵尚可，大矩阵 O(n³) 且并行差）→ **可改造** 而非天然友好。
+- **D3 复杂度**：测地线距离、平行移动、一般 `log|det J|` 都是 O(n³) 起。**改造**：限定有闭式测地线的流形（球面/双曲/SO(3)）；归一化流用三角/耦合层让 log-det 退化成对角和（O(n)）。
+- **D5 低精度**：[~] **最大坑**。矩阵 `exp / log / sqrt`、特征分解、SPD 的仿射不变度量在 bf16/fp16 下 **灾难性不稳定**，常静默地需要 fp32/fp64。流形原语经常"表面能跑、数值早已发散"。
+- **D6 并行与通信**：scaling-and-squaring 的平方链、ODE 积分步、Householder/QR 都有 **串行递推**，难跨 SM/设备 overlap。反例向好：显式辛积分器（leapfrog）高并行 [v]。
+- **D4/D7/D8 显存 / 稀疏 / 融合**：李代数/旋转参数化若限制在 **小矩阵或块对角**（如逐头旋转、SO(3) 的 Rodrigues 闭式），可融进 kernel、走 Tensor Core；大稠密流形算子则要物化大中间张量。
 
 **结论与改造手法（呼应 gpu-friendly-math.md 工具箱）**：
 

package/references/gpu-friendly-math.en.md

@@ -1,9 +1,7 @@
 # GPU-Friendly Math Checklist
 
 > This file is the **single source of truth** for the "GPU-feasibility" acceptance gate.
-> The activator, the 15 thinking lenses, `books/*.md`, and `../agents/math-critic.md` all reference this file; no duplicate definitions elsewhere.
->
-> This file is the single source of truth for the "GPU-feasibility" acceptance gate. The activator, the 15 thinking lenses, the book references, and the math-critic all point here.
+> The activator, the 15 thinking lenses, `books/*.md`, and `../agents/math-critic.en.md` all reference this file; no duplicate definitions elsewhere.
 
 ## Eight-Dimension Abbreviations
 
@@ -18,6 +16,21 @@
 | D7 | Sparse structure |
 | D8 | Operator fusion |
 
+## Quantitative Checklist
+
+When evaluating each dimension, provide concrete numbers rather than just labels:
+
+| Dimension | Quantitative Questions to Answer |
+|-----------|--------------------------------|
+| D1 Tensorization | What are the tensor shapes of core operations? Where is the batch dimension? |
+| D2 GEMM-mappability | How many GEMM/matmul operations? What are the (M,N,K) dimensions of each? |
+| D3 Complexity | Total FLOPs? Ratio vs. baseline (standard attention/MLP)? |
+| D4 Memory | Peak memory (bytes)? Is an n×n matrix materialized? KV-Cache overhead? |
+| D5 Low-precision | Numerical error magnitude under bf16/fp8? Mixed-precision strategy needed? |
+| D6 Parallelism | Theoretical parallelism degree? Communication volume (bytes/step)? All-reduce required? |
+| D7 Sparsity | Sparsity ratio (%)? Sparse format (CSR/BSR/block-sparse)? Dedicated kernel available? |
+| D8 Operator Fusion | Number of fusible kernels? Reduction in kernel launches and memory transfers after fusion? |
+
 ## Core Proposition
 
 **Mathematical beauty ≠ computability.** For a structure to truly enter training and inference on modern GPU clusters, it must simultaneously satisfy two requirements:
@@ -66,13 +79,13 @@ Common techniques for transforming "beautiful but non-computable" into "both bea
 
 ## Worked Example: Tropical Sheaf Attention
 
-Drawn from the auto-research directions cited in `agentic-workflow.md`, demonstrating how a **candidate design enters the 8-dimension verification**:
+Drawn from the auto-research directions cited in `agentic-workflow.en.md`, demonstrating how a **candidate design enters the 8-dimension verification**:
 
 | Component | Mathematical Source | GPU Friendliness |
 |-----------|-------------------|-----------------|
-| Tropical Gating | Tropical semiring, piecewise-linear | Replaces Top-K: element-wise max-plus — dim 1 [v] Tensorization / dim 2 [x] Not a Tensor Core GEMM (runs on CUDA cores) / dim 3 [v] Per-token gating only, sub-quadratic (min-plus matmul is APSP-hard, not sub-quadratic); sub-differentiable, kinks require LogSumExp smoothing (smoothing recovers standard softmax) |
-| Cellular Sheaf Diffusion | Algebraic geometry / topology (sheaves, restriction maps) | Each edge is a low-rank linear transform = small GEMM (dim 2/4) |
-| Čech Cohomology Regularization | Algebraic topology (first cohomology $H^1$) | Local, inexpensive; serves as an algebraic criterion for hallucination (dim 3/8) |
-| Low-Rank Basis KV Compression (Plücker/Grassmannian perspective) | Projective geometry | Store the basis rather than Plücker coordinates (the latter expands when low-rank); block-summary candidate — compression ratio / error / throughput must be benchmarked (dim 4) |
+| Tropical Gating | Tropical semiring, piecewise-linear | Replaces Top-K: element-wise max-plus — D1[v] / D2[x] Not a Tensor Core GEMM (runs on CUDA cores) / D3[v] Per-token gating only, sub-quadratic (min-plus matmul is APSP-hard, not sub-quadratic); sub-differentiable, kinks require LogSumExp smoothing (smoothing recovers standard softmax) |
+| Cellular Sheaf Diffusion | Algebraic geometry / topology (sheaves, restriction maps) | Each edge is a low-rank linear transform = small GEMM (D2/D4) |
+| Čech Cohomology Regularization | Algebraic topology (first cohomology $H^1$) | Local, inexpensive; serves as an algebraic criterion for hallucination (D3/D8) |
+| Low-Rank Basis KV Compression (Plücker/Grassmannian perspective) | Projective geometry | Store the basis rather than Plücker coordinates (the latter expands when low-rank); block-summary candidate — compression ratio / error / throughput must be benchmarked (D4) |
 
 Do not treat the table above as validated conclusions. The correct approach is to enter each component into the test plan: prove or estimate complexity, measure peak memory and throughput, check bf16/fp8 stability, and confirm whether it can be mapped to GEMM / batched GEMM / fused kernels. Only after both empirical benchmarks and theoretical derivations pass should a component be labeled "math beautiful × GPU friendly."

package/references/gpu-friendly-math.md

@@ -20,6 +20,21 @@
 | D7 | 稀疏结构 |
 | D8 | 算子融合 |
 
+## 量化检查清单
+
+评估每个维度时，应填写具体数值而非仅给标签：
+
+| 维度 | 需回答的量化问题 |
+|------|----------------|
+| D1 张量化 | 核心操作的 tensor shape 是什么？batch 维在哪？ |
+| D2 GEMM 可映射 | 操作可分解为几次 GEMM/matmul？每次的 (M,N,K) 是多少？ |
+| D3 复杂度 | FLOPs 总量？与 baseline (标准 attention/MLP) 的比值？ |
+| D4 显存 | 峰值显存 (bytes)？是否物化 n×n 矩阵？KV-Cache 额外开销？ |
+| D5 低精度 | bf16/fp8 下的数值误差量级？是否需要混合精度策略？ |
+| D6 并行 | 理论并行度？通信量 (bytes/step)？是否需要 all-reduce？ |
+| D7 稀疏 | 稀疏度 (%)？稀疏格式 (CSR/BSR/block-sparse)？是否有专用 kernel？ |
+| D8 算子融合 | 可融合的 kernel 数？融合后减少的 kernel launch 次数和显存搬运量？ |
+
 ## 核心命题 / Core Proposition
 
 **数学美 ≠ 可算。** 一个结构要真正进入现代 GPU 集群的训练与推理，必须同时满足两件事：
@@ -72,9 +87,9 @@
 
 | 组件 | 数学来源 | GPU 友好性 |
 |------|---------|-----------|
-| 热带门控 Tropical Gating | 热带半环分段线性 | 替代 Top-K：逐元素 max-plus——dim 1 [v] 张量化 / dim 2 [x] 非 Tensor Core GEMM（落 CUDA core）/ dim 3 [v] 仅逐 token 门控亚二次（min-plus matmul 则 APSP-hard 非亚二次）；次可微，折点需 LogSumExp 软化（软化退回标准 softmax）|
-| 胞腔层扩散 Cellular Sheaf Diffusion | 代数几何/拓扑（层、限制映射）| 每边低秩线性变换 = 小 GEMM（维度 2/4）|
-| Čech 上同调正则 | 代数拓扑（一阶上同调 H¹）| 局部、廉价；作幻觉的代数判据（维度 3/8）|
-| 低秩基底 KV 压缩（Plücker/Grassmannian 视角） | 射影几何 | 存基底而非 Plücker 坐标（后者低秩时反扩张）；块摘要候选，压缩率/误差/吞吐需实测（维度 4）|
+| 热带门控 Tropical Gating | 热带半环分段线性 | 替代 Top-K：逐元素 max-plus——D1[v] / D2[x] 非 Tensor Core GEMM（落 CUDA core）/ D3[v] 仅逐 token 门控亚二次（min-plus matmul 则 APSP-hard 非亚二次）；次可微，折点需 LogSumExp 软化（软化退回标准 softmax）|
+| 胞腔层扩散 Cellular Sheaf Diffusion | 代数几何/拓扑（层、限制映射）| 每边低秩线性变换 = 小 GEMM（D2/D4）|
+| Čech 上同调正则 | 代数拓扑（一阶上同调 H¹）| 局部、廉价；作幻觉的代数判据（D3/D8）|
+| 低秩基底 KV 压缩（Plücker/Grassmannian 视角） | 射影几何 | 存基底而非 Plücker 坐标（后者低秩时反扩张）；块摘要候选，压缩率/误差/吞吐需实测（D4）|
 
 使用时不要把上表当作已验证结论。正确做法是把每个组件写进 testplan：证明或估计复杂度，测峰值显存和吞吐，检查 bf16/fp8 稳定性，确认能否落到 GEMM / batched GEMM / fused kernel。只有实测和推导都通过后，才标为 "math beautiful × GPU friendly"。

package/references/skill-index.en.md

@@ -0,0 +1,106 @@
+# Skill Index: Lens Library, Activation Anchors, Design Translation Prototypes, Workflow Examples
+
+> This file is extracted from `../skills/math-research-activator/SKILL.en.md` for reference when the full catalogs are needed. SKILL.en.md retains condensed summaries with links to this file.
+
+## Language Routing & Mixed-Input Rules (Full Version)
+
+Language routing only determines "which language version to read" and "what language to respond in." It does NOT affect whether the math system triggers or the A/B/C/D/E scenario classification.
+
+### Decision Rules
+
+1. **Judge the natural-language frame first**
+   - If the user's sentence structure, verbs, and mood particles are primarily Chinese, treat as Chinese even if English technical terms are interspersed.
+   - E.g., "帮我 design 一个 attention" / "这个 loss 有没有理论问题" / "能不能用 manifold 做 routing" → Chinese.
+
+2. **English technical terms do not count as English primary language**
+   - attention, loss, routing, embedding, manifold, operator, kernel, KV-cache, transformer, MoE, and similar AI/math/engineering terms are domain terms and do not trigger a switch to English.
+
+3. **Code, paths, and formulas are excluded from language detection**
+   - File paths, function names, variable names, LaTeX formulas, and CLI arguments do not count toward language ratio.
+
+4. **When primary language is unclear, follow the user's last clear language**
+   - If the CN/EN ratio is close and indeterminate, use the user's most recent unambiguous natural language.
+   - If there is no prior context, default to Chinese.
+
+5. **Output language matches primary language**
+   - Chinese primary → read `../skills/math-research-activator/SKILL.md`, respond in Chinese, retaining necessary English terms.
+   - English primary → read `../skills/math-research-activator/SKILL.en.md`, respond in English.
+   - If the user explicitly requests "in English" / "in Chinese," follow the explicit request.
+
+## Lens Library (15 Mathematical Perspectives)
+
+Each lens answers: What is this perspective? What kinds of problems is it suited to diagnose? Which knowledge domains does it route to?
+
+| Lens | File | Core Perspective |
+|------|------|-----------------|
+| Axiomatization | `../lenses/axiomatization.en.md` | Examine consistency/independence/completeness of assumptions |
+| Duality | `../lenses/duality.en.md` | Transform to the dual space to expose constraints and invariants |
+| Symmetry | `../lenses/symmetry.en.md` | Invariants and conservation laws under transformations |
+| Spectral Decomposition | `../lenses/spectral.en.md` | Eigenvalues/singular values reveal dominant structure |
+| Geometric | `../lenses/geometric.en.md` | Metric/curvature/spatial structure on manifolds |
+| Projection & Decomposition | `../lenses/projection.en.md` | Orthogonal decomposition, subspace separation, conflict elimination |
+| Variational | `../lenses/variational.en.md` | Constrained extrema, energy minimization |
+| Local-to-Global | `../lenses/local-to-global.en.md` | Patching local properties into global ones, cohomological obstructions |
+| Topological | `../lenses/topological.en.md` | Invariants under continuous deformation, connectedness, holes |
+| Categorical | `../lenses/categorical.en.md` | Universal properties, functors, natural transformations |
+| Perturbation | `../lenses/perturbation.en.md` | Propagation of small perturbations, stability, robustness |
+| Causal | `../lenses/causal.en.md` | Correlation ≠ causation, interventions, counterfactuals |
+| Game-Theoretic | `../lenses/game.en.md` | Multi-agent strategic interaction, equilibria, mechanism design |
+| Probabilistic & Statistical | `../lenses/probabilistic.en.md` | Quantifying uncertainty, Bayesian updating |
+| Algorithmic | `../lenses/algorithmic.en.md` | Complexity, feasibility, parallelizability |
+
+## Activation Anchors (Organized by Mathematical Domain)
+
+Each anchor is not a closed knowledge card but answers: what math structure to activate, what deeper knowledge it connects to, what AI design actions it translates to, and how to extend when insufficient.
+
+| Domain | Directory | Anchors |
+|--------|-----------|----------------|
+| Matrix Analysis | `../knowledge-base/matrix-analysis/` | projection, spectral-decomposition, low-rank-approximation, positive-semidefinite, matrix-perturbation |
+| Optimization | `../knowledge-base/optimization/` | lagrangian-duality, convex-optimization, constrained-optimization, riemannian-optimization, proximal-method |
+| Differential Geometry | `../knowledge-base/differential-geometry/` | manifold, tangent-space, metric-tensor, geodesic, curvature, connection |
+| Lie Theory | `../knowledge-base/lie-theory/` | group-action, lie-group, lie-algebra, representation, equivariance |
+| Topology | `../knowledge-base/topology/` | persistent-homology, euler-characteristic, fundamental-group |
+| Probability & Information | `../knowledge-base/probability/` | concentration-inequality, entropy, kl-divergence, information-bottleneck, fisher-information |
+| Information Geometry | `../knowledge-base/information-geometry/` | natural-gradient, fisher-metric |
+
+## Design Pattern Library (Organized by AI Component)
+
+Each design pattern answers: Mathematical origin, AI module form, implementable architecture, GPU feasibility, paper-level exposition, risks.
+
+| Component Type | Directory | Patterns |
+|---------------|-----------|----------|
+| Attention | `../design-patterns/attention/` | projection-attention, spectral-attention, equivariant-attention, geometry-aware-attention, information-bottleneck-attention |
+| Loss Functions | `../design-patterns/loss/` | orthogonality-loss, contrastive-loss, variational-loss, information-bottleneck-loss, constraint-penalty |
+| Routing | `../design-patterns/routing/` | optimal-transport-routing, graph-routing, moe-routing, spectral-clustering-routing |
+| Representation | `../design-patterns/representation/` | shared-private-decomposition, manifold-representation, equivariant-split, subspace-alignment |
+| Compression | `../design-patterns/compression/` | low-rank-kv-cache, spectral-token-pruning, topology-preserving-compression, leverage-score-selection |
+
+## Workflow Example
+
+**User**: "Design a new KV Cache compression method that preserves long-range dependencies — I don't want to just do top-k."
+
+```
+Step 1 — Diagnosis: Scenario B (Mechanism Design)
+  Problem type: Sequence memory compression + information preservation + long-range structure
+  Core tension: Compressing token count vs. preserving long-range dependencies
+
+Step 2 — Lens Selection:
+  1. Spectral Decomposition (preserve dominant subspace)
+  2. Information Theory (retain states with maximum mutual information)
+  3. Topological (preserve critical connectivity points of sequential structure)
+
+Step 3 — Activation Anchors:
+  → low-rank-approximation (Matrix Analysis anchor)
+  → leverage-score-selection (Design Pattern: compression)
+  → information-bottleneck (Probability & Information anchor)
+  If existing anchors are insufficient, enter Knowledge Gap Protocol to generate a temporary knowledge card.
+
+Step 4 — Design Translation:
+  Candidate A: Spectral KV Compression (low-rank + leverage score)
+  Candidate B: Information-Preserving Cache (query sensitivity)
+  Candidate C: Topology-Preserving Cache (graph bridge-node retention)
+
+Step 5 — Critic Review:
+  A is most GPU-friendly; B requires estimating future queries with inherent uncertainty; C has prohibitive graph construction cost
+  Recommendation: Prioritize A; use B as a lightweight gate
+```

package/references/skill-index.md

@@ -0,0 +1,106 @@
+# Skill 索引：透镜库、激活锚点、设计翻译原型、工作流范例
+
+> 本文件从 `../skills/math-research-activator/SKILL.md` 提取，供需要查阅完整目录时使用。SKILL.md 保留精简摘要和指向本文件的链接。
+
+## 语言路由与混合输入规则（完整版）
+
+语言路由只决定"读取哪个语言版本的说明"和"最终用什么语言回答"，不参与数学系统是否触发、也不参与 A/B/C/D/E 场景判断。
+
+### 判定规则
+
+1. **先判断自然语言主框架**
+   - 如果用户的请求句式、动词、语气词主要是中文，即使夹杂英文技术词，也按中文处理。
+   - 例如："帮我 design 一个 attention""这个 loss 有没有理论问题""能不能用 manifold 做 routing"均按中文处理。
+
+2. **英文技术词不计入英文主语言**
+   - attention、loss、routing、embedding、manifold、operator、kernel、KV-cache、transformer、MoE 等 AI/数学/工程术语视为领域术语，不作为切换到英文的依据。
+
+3. **代码、路径、公式不参与语言判定**
+   - 文件路径、函数名、变量名、LaTeX 公式、命令行参数不计入语言比例。
+
+4. **主语言不明显时，沿用用户上一轮主要语言**
+   - 若中英文比例接近且无法判断，以用户最近一次明确使用的自然语言为准。
+   - 若没有上下文，默认中文。
+
+5. **输出语言与主语言一致**
+   - 中文主语言 → 读取中文 `../skills/math-research-activator/SKILL.md`，用中文回答，保留必要英文术语。
+   - 英文主语言 → 读取 `../skills/math-research-activator/SKILL.en.md`，用英文回答。
+   - 用户明确要求"用英文/用中文"时，以用户显式要求为准。
+
+## 透镜库（15 个数学视角）
+
+每个透镜回答：这是什么视角？适合诊断什么问题？会路由到哪些知识域？
+
+| 透镜 | 文件 | 核心视角 |
+|------|------|---------|
+| 公理化 | `../lenses/axiomatization.md` | 审查假设的相容性/独立性/完备性 |
+| 对偶 | `../lenses/duality.md` | 转换到对偶空间暴露约束与不变量 |
+| 对称性 | `../lenses/symmetry.md` | 变换下的不变量与守恒律 |
+| 谱分解 | `../lenses/spectral.md` | 特征值/奇异值揭示主导结构 |
+| 几何 | `../lenses/geometric.md` | 度量/曲率/流形上的空间结构 |
+| 投影与分解 | `../lenses/projection.md` | 正交分解、子空间分离、冲突消除 |
+| 变分 | `../lenses/variational.md` | 约束下极值、能量最小化 |
+| 局部到整体 | `../lenses/local-to-global.md` | 局部性质拼接为全局、层上同调障碍 |
+| 拓扑 | `../lenses/topological.md` | 连续变形不变量、连通性、空洞 |
+| 范畴化 | `../lenses/categorical.md` | 泛性质、函子、自然变换 |
+| 扰动 | `../lenses/perturbation.md` | 小扰动的传播、稳定性、鲁棒性 |
+| 因果 | `../lenses/causal.md` | 相关≠因果、干预、反事实 |
+| 博弈 | `../lenses/game.md` | 多方策略互动、均衡、机制设计 |
+| 概率统计 | `../lenses/probabilistic.md` | 量化不确定性、贝叶斯更新 |
+| 算法 | `../lenses/algorithmic.md` | 复杂度、可行性、并行性 |
+
+## 激活锚点（按数学领域组织）
+
+每个锚点不是封闭知识卡，而是回答：激活什么数学结构、连接哪些更深知识、可翻译成哪些 AI 设计动作、不足时如何扩展。
+
+| 领域 | 目录 | 锚点 |
+|------|------|---------|
+| 矩阵分析 | `../knowledge-base/matrix-analysis/` | projection, spectral-decomposition, low-rank-approximation, positive-semidefinite, matrix-perturbation |
+| 最优化 | `../knowledge-base/optimization/` | lagrangian-duality, convex-optimization, constrained-optimization, riemannian-optimization, proximal-method |
+| 微分几何 | `../knowledge-base/differential-geometry/` | manifold, tangent-space, metric-tensor, geodesic, curvature, connection |
+| 李理论 | `../knowledge-base/lie-theory/` | group-action, lie-group, lie-algebra, representation, equivariance |
+| 拓扑 | `../knowledge-base/topology/` | persistent-homology, euler-characteristic, fundamental-group |
+| 概率与信息 | `../knowledge-base/probability/` | concentration-inequality, entropy, kl-divergence, information-bottleneck, fisher-information |
+| 信息几何 | `../knowledge-base/information-geometry/` | natural-gradient, fisher-metric |
+
+## 设计模式库（按 AI 组件组织）
+
+每个设计模式回答：数学来源、AI 模块形式、可实现结构、GPU 可行性、论文表述、风险。
+
+| 组件类型 | 目录 | 模式 |
+|---------|------|------|
+| 注意力 | `../design-patterns/attention/` | projection-attention, spectral-attention, equivariant-attention, geometry-aware-attention, information-bottleneck-attention |
+| 损失函数 | `../design-patterns/loss/` | orthogonality-loss, contrastive-loss, variational-loss, information-bottleneck-loss, constraint-penalty |
+| 路由 | `../design-patterns/routing/` | optimal-transport-routing, graph-routing, moe-routing, spectral-clustering-routing |
+| 表示 | `../design-patterns/representation/` | shared-private-decomposition, manifold-representation, equivariant-split, subspace-alignment |
+| 压缩 | `../design-patterns/compression/` | low-rank-kv-cache, spectral-token-pruning, topology-preserving-compression, leverage-score-selection |
+
+## 工作流范例
+
+**用户**："设计新的 KV Cache 压缩方法，保留长期依赖，不想只做 top-k"
+
+```
+第一步 诊断：场景 B（机制设计）
+  问题类型：序列记忆压缩 + 信息保留 + 长程结构
+  核心张力：压缩 token 数量 vs 不破坏长期依赖
+
+第二步 透镜选择：
+  1. 谱分解（保留主导子空间）
+  2. 信息论（保留最大互信息状态）
+  3. 拓扑（保留序列结构关键连接点）
+
+第三步 激活锚点：
+  → low-rank-approximation（矩阵分析锚点）
+  → leverage-score-selection（压缩设计模式）
+  → information-bottleneck（概率与信息锚点）
+  若现有锚点不足，进入 Knowledge Gap Protocol 生成临时知识卡。
+
+第四步 设计翻译：
+  候选 A：Spectral KV Compression（低秩 + leverage score）
+  候选 B：Information-Preserving Cache（query sensitivity）
+  候选 C：Topology-Preserving Cache（图桥接节点保留）
+
+第五步 Critic 审查：
+  A 最 GPU 友好，B 需估计未来 query 有不确定性，C 图构建成本过高
+  建议：优先 A，B 作轻量 gate
+```