@kernaut/kernaut-resources 0.1.17-beta.1 → 0.1.18

package/resources/kop-2026-10-23-0.1.17-beta.1/product_capabilities/kernel_optflow/skills/KernelWiki-private/governance/specs/paths.md

@@ -1,60 +0,0 @@
-# Knowledge paths and ownership
-
-This specification defines the adopted destination; see [deployment status](../README.md#adoption-and-deployment) before writing to the current corpus.
-
-## Body ownership
-
-```text
-knowledge/<vendor>/<technology-or-shared>/<profile>/[<topic>/...]<slug>.md
-```
-
-Vendor separates implementations with substantially different architectures. Technology separates DSL/API semantics; `shared` contains genuinely technology-independent knowledge within that vendor. Product generations use `platforms`, not repeated DSL trees. Ascend's current construction focus is AscendC and Triton; retained material for other stacks keeps its original support and evidence boundaries.
-
-A card has one body owner. Topic indexes may link multiple Profiles; a source batch, date, compiler release or migration run is provenance, not a new permanent content taxonomy. Registered future vendors or technologies do not require empty directory trees.
-
-`generic/shared/` uses the [generic admission policy](generic-bucket-budget.md). Unknown ownership is not generic. Vendor-independent applicability must be justified; identical DSL names across vendors do not imply identical semantics.
-
-## Card Profiles
-
-The following mappings are the adopted authoring contract. They must be encoded in the shared machine declarations before new-format production is enabled.
-
-| Path Profile | OKF `type` | Owns | Excludes |
-| --- | --- | --- | --- |
-| `concepts` | `Concept` | Hardware facts, execution and programming models | API argument references or diagnostic procedures |
-| `apis` | `API` | Exact interface, signature, semantics and support conditions | Whole operator implementations |
-| `examples` | `Example` | Minimal examples with inputs, outputs and validation | Uncontextualized snippets |
-| `guides` | `Guide` | A development task, its reasoning, steps and checks | A pure directory listing |
-| `operators` | `Operator` | An operator's semantics, host/tiling/kernel and dependencies | A general method without an operator implementation |
-| `optimizations` | `Optimization` | Reusable mechanism, trigger, costs, counterexamples and checks | Unattributed single-run speedups |
-| `runbooks` | `Runbook` | Symptoms/tasks, discriminating checks, actions and exit criteria | Background explanation without an actionable procedure |
-| `interoperability` | `Interoperability` | Explicit source/target technology mapping, under `shared` | Assumed equivalence based on similar names |
-| `glossaries` | `Glossary` | Terms, definitions and aliases | Tutorials or measurement records |
-| `experiences` | `Experience` | A scoped experiment, observations and durable evidence | An unsupported universal recommendation |
-
-A legacy `type: skill` technical page is knowledge, not automatically an agent workflow. Classify its actual content. Reusable methods link to experiments without erasing the experiments' absolute measurements, negative results or conditions.
-
-## Names, identity and navigation
-
-New card/topic slugs use lowercase `snake_case`; registered vendor/technology identifiers retain their registered spelling. Collection Profiles are plural. Legacy filenames remain resolvable until checked migration maps and consumers are deployed.
-
-The Concept ID for `knowledge/ascend/shared/concepts/resource_model.md` is `ascend/shared/concepts/resource_model`. It is not a legacy `id` or `knowledge_id`. A move changes Concept identity: record the mapping, repair links and confirm old lookup behavior before cutover. Splits and merges identify all replacement targets; do not resolve an old ID to an arbitrary surviving fragment.
-
-Card links use ordinary relative Markdown paths or paths beginning with `/` resolved from the Bundle root. Preserve link meaning when moving cards.
-
-Each materialized Bundle directory has a generated thin `index.md` listing direct children and useful summaries. It is navigation, excluded from card search and card-link graph nodes. A topic aggregation with substantive technical conclusions is a named card, not a disguised index. `log.md` remains reserved for history and is not used as a card filename.
-
-## Outside the Bundle
-
-| Location | Owns |
-| --- | --- |
-| `sources/` | Normalized source records and original source identity |
-| `corpus/` | Original material or a pinned, recoverable source description |
-| `artifacts/` | Durable source code, workloads, measurements and provenance |
-| `queries/` | Generated retrieval/navigation views, including platform views |
-| `governance/` | Construction rules and their implementation inventory |
-| `agent/` | Existing producers and their workflow-specific references |
-| `docs/` | Current operation/design explanations; completed plans under `history/` |
-| `provenance/` | Migration disposition, release receipts and historical inventories |
-| Ignored work/cache directories | Rebuildable outputs and transient task state |
-
-The existing [production-tool inventory](../../agent/README.md) remains authoritative about which workflows are current. Moving a producer's files requires dependency adaptation. Governance and historical plans do not enter kernel-knowledge search; evidence remains explicitly distinguishable from synthesized knowledge.

package/resources/kop-2026-10-23-0.1.17-beta.1/product_capabilities/kernel_optflow/skills/KernelWiki-private/governance/specs/schemas.md

@@ -1,35 +0,0 @@
-# Schema, registry and contract maintenance
-
-Governance adopts CANNBot's separation of responsibilities while reusing KernelWiki's deployed authorities. The `schemas/` and `contracts/` directories currently index existing declarations and implementations; they do not contain duplicate copies or claim to be a new executable checker.
-
-## One owner per rule
-
-| Layer | Owns | Must not become |
-| --- | --- | --- |
-| Schema / Profile declaration | Field shape, required combinations and path-to-type mapping | Independent copies in every producer template |
-| Registry | Canonical vendors, technologies, platforms, aliases, tags and source identities | A list inferred separately by each consumer |
-| Contract implementation | Path/metadata consistency, identities, links, cross-file evidence and compatibility | A second taxonomy hardcoded beside the declarations |
-| Spec | Meaning, rationale, examples, deployment and migration boundaries | A claim that an unimplemented rule is already enforced |
-
-Current declaration owners are listed in [schemas](../schemas/README.md); current implementations are listed in [contracts](../contracts/README.md). Existing declarations and code are not perfectly factored. Reconciliation must update the actual consumers rather than choosing whichever old document makes a proposed write pass.
-
-Profile, platform and OKF checking must be integrated into these shared authorities during the migration pilot. A later coordinated relocation into `governance/` must update imports, script entry points, producer inputs, tests and documentation together. Keep one public validation entry point; do not keep an old and a new full validator with divergent rules.
-
-## Shared consumer obligations
-
-Producers validate candidates against the same declarations before publication; query and lint agree on classification, applicable scope, source identity and navigation exclusion. Generated indexes and graphs are derived, never additional hand-maintained facts. A legacy adapter must report unresolved metadata and preserve old lookups rather than silently widening applicability.
-
-Strict Profile rejection and general OKF reading are different contracts. In particular, unknown optional extensions in an imported Bundle are not automatically invalid OKF; our producer may still need an explicit conversion before accepting that card into KernelWiki. A failing changed card cannot silently cause unrelated knowledge to disappear from search. Define and test batch-level rejection/reporting before switching consumers.
-
-## Change checklist
-
-For a rule change, record the affected scope and complete the relevant items:
-
-1. State the engineering need, existing owner, compatibility effect and deployment stage. Add a field or registry entry only for a real card or concrete consumer.
-2. Update the authoritative declaration and all affected shared consumers. Do not maintain a separate enum in a template, prompt or lint helper.
-3. Add meaningful accepting and rejecting examples for executable behavior: valid scope versus wrong vendor/platform; real source versus broken reference; valid identity versus conflicting mapping. Documentation-only clarification needs link and consistency checks, not synthetic implementation-mirroring tests.
-4. For moves, splits or merges, check old IDs, new Concept IDs, links, navigation, source/artifact delivery and duplicate retrieval. Preserve technical content and evidence identity unless the change explicitly reviews them.
-5. Run the affected existing checks and the pilot's retrieval/read regression. Report pre-existing failures separately; do not weaken limits or hide baseline failures to claim success.
-6. Record migration disposition and the release's content/tool revisions. Switch writers/readers only for validated batches; retain a reproducible prior revision for rollback.
-
-Review covers the evidence and applicability that machine checks cannot establish. Record the concrete candidate, disposition and supporting checks in Git so that maintenance and publication decisions remain reviewable.

package/resources/kop-2026-10-23-0.1.17-beta.1/product_capabilities/kernel_optflow/skills/KernelWiki-private/index.md

@@ -1,100 +0,0 @@
-# GPU Kernel Optimization Knowledge Base
-
-> Multi-vendor knowledge base for GPU kernel optimization. Currently covering NVIDIA Blackwell (SM100), Hopper (SM90), and general CUDA.
-> Optimized for LLM agent retrieval. See [CLAUDE.md](CLAUDE.md) for schema and conventions.
-> **For Claude Code agents**: this repository is a Claude Code skill — see [SKILL.md](SKILL.md).
-
-## Recommended Query Tools (for LLM agents)
-
-```bash
-python3 scripts/query.py "<natural language>" [--tag <t>] [--type <kernel|technique|pr|...>]
-python3 scripts/get_page.py <page-id-or-path> [--follow-sources]
-python3 scripts/grep_wiki.py "<regex>" [--only wiki|sources]
-```
-
-See [references/examples.md](references/examples.md) for 10 worked query patterns.
-
-## Quick Navigation
-
-| I want to... | Go to |
-|---|---|
-| Fix a performance problem | [queries/by-problem.md](queries/by-problem.md) |
-| Learn a specific technique | [queries/by-technique.md](queries/by-technique.md) |
-| Use a hardware feature | [queries/by-hardware-feature.md](queries/by-hardware-feature.md) |
-| See what a repo contributed | [queries/by-repo.md](queries/by-repo.md) |
-| Write a specific kernel type | [queries/by-kernel-type.md](queries/by-kernel-type.md) |
-| Use a specific language/DSL | [queries/by-language.md](queries/by-language.md) |
-| Browse by vendor | [queries/by-vendor.md](queries/by-vendor.md) |
-| Learn a foundational skill | `wiki/nvidia/foundations/` |
-| Route an operator | `wiki/nvidia/operator-routing/` |
-| Look up an API | `wiki/nvidia/api-definitions/` |
-
-## Hardware Features
-
-- [hw-tcgen05-mma](wiki/nvidia/hardware/tcgen05-mma.md) — Blackwell MMA instruction (replaces wgmma)
-- [hw-tmem](wiki/nvidia/hardware/tmem.md) — Tensor Memory (256KB dedicated accumulator storage)
-- [hw-clc](wiki/nvidia/hardware/clc.md) — Cluster Launch Control (dynamic tile scheduling)
-- [hw-tma](wiki/nvidia/hardware/tma.md) — Tensor Memory Accelerator (async bulk loads)
-- [hw-2sm-cooperative](wiki/nvidia/hardware/2sm-cooperative.md) — Two-SM cooperative MMA
-- [hw-nvfp4](wiki/nvidia/hardware/nvfp4.md) — NVFP4 and block-scaled narrow precision
-- [hw-pdl-gdc](wiki/nvidia/hardware/pdl-gdc.md) — Programmatic Dependent Launch / Grid Dependency Control
-
-## Optimization Techniques
-
-- [technique-warp-specialization](wiki/nvidia/techniques/warp-specialization.md) — Warp role assignment
-- [technique-persistent-kernels](wiki/nvidia/techniques/persistent-kernels.md) — Persistent kernel patterns with CLC
-- [technique-swizzling](wiki/nvidia/techniques/swizzling.md) — Shared memory swizzling
-- [technique-pipeline-stages](wiki/nvidia/techniques/pipeline-stages.md) — Software pipelining
-- [technique-epilogue-fusion](wiki/nvidia/techniques/epilogue-fusion.md) — Fusing epilogue with mainloop
-- [technique-tile-scheduling](wiki/nvidia/techniques/tile-scheduling.md) — Tile scheduling strategies
-- [technique-double-buffering](wiki/nvidia/techniques/double-buffering.md) — Double/multi-buffering
-- [technique-software-exp](wiki/nvidia/techniques/software-exp.md) — Software-emulated exponential
-- [technique-fine-grained-quant](wiki/nvidia/techniques/fine-grained-quantization.md) — Fine-grained FP8/FP4 quantization
-- [technique-vectorized-loads](wiki/nvidia/techniques/vectorized-loads.md) — Wide vectorized loads and cache policies
-
-## Kernel Case Studies
-
-- [kernel-flash-attention-4](wiki/nvidia/kernels/flash-attention-4.md) — FlashAttention-4 (1605 TFLOPS on B200)
-- [kernel-deepgemm](wiki/nvidia/kernels/deepgemm.md) — DeepGEMM FP8 GEMM (1550 TFLOPS on H800)
-- [kernel-flashmla](wiki/nvidia/kernels/flashmla.md) — FlashMLA sparse/dense MLA decoding
-- [kernel-nsa](wiki/nvidia/kernels/nsa.md) — Native Sparse Attention (9x fwd speedup)
-- [kernel-gated-delta-net](wiki/nvidia/kernels/gated-delta-net.md) — Gated Delta Net linear attention
-- [kernel-nvfp4-gemm](wiki/nvidia/kernels/nvfp4-gemm.md) — NVFP4 GEMM from GPU Mode hackathon
-- [kernel-nvfp4-gemv](wiki/nvidia/kernels/nvfp4-gemv.md) — NVFP4 batched GEMV optimization
-- [kernel-grouped-gemm](wiki/nvidia/kernels/grouped-gemm.md) — Grouped GEMM for MoE
-- [kernel-fused-moe](wiki/nvidia/kernels/fused-moe.md) — Fused MoE with FP8
-
-## Problem → Solution Patterns
-
-- [pattern-low-sm-utilization](wiki/nvidia/patterns/low-sm-utilization.md) — SM utilization is low
-- [pattern-memory-bound](wiki/nvidia/patterns/memory-bound.md) — Kernel is memory bandwidth limited
-- [pattern-register-pressure](wiki/nvidia/patterns/register-pressure.md) — Too many registers → low occupancy
-- [pattern-compute-bound](wiki/nvidia/patterns/compute-bound.md) — Not reaching peak FLOPS
-- [pattern-tail-effect](wiki/nvidia/patterns/tail-effect.md) — Last wave underutilizes GPU
-
-## Languages & DSLs
-
-- [lang-cute-dsl](wiki/nvidia/languages/cute-dsl.md) — CuTe DSL for Blackwell
-- [lang-cuda-cpp](wiki/nvidia/languages/cuda-cpp.md) — CUDA C++ with PTX inline
-- [lang-ptx](wiki/nvidia/languages/ptx-sm100.md) — PTX instructions for SM100
-- [lang-triton](wiki/nvidia/languages/triton-blackwell.md) — Triton on Blackwell
-
-## Migration Guides
-
-- [migration-wgmma-to-tcgen05](wiki/nvidia/migration/wgmma-to-tcgen05.md) — Hopper wgmma → Blackwell tcgen05
-- [migration-register-to-tmem](wiki/nvidia/migration/register-to-tmem.md) — Register accumulators → TMEM
-
-## Source Repositories
-
-| Repository | Focus |
-|---|---|
-| [NVIDIA/cutlass](queries/by-repo.md#nvidiacutlass) | CUTLASS 4.x Blackwell support |
-| [sgl-project/sglang](queries/by-repo.md#sgl-projectsglang) | SGLang Blackwell integration |
-| [vllm-project/vllm](queries/by-repo.md#vllm-projectvllm) | vLLM Blackwell support |
-| [flashinfer-ai/flashinfer](queries/by-repo.md#flashinfer-aiflashinfer) | FlashInfer Blackwell kernels |
-| [pytorch/pytorch](queries/by-repo.md#pytorchpytorch) | PyTorch/Inductor Blackwell |
-
-## Competitions
-
-- [GPU Mode NVFP4 Hackathon](sources/contests/gpu-mode-nvfp4/) — 4 NVFP4 kernel challenges on B200
-- [FlashInfer MLSys 2026](sources/contests/flashinfer-mlsys26/) — MoE, Sparse Attention, GatedDeltaNet

package/resources/kop-2026-10-23-0.1.17-beta.1/product_capabilities/kernel_optflow/skills/KernelWiki-private/provenance/README.md

@@ -1,10 +0,0 @@
-# Provenance and historical evidence
-
-- [Original kernel-kb-mvp migration inventory](history/kernel-kb-mvp/migration_inventory.tsv) and [deletion ledger](history/kernel-kb-mvp/deletion_ledger.tsv): immutable historical rows; their recorded paths describe the original migration.
-- [2026-10-10 ATB delivery receipt](releases/contract-atb/2026-10-10-validation.json): a fixed batch result, not current configuration or a new validation run.
-- [2026-10-20 experience page naming](migrations/20261020-experience-page-naming.json): semantic page names and compatible old ID/path mappings; historical corpus snapshots remain unchanged.
-- [2026-10-20 organization ledger](migrations/20261020-organization-cleanup.json): old/new paths, hashes, reasons and deferred material for this cleanup.
-
-New invocations of `scripts/migrate_kb_mvp.py` write their default inventory under `provenance/migrations/kernel-kb-mvp/`, separate from the preserved historical inventory. An explicit `--output-tsv` still selects another destination.
-
-This tree stores small audit metadata. Source material and runtime evidence remain in `corpus/` and `artifacts/`; they have not been reclassified as disposable cache. The active DoD fixtures, size budget and plan-body baselines remain at their existing `data/` paths because scripts still consume them.

package/resources/kop-2026-10-23-0.1.17-beta.1/product_capabilities/kernel_optflow/skills/KernelWiki-private/provenance/cannbench-branch-review-20261020/README.md

@@ -1,110 +0,0 @@
-# CANNBench 同事分支知识审查（2026-10-20）
-
-本次从已合并此前 PR 的 KernelWiki main `edca1de2` 开始，审查 KOP `feature/ascendc_harry_910c` 的固定提交 `73923adac6a6770284086e749443ab3ac066cfc7`。当前审查分支为 `feat/cannbench-branch-experience-review-20261020`。
-
-## 结论与入库边界
-
-不是把整个同事分支当成更新版本：57 份顶层文档中 50 份已有保存副本，很多副本还包含更晚的追加记录。本文区分“源资料已保存但未提炼为直接可查的机制”“新增结构化资格证据”“已由较新结论覆盖”和“历史辅助材料”。逐机制对比 57 份顶层文档，并清点 15 份早期 AAP3D/Prolog lane 文档、43 份 compact evidence 文件和 best 树的 Git 对象；嵌套 lane 不作全量重新归因；没有重新分析全部原始会话或运行完整优化任务。
-
-新增 7 个有明确适用条件的机制/反例页；向 5 个已有机制页追加条件性负结果或最终判决，另为 Cummin 主入口增加不同机制的跳转。旧页 ID、正文和旧 corpus 均保留，53 条公开分数/通过数/原始 chronicle 链接逐项不变。
-
-| 内容 | 处理 | 对当前公开集的意义 |
-|---|---|---|
-| MhcSinkhorn MH802S | 新增周期索引播种页；补上最终源绑定及 CSV 舍入订正 | 7 条公开路由命中，本地全20收益；代理分不是新增线上分 |
-| ResizeBilinear RB803 | 新增 Axpy 指令开销页，保留非 FMA 证据和继承精度债 | case13 本地快；公开120/120与隐藏失败分别记录，不要求隐藏评测 |
-| RoPE AP803 / Foreach FS802 / Cummin CM801 | 新增有源码和统计边界的局部域经验页 | **当前公开20例零新路由命中，收益不能迁移为刷公开的预计加分** |
-| SwiGlu SG901R / NMS NSM801 | 新增全任务负收益或局部 screen/oracle 反例 | 避免将 micro收益、FP32直觉或局部正确性当作全量收益 |
-| GridSampler3D / Unique / TopK | 向原页追加有版本区分的负结果和证据 | 不把局部 NO_GO 泛化为技术死路；UN801 与 UN901 分开 |
-| Conv3D / StridedSlice | 在原页补最终 CPU 判决或局部改动条件 | pending 不覆盖终态；CPU证明不冒充NPU实测 |
-
-## 新旧与冲突判定
-
-1. 导入时间、Git 最后修改时间、实验发生时间分别保留。文件更长或 commit 更晚不证明性能结论更新；逐机制、dtype/shape、设备、基线包和测量口径比较。
-2. 保留已审核的 2026-10-16 公开 API 参考。`kernels_best_910c/provenance.json` 的 August-era `official` 数字缺少足以替换该快照的侧别/时间/包绑定；不按更大数字覆盖。
-3. TopK 的 branch canonical 是 GT13-r2 隐藏修复版，自己的公开62.67000581低于 GT12 公开62.69159418。canonical 更晚不等于公开更优；不覆盖 GT12。
-4. 已合格但 held、局部 screen 负、CPU-only、预测值和实际线上分分开。维护公告及旧认领/提交流程只描述当时状态，不是当前执行指令。
-5. 公开与隐藏分开；隐藏失败不是否定独立公开通过的充分理由，也不为公开工作流添加隐藏资格门。真实语义/源完整性债继续保留。
-
-## 证据与验证
-
-所有新增经验仍为 `source-reported`，没有声明独立设备复测或人工背书。源 Git blob、原始输入 SHA256、完整源码成员哈希在 [inventory](../../corpus/ascend/cannbench/branch-review-20261020/inventory.json) 与各 `source-members/*/ARCHIVE.json`。四个候选包按原始档案哈希核对后仅保留相关源码文本，编译产物、ZIP、PT 和 profiler 原始大文件不重复入库。MH802S 的辅助审计明确来自已推送的 KOP checkpoint `00d90d029`，没有把残留未跟踪目录误当同事分支的内容。
-
-离线验证将核对证据身份、公开参考及原有资料不变，并重算保存下来的计时样本/汇总。其通过只证明内部一致性，不升级为独立 NPU 复测。验证结果见 `verification.json`。已有全库10个 ATB 资产错误作基线比较，不将其混作本次新增错误。
-
-## 逐文档处置
-
-完整数据见 [document-comparison.json](document-comparison.json)。下表列出所有顶层文档；diff行数只辅助定位，不用于推断知识更新。
-
-| 文档 | branch行数 | 相对已有源新增/移除 | 处置 |
-|---|---:|---:|---|
-| README | 226 | +0 / −0 | covered-or-historical-auxiliary |
-| aap3d_harness_REPORT | 370 | +0 / −0 | covered-or-historical-auxiliary |
-| adaptive_avg_pool_3d | 1122 | +0 / −24 | retain-existing-no-replacement |
-| add_rms_norm_dynamic_quant | 242 | +6 / −109 | retain-existing-no-replacement |
-| apply_adam_w | 233 | +0 / −34 | retain-existing-no-replacement |
-| apply_rotary_pos_emb | 304 | +1 / −6 | new-queryable-mechanism-without-source-replacement |
-| arg_max | 122 | +0 / −5 | retain-existing-no-replacement |
-| conv_2d | 1121 | +1014 / −467 | retain-existing-no-replacement |
-| conv_3d_backprop_filter | 2735 | +122 / −591 | append-evidence-to-existing-owner |
-| conv_3d_backprop_filter_d3_spec | 413 | +0 / −0 | covered-or-historical-auxiliary |
-| cross_entropy_loss | 338 | +1 / −72 | retain-existing-no-replacement |
-| cummin | 4259 | +18 / −930 | new-queryable-mechanism-without-source-replacement |
-| depthwise_conv_2d | 212 | +131 / −999 | retain-existing-no-replacement |
-| dequant_swiglu_quant | 3015 | +0 / −30 | retain-existing-no-replacement |
-| dilation_2d | 244 | +1 / −58 | retain-existing-no-replacement |
-| dynamic_quant | 27 | +20 / −405 | retain-existing-no-replacement |
-| engram_gate_fusion | 545 | +380 / −1144 | retain-existing-no-replacement |
-| exp | 93 | +0 / −5 | retain-existing-no-replacement |
-| foreach_addcdiv_scalar | 81 | +0 / −61 | new-queryable-mechanism-without-source-replacement |
-| foreach_norm | 256 | +0 / −55 | retain-existing-no-replacement |
-| gather | 452 | +206 / −1022 | retain-existing-no-replacement |
-| gcd | 490 | +0 / −145 | retain-existing-no-replacement |
-| gelu | 34 | +7 / −619 | retain-existing-no-replacement |
-| gqa | 1702 | +1643 / −105 | retain-existing-no-replacement |
-| grid_sampler_3d | 2594 | +18 / −230 | append-evidence-to-existing-owner |
-| group_norm | 552 | +20 / −914 | retain-existing-no-replacement |
-| grouped_matmul | 2853 | +2746 / −963 | retain-existing-no-replacement |
-| gru | 699 | +0 / −43 | retain-existing-no-replacement |
-| lstm | 2561 | +2 / −347 | retain-existing-no-replacement |
-| lstm_archive | 5870 | +0 / −0 | covered-or-historical-auxiliary |
-| maximum | 127 | +0 / −98 | retain-existing-no-replacement |
-| mha | 278 | +257 / −480 | retain-existing-no-replacement |
-| mhc_sinkhorn | 36 | +17 / −126 | new-queryable-mechanism-without-source-replacement |
-| mish | 208 | +1 / −95 | retain-existing-no-replacement |
-| mla | 508 | +0 / −5 | retain-existing-no-replacement |
-| mla_prolog | 2106 | +91 / −444 | retain-existing-no-replacement |
-| moe_finalize_routing | 123 | +0 / −5 | retain-existing-no-replacement |
-| moe_gating_top_k_softmax | 1721 | +1626 / −387 | retain-existing-no-replacement |
-| moe_re_routing | 70 | +0 / −20 | retain-existing-no-replacement |
-| nms | 440 | +18 / −98 | new-queryable-mechanism-without-source-replacement |
-| opt6_low_yield_lanes | 107 | +0 / −0 | covered-or-historical-auxiliary |
-| quant_matmul | 200 | +0 / −10 | retain-existing-no-replacement |
-| resize_bilinear | 630 | +2 / −73 | new-queryable-mechanism-without-source-replacement |
-| rms_norm | 356 | +135 / −206 | retain-existing-no-replacement |
-| roi_align | 3390 | +2690 / −549 | retain-existing-no-replacement |
-| scatter | 1043 | +735 / −918 | retain-existing-no-replacement |
-| softmax | 642 | +0 / −43 | retain-existing-no-replacement |
-| softmax_strided_tiling | 80 | +0 / −0 | covered-or-historical-auxiliary |
-| sparse_flash_attention | 1760 | +556 / −647 | retain-existing-no-replacement |
-| strided_slice | 251 | +9 / −85 | append-evidence-to-existing-owner |
-| swi_glu | 189 | +8 / −109 | new-queryable-mechanism-without-source-replacement |
-| top_k | 530 | +114 / −674 | append-evidence-to-existing-owner |
-| transpose | 405 | +1 / −33 | retain-existing-no-replacement |
-| unique | 1866 | +1 / −245 | append-evidence-to-existing-owner |
-| unsorted_segment_sum | 333 | +1 / −6 | retain-existing-no-replacement |
-| unsorted_segment_sum_small_output | 179 | +0 / −0 | covered-or-historical-auxiliary |
-| weight_quant_batch_matmul | 1178 | +2 / −19 | retain-existing-no-replacement |
-
-## 验证结果
-
-- 44 份选入原文的 SHA256、4 个原始源码包的哈希绑定和 51 份选定源码成员通过核对。
-- 对保留的 1,096 组 kernel 中位数（5,480 个计时样本）按原报告 0.01us 分辨率复核；AP803/FS802/CM801 的完整 family/有符号统计重算一致。此处是 CPU 内部一致性检查，不是新的设备复测。
-- 从 MH802S 最终源中读取的周期权重经 1,024 个索引字、512 个逆置换检查通过。
-- 53 条公开参考、97 个既有 corpus 文件及 6 个原页面的完整正文/既有 metadata 值保持不变。
-- 7 条新增机制查询和 8 条原有机制回归查询均命中目标第1位；这证明本组检索样例可用，不代表所有问题的召回率或完整优化任务收益。
-- 7 页真实 CLI 直接读取与 source-following 检查通过；实际 query CLI 命中目标。
-- 全库 validator 仍为原有 10 个 ATB 资产错误，新增错误0；7 条新增 unknown SDK 警告明确保留，未伪造官方/本地相同的 CANN 补丁版本。
-
-TopK 源身份另按 11 份实际 Git 文件与 promotion 记录逐个 SHA256 核对，均匹配 GT13-r2；GT12 公开赢家的 `job_ea9753e9c561` 与 ZIP `3426273614e550cc9ecd21b5acfd91253e7ed61aea6fe884aef801eb9b6e9e48` 单列，见 `topk-source-identity.json`。这补充了原先公开分数快照自身没有的源码包身份。
-
-源格式边界：原文及源码成员中的 32 处历史空白提示（patch上下文空格、源码末尾空行）按原字节/哈希保留；逐路径排除这些原始来源后，新增撰写内容的 whitespace 检查通过。

package/resources/kop-2026-10-23-0.1.17-beta.1/product_capabilities/kernel_optflow/skills/KernelWiki-private/provenance/cannbench-branch-review-20261020/direct-reader-checks.json

@@ -1,47 +0,0 @@
-{
-  "reads": [
-    {
-      "path": "wiki/ascend/experience/mhc-sinkhorn-periodic-index-seeding.md",
-      "returncode": 0,
-      "returned_chars": 9519,
-      "source_page_returned": true
-    },
-    {
-      "path": "wiki/ascend/experience/resize-bilinear-axpy-issue-cost.md",
-      "returncode": 0,
-      "returned_chars": 6136,
-      "source_page_returned": true
-    },
-    {
-      "path": "wiki/ascend/experience/rope-row-stream-stride-overflow.md",
-      "returncode": 0,
-      "returned_chars": 10366,
-      "source_page_returned": true
-    },
-    {
-      "path": "wiki/ascend/experience/foreach-addcdiv-single-tile-queue.md",
-      "returncode": 0,
-      "returned_chars": 10457,
-      "source_page_returned": true
-    },
-    {
-      "path": "wiki/ascend/experience/cummin-axis-one-bitcopy.md",
-      "returncode": 0,
-      "returned_chars": 10363,
-      "source_page_returned": true
-    },
-    {
-      "path": "wiki/ascend/experience/swiglu-reciprocal-full-task-regression.md",
-      "returncode": 0,
-      "returned_chars": 5423,
-      "source_page_returned": true
-    },
-    {
-      "path": "wiki/ascend/experience/nms-prefetch-and-oracle-precision.md",
-      "returncode": 0,
-      "returned_chars": 5369,
-      "source_page_returned": true
-    }
-  ],
-  "actual_query_cli_target_found": true
-}