@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/tasks/build-vectorized-access.yaml

@@ -1,50 +0,0 @@
-task_id: 2026-04-16-vectorized-access
-task_type: build-skill
-target_path: wiki/nvidia/foundations/memory/vectorized-access/
-
-hardware:
-  device: H200
-  sm: "9.0a"
-  cuda_runtime: "12.9"
-
-upstream_scope:
-  - <local-path>
-  - <local-path> Programming Guides/cuda-c-best-practices-guide/
-  - <local-path>
-
-references:
-  - reasoning/api-probing.md
-  - reasoning/hardware-microbench.md
-  - reasoning/benchmark-protocol.md
-  - reasoning/bottleneck-triage.md
-
-success_criteria:
-  - file_exists: skill.md
-  - file_exists: pitfalls.md
-  - file_exists: apis.md
-  - lint_clean: true
-  - at_least_one_hw_probe_linked: true
-
-related_skills: [coalescing, bank-conflict]
-
-notes: |
-  Cover vectorized memory access: using float4/int4/uint4 vector types to
-  issue 128-bit wide loads/stores instead of scalar 32-bit ones, achieving
-  higher memory throughput on memory-bound kernels.
-
-  Key topics:
-  - float4, float2, int4, uint4, longlong2 vector types
-  - reinterpret_cast<float4*> pattern for type-punning arrays
-  - Alignment requirements (16-byte alignment for float4)
-  - When to use: memory-bound kernels where N is divisible by vector width
-  - Tail handling when N is not a multiple of 4
-  - Interaction with coalescing (vectorized loads are inherently coalesced)
-  - LDG.128 / STG.128 PTX instructions generated by the compiler
-
-  apis.md is REQUIRED: float4, float2, int4, uint4, make_float4, etc.
-
-  Microbench probe: compare scalar float load bandwidth vs float4 vectorized
-  load bandwidth on HBM. Simple kernel that reads N floats from global memory:
-    (a) scalar: thread i reads array[i] one float at a time
-    (b) float4: thread i reads ((float4*)array)[i], processes 4 floats at once
-  Measure effective bandwidth (GB/s) for both. Expect ~2-4x improvement.

package/resources/kop-2026-10-23-0.1.17-beta.1/product_capabilities/kernel_optflow/skills/KernelWiki-private/tasks/build-warp-primitives.yaml

@@ -1,44 +0,0 @@
-task_id: 2026-04-16-warp-primitives
-task_type: build-skill
-target_path: wiki/nvidia/foundations/compute/warp-primitives/
-
-hardware:
-  device: H200
-  sm: "9.0a"
-  cuda_runtime: "12.9"
-
-upstream_scope:
-  - cuda-official/toolkit-docs-13.2
-  - source-code/cuda-samples
-
-references:
-  - reasoning/api-probing.md
-  - reasoning/hardware-microbench.md
-  - reasoning/benchmark-protocol.md
-  - reasoning/bottleneck-triage.md
-
-success_criteria:
-  - file_exists: skill.md
-  - file_exists: pitfalls.md
-  - file_exists: apis.md
-  - lint_clean: true
-  - at_least_one_hw_probe_linked: true
-
-related_skills: [coalescing, vectorized-access, bank-conflict]
-
-notes: |
-  Cover the warp shuffle family (shfl.sync.idx / up / down / bfly) and
-  the warp vote family (ballot / all / any). Include one classical
-  warp-reduce example: block-level sum of 1024 fp32 elements using
-  butterfly reduction (__shfl_xor_sync).
-
-  Skip match.sync and reduce.sync for now — they are deferred post-MVP.
-
-  The agent must write its own compute-latency microbench probe for
-  shfl.sync.bfly following hardware-microbench.md protocol. The canonical
-  baseline (4096 trials, 128-unroll dependent chain) measured 23.83 cycles
-  on H200 — the skill-specific probe should exercise the exact warp-reduce
-  pattern and may show a different number.
-
-  apis.md is REQUIRED for this skill because it touches __shfl_sync,
-  __ballot_sync, and related intrinsics.

package/resources/kop-2026-10-23-0.1.17-beta.1/product_capabilities/kernel_optflow/skills/KernelWiki-private/tasks/extract-cutlass-persistent-kernel.yaml

@@ -1,51 +0,0 @@
-task_id: 2026-04-28-cutlass-persistent-kernel
-task_type: build-classical-algo   # see reasoning/task-packet.md
-target_path: wiki/nvidia/techniques/persistent-kernel/
-also_writes:
-  - wiki/nvidia/code-walkthroughs/cutlass-cute/persistent-kernel/
-  - sources/experience/api-probes/gemm/
-
-hardware:
-  device: H200
-  sm: "9.0a"
-  cuda_runtime: "12.9"
-
-upstream_scope:
-  - <local-path>
-  - <local-path>
-upstream_repo_pin: "cutlass@f74fea9c"
-
-references:
-  - templates/frontmatter/classical-algo.yaml
-  - reasoning/benchmark-protocol.md
-  - docs/15-代码仓库抽取知识流程_update_en.md  # AC-6 (persistent kernel)
-
-source_anchors:
-  - include/cutlass/gemm/kernel/sm90_gemm_tma_warpspecialized_pingpong.hpp
-  - include/cutlass/gemm/kernel/tile_scheduler.hpp
-  - corpus/nvidia/blogs/colfax/cutlass-tutorial-persistent-kernels-and-stream-k
-
-success_criteria:
-  - file_exists: wiki/nvidia/techniques/persistent-kernels.md
-  - file_exists: wiki/nvidia/techniques/persistent-kernel/pitfalls.md
-  - file_exists: wiki/nvidia/code-walkthroughs/cutlass-cute/persistent-kernel/build.sh
-  - file_exists: wiki/nvidia/code-walkthroughs/cutlass-cute/persistent-kernel/persistent_pingpong_ab.cu
-  - dir_nonempty: sources/experience/api-probes/gemm/   # date-stamped *-persistent-ablation.md at execution time
-  - lint_clean: true
-  - skill_status: verified
-  - skill_evidence_level: measured
-  - ablation_ab_data_present: true
-
-depends_on:
-  - build-cutlass-gemm-aligned.yaml      # AC-4 baseline
-  - extract-cutlass-warp-specialize.yaml # AC-6 prereq
-
-blockers_for_round_1:
-  - reason: "H200 A/B; depends on aligned harness + warp-specialize skeleton."
-
-notes: |
-  AC-6 partial — persistent-kernel / pingpong half. Distill the tile-scheduler
-  loop driving a fixed grid that consumes work tiles until exhaustion.
-  A/B: toggle persistent vs one-grid-launch-per-tile; measure SM idle time
-  and total kernel time. Stream-K is a stretch goal; do NOT block this task
-  on it.

package/resources/kop-2026-10-23-0.1.17-beta.1/product_capabilities/kernel_optflow/skills/KernelWiki-private/tasks/extract-cutlass-warp-specialize.yaml

@@ -1,51 +0,0 @@
-task_id: 2026-04-28-cutlass-warp-specialize
-task_type: build-classical-algo   # see reasoning/task-packet.md
-target_path: wiki/nvidia/techniques/warp-specialization/
-also_writes:
-  - wiki/nvidia/code-walkthroughs/cutlass-cute/warp-specialization/
-  - sources/experience/api-probes/gemm/
-
-hardware:
-  device: H200
-  sm: "9.0a"
-  cuda_runtime: "12.9"
-
-upstream_scope:
-  - <local-path>
-  - <local-path>
-upstream_repo_pin: "cutlass@f74fea9c"
-
-references:
-  - templates/frontmatter/classical-algo.yaml
-  - reasoning/benchmark-protocol.md
-  - docs/15-代码仓库抽取知识流程_update_en.md  # AC-6 (warp-specialization)
-
-source_anchors:
-  - examples/48_hopper_warp_specialized_gemm/
-  - include/cutlass/gemm/kernel/sm90_gemm_tma_warpspecialized.hpp
-
-success_criteria:
-  - file_exists: wiki/nvidia/techniques/warp-specialization.md
-  - file_exists: wiki/nvidia/techniques/warp-specialization/pitfalls.md
-  - file_exists: wiki/nvidia/code-walkthroughs/cutlass-cute/warp-specialization/build.sh
-  - file_exists: wiki/nvidia/code-walkthroughs/cutlass-cute/warp-specialization/warp_specialize_ab.cu
-  - dir_nonempty: sources/experience/api-probes/gemm/   # date-stamped *-warp-specialize-ablation.md at execution time
-  - lint_clean: true
-  - skill_status: verified
-  - skill_evidence_level: measured
-  - applies_to_pattern_class_contains: tensor-core
-  - applies_to_pattern_class_or_ops_match: gemm
-  - ablation_ab_data_present: true
-
-depends_on:
-  - build-cutlass-gemm-aligned.yaml  # AC-4 baseline
-
-blockers_for_round_1:
-  - reason: "H200 A/B run; depends on aligned harness."
-
-notes: |
-  AC-6 partial — warp-specialize half. Distill the dual-warpgroup producer/consumer
-  pattern (load warpgroup driving TMA + mbarrier; compute warpgroup driving wgmma)
-  into a readable skeleton, NOT a verbatim copy of cutlass templates.
-  A/B: toggle warp-specialization vs a single-warpgroup baseline at the same
-  tile/stage configuration; report end-to-end gain.

package/resources/kop-2026-10-23-0.1.17-beta.1/product_capabilities/kernel_optflow/skills/KernelWiki-private/tasks/probe-batch-1.yaml

@@ -1,27 +0,0 @@
-# Probe batch 1 — curated high-value APIs (X3 pitfall criterion)
-# 13 APIs across 4 thematic sub-batches, parallelizable.
-
-# Sub-batch A: Warp shuffle family — completes the `__shfl_*_sync` set
-# (sibling of already-probed __shfl_xor_sync).
-- __shfl_sync
-- __shfl_up_sync
-- __shfl_down_sync
-
-# Sub-batch B: Cache-hint loads — 5 variants of __ldg with different L1/L2
-# caching policies. Non-obvious perf implications on H200.
-- __ldcg
-- __ldca
-- __ldcs
-- __ldcv
-- __ldlu
-
-# Sub-batch C: Async-copy primitives — core APIs for LDGSTS prefetching.
-# Already have end-to-end async-copy probe; these are the building blocks.
-- __pipeline_memcpy_async
-- __pipeline_commit
-- __pipeline_wait_prior
-
-# Sub-batch D: Occupancy + function attribute APIs — most-referenced across
-# skills (compiler-hints, register-pressure, occupancy-tuning all touch).
-- cudaOccupancyMaxActiveBlocksPerMultiprocessor
-- cudaFuncSetAttribute

package/resources/kop-2026-10-23-0.1.17-beta.1/product_capabilities/kernel_optflow/skills/KernelWiki-private/tasks/probe-runtime-all-sync.yaml

@@ -1,68 +0,0 @@
-task_id: 2026-04-17-probe-runtime-all-sync
-task_type: probe-api
-target_path: wiki/nvidia/api-definitions/runtime/__all_sync.md
-hardware:
-  device: H200
-  sm: 9.0a
-  cuda_runtime: '12.9'
-upstream_scope:
-- corpus/nvidia/
-references:
-- reasoning/api-probing.md
-- reasoning/benchmark-protocol.md
-success_criteria:
-- file_exists: sources/experience/api-probes/2026-04-17-runtime-all-sync.md
-- file_updated: wiki/nvidia/api-definitions/runtime/__all_sync.md
-- has_end_to_end_example_flipped: true
-- 'lint_clean: true'
-api:
-  name: __all_sync
-  namespace: runtime
-referenced_in_corpus:
-- path: corpus/nvidia/cuda-official/cuda-toolkit-documentation-13.2/CUDA API References/cccl/cccl.md
-  anchor: L22834-L22834
-  title: Removed functions and classes
-  scope: cuda-official
-- path: corpus/nvidia/cuda-official/cuda-toolkit-documentation-13.2/CUDA Programming
-    Guides/cuda-programming-guide/cuda_cuda-programming-guide_index.html.md
-  anchor: L23872-L23872
-  title: 5.4.6.2. Warp Vote Functions
-  scope: cuda-official
-- path: corpus/nvidia/cuda-official/cuda-toolkit-documentation-13.2/CUDA Programming
-    Guides/cuda-programming-guide/cuda_cuda-programming-guide_index.html.md
-  anchor: L23880-L23880
-  title: 5.4.6.2. Warp Vote Functions
-  scope: cuda-official
-- path: corpus/nvidia/cuda-official/cuda-toolkit-documentation-13.2/CUDA Programming
-    Guides/cuda-programming-guide/cuda_cuda-programming-guide_index.html.md
-  anchor: L24232-L24232
-  title: 5.4.6.6. Warp __sync Intrinsic Constraints
-  scope: cuda-official
-- path: corpus/nvidia/cuda-official/cuda-toolkit-documentation-13.2/CUDA Programming
-    Guides/cuda-programming-guide/cuda_cuda-programming-guide_index.html.md
-  anchor: L24238-L24238
-  title: 5.4.6.6. Warp __sync Intrinsic Constraints
-  scope: cuda-official
-- path: corpus/nvidia/source-code/cuda-samples/Samples/0_Introduction/README.md
-  anchor: L116-L116
-  title: '[simpleVoteIntrinsics](./simpleVoteIntrinsics)'
-  scope: source-code
-- path: corpus/nvidia/source-code/cuda-samples/Samples/0_Introduction/simpleVoteIntrinsics/README.md
-  anchor: L5-L5
-  title: Description
-  scope: source-code
-- path: corpus/nvidia/source-code/cuda-samples/Samples/0_Introduction/simpleVoteIntrinsics/simpleVote_kernel.cuh
-  anchor: L58-L58
-  title: define SIMPLEVOTE_KERNEL_CU
-  scope: source-code
-- path: corpus/nvidia/source-code/cuda-samples/Samples/0_Introduction/simpleVoteIntrinsics/simpleVote_kernel.cuh
-  anchor: L76-L76
-  title: define SIMPLEVOTE_KERNEL_CU
-  scope: source-code
-notes: "F10 probe for __all_sync. Follow api-probing.md 6-step protocol.\nPrefetched\
-  \ corpus hits are provided in referenced_in_corpus — prefer reading those via source_read\
-  \ before writing new probes.\nDeliverables:\n  1. sources/experience/api-probes/2026-04-17-runtime-all-sync.md\
-  \ (probe record with experience.yaml frontmatter)\n  2. artifacts/experience/api-probes/__all_sync_probe.cu\
-  \ (minimal runnable kernel, compiled on H200 via SSH)\n  3. Update wiki/nvidia/api-definitions/runtime/__all_sync.md:\
-  \ flip has_end_to_end_example → true, set probed_by to the probe record path, append\
-  \ ## End-to-End Example section\nRun scripts/validate.py before declaring done."

package/resources/kop-2026-10-23-0.1.17-beta.1/product_capabilities/kernel_optflow/skills/KernelWiki-private/tasks/probe-runtime-any-sync.yaml

@@ -1,67 +0,0 @@
-task_id: 2026-04-17-probe-runtime-any-sync
-task_type: probe-api
-target_path: wiki/nvidia/api-definitions/runtime/__any_sync.md
-hardware:
-  device: H200
-  sm: 9.0a
-  cuda_runtime: '12.9'
-upstream_scope:
-- corpus/nvidia/
-references:
-- reasoning/api-probing.md
-- reasoning/benchmark-protocol.md
-success_criteria:
-- file_exists: sources/experience/api-probes/2026-04-17-runtime-any-sync.md
-- file_updated: wiki/nvidia/api-definitions/runtime/__any_sync.md
-- has_end_to_end_example_flipped: true
-- 'lint_clean: true'
-api:
-  name: __any_sync
-  namespace: runtime
-referenced_in_corpus:
-- path: corpus/nvidia/cuda-official/cuda-toolkit-documentation-13.2/CUDA API References/cccl/cccl.md
-  anchor: L22835-L22835
-  title: Removed functions and classes
-  scope: cuda-official
-- path: corpus/nvidia/cuda-official/cuda-toolkit-documentation-13.2/CUDA Programming
-    Guides/cuda-programming-guide/cuda_cuda-programming-guide_index.html.md
-  anchor: L23864-L23864
-  title: 5.4.6.1. Warp Active Mask
-  scope: cuda-official
-- path: corpus/nvidia/cuda-official/cuda-toolkit-documentation-13.2/CUDA Programming
-    Guides/cuda-programming-guide/cuda_cuda-programming-guide_index.html.md
-  anchor: L23873-L23873
-  title: 5.4.6.2. Warp Vote Functions
-  scope: cuda-official
-- path: corpus/nvidia/cuda-official/cuda-toolkit-documentation-13.2/CUDA Programming
-    Guides/cuda-programming-guide/cuda_cuda-programming-guide_index.html.md
-  anchor: L23882-L23882
-  title: 5.4.6.2. Warp Vote Functions
-  scope: cuda-official
-- path: corpus/nvidia/source-code/cuda-samples/Samples/0_Introduction/README.md
-  anchor: L116-L116
-  title: '[simpleVoteIntrinsics](./simpleVoteIntrinsics)'
-  scope: source-code
-- path: corpus/nvidia/source-code/cuda-samples/Samples/0_Introduction/simpleVoteIntrinsics/README.md
-  anchor: L5-L5
-  title: Description
-  scope: source-code
-- path: corpus/nvidia/source-code/cuda-samples/Samples/0_Introduction/simpleVoteIntrinsics/simpleVote_kernel.cuh
-  anchor: L46-L46
-  title: define SIMPLEVOTE_KERNEL_CU
-  scope: source-code
-- path: corpus/nvidia/source-code/cuda-samples/Samples/0_Introduction/simpleVoteIntrinsics/simpleVote_kernel.cuh
-  anchor: L70-L70
-  title: define SIMPLEVOTE_KERNEL_CU
-  scope: source-code
-- path: corpus/nvidia/source-code/cutlass/examples/77_blackwell_fmha/collective/sm100_fmha_fwd_mainloop_tma_warpspecialized.hpp
-  anchor: L1019-L1019
-  title: endif
-  scope: source-code
-notes: "F10 probe for __any_sync. Follow api-probing.md 6-step protocol.\nPrefetched\
-  \ corpus hits are provided in referenced_in_corpus — prefer reading those via source_read\
-  \ before writing new probes.\nDeliverables:\n  1. sources/experience/api-probes/2026-04-17-runtime-any-sync.md\
-  \ (probe record with experience.yaml frontmatter)\n  2. artifacts/experience/api-probes/__any_sync_probe.cu\
-  \ (minimal runnable kernel, compiled on H200 via SSH)\n  3. Update wiki/nvidia/api-definitions/runtime/__any_sync.md:\
-  \ flip has_end_to_end_example → true, set probed_by to the probe record path, append\
-  \ ## End-to-End Example section\nRun scripts/validate.py before declaring done."

package/resources/kop-2026-10-23-0.1.17-beta.1/product_capabilities/kernel_optflow/skills/KernelWiki-private/tasks/probe-skip.yaml

@@ -1,33 +0,0 @@
-# Persistent skip list for probe_gap_scan / probe_loop.
-#
-# APIs listed here are filtered from the `dangling-body-ref` scan output on
-# top of the hardcoded _SKILL_BODY_SKIP set in tools/probe_gap_scan.py.
-#
-# Use this file for project-specific skips that don't belong in the
-# hardcoded core set. Edit freely; format is either a top-level list or
-# {skip_apis: [...]} mapping.
-#
-# Guidelines:
-#   - Skip things that are NOT function-like: enum values, type aliases,
-#     storage qualifiers, preprocessor macros.
-#   - Skip things that are function-like but TRIVIAL (cudaMalloc, atomicAdd,
-#     cudaMemcpy) — these are better served by direct source_corpus lookup
-#     than a probe record.
-#   - DO NOT skip APIs with non-obvious semantics (mask rules, alignment
-#     constraints, H200-specific behavior). Those warrant a probe record.
-#
-# To un-skip something, delete its line here. Hardcoded skips can only be
-# changed in probe_gap_scan.py.
-
-- cudaMalloc                  # trivial; every CUDA dev already knows
-- cudaMallocManaged           # trivial; use source_corpus.cli for signature
-- cudaFree                    # trivial
-- cudaMemcpy                  # trivial
-- cudaMemcpyAsync             # trivial
-- cudaMemcpyToSymbol          # trivial
-- cudaHostAlloc               # trivial
-- cudaDeviceSynchronize       # trivial
-- cudaGetLastError            # trivial
-- atomicAdd                   # basic atomic, covered by C++ atomic semantics
-- atomicInc                   # basic atomic
-- cudaGetDeviceProperties     # covered by kp_introspect device-static

package/resources/kop-2026-10-23-0.1.17-beta.1/product_capabilities/kernel_optflow/skills/KernelWiki-private/tasks/validate-probe-all-sync.yaml

@@ -1,34 +0,0 @@
-task_id: 2026-04-17-validate-probe-all-sync
-task_type: probe-api
-target_path: wiki/nvidia/api-definitions/runtime/
-
-hardware:
-  device: H200
-  sm: "9.0a"
-  cuda_runtime: "12.9"
-
-upstream_scope:
-  - <local-path>
-  - <local-path>
-
-references:
-  - reasoning/api-probing.md
-  - reasoning/benchmark-protocol.md
-
-success_criteria:
-  - file_exists: wiki/nvidia/api-definitions/runtime/__all_sync.md
-  - source_corpus_used: true
-  - final_report_mentions_probe: true
-
-notes: |
-  Validate the new source corpus integration by probing the runtime intrinsic
-  `__all_sync`.
-
-  Requirements:
-  1. Use `source_search` and `source_read` to find the original semantics first.
-  2. If the API is sufficiently documented, create or update
-     `wiki/nvidia/api-definitions/runtime/__all_sync.md`.
-  3. If you can complete the minimal runnable probe, also write the probe record
-     under `sources/experience/api-probes/`.
-  4. If the task cannot finish within the current turn budget, still report the
-     retrieval hits you used and what remains blocked.

package/resources/kop-2026-10-23-0.1.17-beta.1/product_capabilities/kernel_optflow/skills/KernelWiki-private/templates/frontmatter/api-raw.yaml

@@ -1,31 +0,0 @@
-# api-raw entry template
-#
-# Used by 10-api-raw/<namespace>/<func>.md
-# One file per API function.
-# Enforced by tools/lint_knowledge.py --layer api-raw
-
-# ==== Identity ====
-func_name: ""                          # REQUIRED. e.g. "__shfl_sync"
-namespace: ""                          # REQUIRED. runtime | driver | cublas | cub | thrust | ptx | math | libcupp
-header: ""                             # REQUIRED. e.g. "cuda_runtime.h" or "<cub/cub.cuh>" or "PTX"
-signature: ""                          # REQUIRED. Full C/C++ signature on one line.
-since_cuda: ""                         # Optional. Minimum CUDA toolkit version, e.g. "11.0"
-
-# ==== Status ====
-status: draft                          # REQUIRED. draft | verified
-has_end_to_end_example: false          # REQUIRED. Flipped true once a probe back-fills the page.
-
-# ==== Semantics hints ====
-parameters: []                         # Optional. List of {name, type, description}
-return: ""                             # Optional. Return type + meaning.
-preconditions: []                      # Optional. Caller must ensure ...
-error_modes: []                        # Optional. What can go wrong at runtime.
-
-# ==== Provenance (REQUIRED: at least one entry) ====
-source:
-  - path: ""                           # REQUIRED. absolute upstream path
-    anchor: ""                         # REQUIRED. heading or line range
-    excerpt: ""                        # Optional. short quote.
-
-# ==== Probe linkage ====
-probed_by: ""                          # Optional. relative path to 80-experience/api-probes/<date>-<ns>-<func>.md

package/resources/kop-2026-10-23-0.1.17-beta.1/product_capabilities/kernel_optflow/skills/KernelWiki-private/templates/frontmatter/classical-algo.yaml

@@ -1,55 +0,0 @@
-# 50-classical-algo/<algo>/skill.md frontmatter template
-#
-# Used by 50-classical-algo/<algo>/skill.md
-# Enforced by tools/lint_knowledge.py --layer classical-algo
-#
-# Classical-algo skills are skill.md-shaped: they document an end-to-end
-# optimization algorithm (warp-specialization, persistent-kernel, online-softmax,
-# split-k, decoupled-lookback-scan, …) extracted from upstream code. The
-# algorithm-skeleton code itself lives under 60-code/<source-repo>/<algo>/;
-# this layer is the prose + measured A/B story.
-#
-# Schema mirrors skill.yaml; the only deltas are:
-#   * applies_to_pattern_class is open (cuda-core / tensor-core / fused-…)
-#     because classical algorithms span all three.
-#   * status: verified MUST come with evidence_level: measured plus an A/B
-#     experiment record under 80-experience/api-probes/<op>/ that toggles
-#     the algorithm on/off.
-
-# ==== Core identity ====
-title: ""                              # REQUIRED. e.g. "Warp-specialized GEMM mainloop"
-status: draft                          # REQUIRED. draft | verified | deprecated
-evidence_level: spec                   # REQUIRED.
-
-# ==== Applicability ====
-applies_to_pattern_class:              # REQUIRED. At least one entry.
-  - tensor-core
-applies_to_ops: []                     # REQUIRED.
-requires_sm: ">=9.0"                   # REQUIRED.
-requires_features: []                  # Optional.
-single_kernel_useful: true             # REQUIRED.
-
-# ==== Environment (REQUIRED-IF evidence_level == measured) ====
-cuda_version_tested: ""
-driver_version_tested: ""
-toolchain: ""
-measured_on: ""
-
-# ==== Provenance ====
-source:
-  - path: ""
-    anchor: ""
-    excerpt: ""
-
-# ==== Artifacts (REQUIRED-IF status == verified AND evidence_level == measured) ====
-artifacts:
-  code: ""                             # relative path to algorithm skeleton (often under 60-code/<source-repo>/<algo>/)
-  build: ""
-  introspection: ""
-  profile: ""                          # Optional.
-  ablation: ""                         # Optional. Path to A/B experiment record under 80-experience/.
-
-# ==== Cross-references ====
-related_apis: []
-related_skills: []
-upstream_repo: ""                      # Optional. e.g. "cutlass@f74fea9c"

package/resources/kop-2026-10-23-0.1.17-beta.1/product_capabilities/kernel_optflow/skills/KernelWiki-private/templates/frontmatter/experience.yaml

@@ -1,52 +0,0 @@
-# Experience / Probe record template
-#
-# Used by 80-experience/api-probes/<YYYY-MM-DD>-<ns>-<func-slug>.md
-# Enforced by tools/lint_knowledge.py --layer experience
-
-# ==== Identity ====
-api: ""                                # REQUIRED. e.g. "__shfl_xor_sync"
-namespace: ""                          # REQUIRED. runtime | driver | cublas | cub | thrust | ptx | math | libcupp
-probe_slug: ""                         # REQUIRED. kebab-case slug used in the filename
-
-# ==== Status ====
-status: draft                          # REQUIRED. draft | verified | failed | blocked
-kind: documented                       # REQUIRED. documented | undocumented
-trigger: init-sweep                    # REQUIRED. init-sweep | runtime-demand | skill-build
-evidence_level: inferred               # REQUIRED. inferred | measured
-clock_policy: unknown                  # REQUIRED. unknown | locked | auto | stable-auto
-
-# ==== Hardware context (REQUIRED; filled by kp_introspect) ====
-measured_on:
-  device: ""                           # e.g. "NVIDIA H200"
-  sm: ""                               # e.g. "9.0a"
-  gpu_uuid: ""                         # Optional. from kp_introspect device-static.
-  cuda_runtime: ""                     # e.g. "12.4"
-  driver: ""                           # e.g. "550.54.15"
-
-# ==== Artifacts (REQUIRED when evidence_level == measured) ====
-artifacts:
-  code: ""                             # relative path to probe .cu
-  build: ""                            # relative path to build.sh OR literal nvcc command
-  introspection: ""                    # kp_introspect bundle json path
-  profile: ""                          # Optional. ncu-rep path.
-
-# ==== Upstream grounding (REQUIRED) ====
-source_refs: []                        # List of {source_id, path, anchor} MANIFEST-backed refs
-source:
-  - path: ""                           # REQUIRED (unless kind == undocumented)
-    anchor: ""
-    excerpt: ""
-
-# ==== Measurement summary (REQUIRED when status == verified) ====
-conclusions:
-  max_abs_err: null                    # numeric; pass/fail vs. tolerance
-  latency_ms_median: null
-  latency_ms_p10: null
-  latency_ms_p90: null
-  baseline_name: ""                    # e.g. "torch.sum" or "cub::BlockReduce"
-  baseline_ms: null
-  ratio: null                          # baseline_ms / latency_ms_median
-
-# ==== Follow-ups ====
-back_filled_into: []                   # relative paths to 10-api-raw entries that gained ## End-to-End Example
-open_questions: []                     # free-form strings for human review

package/resources/kop-2026-10-23-0.1.17-beta.1/product_capabilities/kernel_optflow/skills/KernelWiki-private/templates/frontmatter/hardware-feature.yaml

@@ -1,52 +0,0 @@
-# 40-hardware-feature/<feature>/skill.md frontmatter template
-#
-# Used by 40-hardware-feature/<feature>/skill.md
-# Enforced by tools/lint_knowledge.py --layer hardware-feature
-#
-# Hardware-feature skills are skill.md-shaped: they document a single
-# hardware-new-feature (TMA, wgmma, DSMEM, async-pipeline, setmaxnreg, …)
-# with evidence drawn from on-target measurement on H200 (sm_90a) or newer.
-# The schema mirrors skill.yaml; the only deltas are:
-#   * applies_to_pattern_class defaults to ["tensor-core"] but may include
-#     ["cuda-core", "tensor-core", "fused-cuda-tensor-core"] depending on
-#     where the feature plugs in.
-#   * requires_sm must be a Hopper-or-newer SM gate (>=9.0 / >=9.0a / >=10.0).
-#   * status: verified MUST come with evidence_level: measured AND a populated
-#     artifacts: {code, build, introspection} block plus measured_on.
-
-# ==== Core identity ====
-title: ""                              # REQUIRED. e.g. "TMA: cp.async.bulk.tensor end-to-end"
-status: draft                          # REQUIRED. draft | verified | deprecated
-evidence_level: spec                   # REQUIRED. spec | measured | inferred | anecdotal
-
-# ==== Applicability ====
-applies_to_pattern_class:              # REQUIRED. At least one entry.
-  - tensor-core                        #   cuda-core | tensor-core | fused-cuda-tensor-core
-applies_to_ops: []                     # REQUIRED. e.g. [gemm, attention]
-requires_sm: ">=9.0"                   # REQUIRED. Hopper or newer for this layer.
-requires_features: []                  # Optional. e.g. [tma, wgmma, mbarrier]
-single_kernel_useful: true             # REQUIRED.
-
-# ==== Environment (REQUIRED-IF evidence_level == measured) ====
-cuda_version_tested: ""
-driver_version_tested: ""
-toolchain: ""
-measured_on: ""                        # e.g. "H200-SXM | sm_90a | cuda 12.9 | driver 555.42"
-
-# ==== Provenance (REQUIRED: at least one entry) ====
-source:
-  - path: ""                           # absolute upstream path or knowledge/-relative
-    anchor: ""                         # heading slug or "Lxxx-Lyyy"
-    excerpt: ""                        # Optional, ≤200 chars
-
-# ==== Artifacts (REQUIRED-IF status == verified AND evidence_level == measured) ====
-artifacts:
-  code: ""                             # relative path to probe/microbench .cu (often under 60-code/<source-repo>/<topic>/)
-  build: ""                            # build.sh path or literal nvcc command
-  introspection: ""                    # kp_introspect bundle json path
-  profile: ""                          # Optional. ncu-rep / nsys-rep path
-
-# ==== Cross-references ====
-related_apis: []                       # Optional. slugs into 10-api-raw/
-related_skills: []                     # Optional. slugs into 30-skill/, 50-classical-algo/
-upstream_repo: ""                      # Optional. e.g. "cutlass@f74fea9c" — pin the commit when extracting.