feat(kernel_workflow): warm-start from a local kb_artifacts

.github/workflows/ci-l0-checks.yml

@@ -55,10 +55,13 @@ jobs:
       - uses: actions/setup-python@v5
         with:
           python-version: "3.11"
-      # pyyaml is required, not optional: run_e2e.py parses the recipe envs: block
-      # with it, and the recipe-env tests assert the string-preserving BaseLoader
-      # path (typed scalars, folded multi-line EXTRA_<BE>_ARGS). Without it those
-      # tests would silently exercise the degraded line-scanner fallback instead.
+      # pyyaml is required, not optional, and two suites need it. run_e2e.py parses
+      # the recipe envs: block with it, and the recipe-env tests assert the
+      # string-preserving BaseLoader path (typed scalars, folded multi-line
+      # EXTRA_<BE>_ARGS); without it those tests would silently exercise the degraded
+      # line-scanner fallback instead. The experience-store tests read/write the
+      # on-disk meta.yaml format (that module degrades to JSON without pyyaml, but
+      # then the test would be asserting a format the real store never uses).
       - run: pip install pytest pytest-cov coverage pyyaml
       # Stdlib-only tests (import run_e2e via importlib; no claude_agent_sdk / GPU).
       # Files are listed explicitly rather than discovered, so that adding a test
@@ -95,6 +98,12 @@ jobs:
             e2e_workflow/scripts/tests/test_leg_runner.py \
             e2e_workflow/scripts/tests/test_server_teardown.py \
             e2e_workflow/scripts/tests/test_bench_e2e_teardown_lookup.py \
+            e2e_workflow/scripts/tests/test_e2e_store.py \
+            e2e_workflow/scripts/tests/test_kb_retract.py \
+            kernel_workflow/scripts/tests/test_experience_store.py \
+            kb/tests/test_store_local.py \
+            kb/tests/test_attest.py \
+            kernel_workflow/scripts/tests/test_kb_loop_offline.py \
             ci/node/test_run_model_material_audit.py \
             geak/test_bootstrap.py
 

.gitignore

@@ -224,6 +224,11 @@ exp/
 .torch_ext/
 # bench_e2e.sh default output dir (OUT_DIR=$(pwd)/e2e_bench_out) — runtime measurements
 e2e_bench_out/
+# Machine-produced experience KB (warm-start store, kernel_workflow Part 4/1.7): runtime-accumulated
+# best patches + meta; volume is unbounded, so default-ignore and commit selectively via a whitelist.
+kb_artifacts/
+# Default root of the on-disk KB Store plane (kb/store_local.py): runtime records + artifacts.
+kb_store_local/
 
 # rocprofv3 scratch: written next to the workflow scripts when a profiling run is
 # launched from this directory. Build output, never source.

README.md

@@ -183,6 +183,7 @@ GEAK/
 │   ├── roles/  knowledge/  scripts/   # gpu_lock.sh, profile_kernel.sh
 │   └── README.md
 ├── perf_knowledge/      # AMD operator × backend SOTA knowledge base (REFERENCE ONLY)
+├── kb_artifacts/        # machine-produced best patches (code-carrying, gitignored, warm-start source)
 ├── examples/            # Example kernel tasks, benchmark comparisons, real e2e runs
 └── exp/                 # Experiment outputs (timestamped per run)
 ```

docs/reference/api-reference.md

@@ -101,6 +101,8 @@ Related timing/tuning args: `fast_head_deadline_ms`, `fast_head_workflow_ms`, `d
 | `ab_finish_retries` | `3` | A/B leg completion retries. |
 | `use_expert_skills` | `false` | Consult `perf_knowledge/expert_skills` (advisory priors). OFF = byte-identical. |
 | `perf_knowledge_dir` | sibling `perf_knowledge/` | Authoring knowledge base. |
+| `kb_artifacts_dir` | sibling `kb_artifacts/` | Warm-start patch store, forwarded to every recursive kernel lane so one run shares one store. |
+| `warm_start`, `warm_start_match`, `warm_start_min_speedup`, `kb_mode`, `kb_store_dir`, `kb_framework_version` | see the kernel-layer table | Forwarded verbatim to the kernel lanes. Corrective re-authors are forced to `reference` (they must keep the isolated win). |
 | `time_budget_s`, `initial_extra_server_args`, `initial_extra_env`, `tracelens`, `agent_timeout_ms` | — | Forwarded from the external orchestrator. |
 
 ### Example
@@ -142,6 +144,13 @@ budget-controlled, each patch independently verified.
 | `target_language` | `triton` | Author-mode language: `triton` \| `flydsl` \| `hip` \| `ck`. |
 | `op_spec` | `{}` | Op specification (author mode). |
 | `perf_knowledge_dir` | sibling `perf_knowledge/` | Knowledge base. |
+| `warm_start` | `on` | Local experience reuse from `kb_artifacts/`. `on` = read the top same-arch patches (one per optimization `direction`) through the verify gate; `reference` = prose only, no patch apply; `return_after_read` = apply+validate a candidate then return; `off` = cold start (byte-equivalent to pre-warm-start). |
+| `kb_artifacts_dir` | sibling `kb_artifacts/` | Lossless best-patch sink (machine-produced, code-carrying; gitignored, runtime-accumulated). |
+| `warm_start_match` | `fuzzy` | How a kernel name reaches a stored page: `exact` (canonicalized name only) \| `normalized` \| `fuzzy` (closest containing page; refuses when two pages fit equally well). Names are canonicalized first, so `fused_moe_kernel_task`, `triton_fused_moe_kernel.py` and a full task path all reach the same page. |
+| `warm_start_min_speedup` | `1.05` | Floor on a stored entry's recorded speedup before it is worth a verify slot. |
+| `kb_mode` | `local` | Which plane warm start reads and writes. `local` = the curated `kb_artifacts/` tree, addressed by slug. `store` = a KB Store on disk in the shape the KernelForge service uses, addressed by canonical id (`kernel:geak:<name>:rocm:<ver>:<triton\|hip\|ck>:mi355x`); the write records BOTH planes, so they cannot drift. Phases, schemas and the verify gate are identical either way. |
+| `kb_store_dir` | sibling `kb_store_local/` | Root of the `store`-mode plane. Ignored when `kb_mode=local`. |
+| `kb_framework_version` | measured stack | ROCm `<major>.<minor>` for the key's `framework_version` segment. Set it on a box with no `/opt/rocm`, where the measured stack is empty and every record would otherwise file under `unspecified` — splitting one kernel's history across two keys. Affects the key only, never the recorded stack. |
 | `update_experience` | `on` | Curate ONE distilled card into `kernel_workflow/knowledge/learned/` after a measured win (every lane run; once centrally per bake-off). `off` skips it. The sink is always this workflow's own `learned/` — an e2e-opened lane never writes into `e2e_workflow/knowledge/learned/`. Cards carry a skill-style discovery header (`name`/`description`/`keywords`/`kernels`/`platforms`/`kernel_class`/`regime`) and `learned/INDEX.md` is regenerated from it by `kernel_workflow/scripts/kb.py ... index` (`--check` = fail-if-stale). Numbers on a card are relative only (ratios, % of achievable peak) — never wall-clock. |
 | `workload_spec_path` | — | Workload-alignment spec; makes the primary metric the time-weighted ratio-of-sums. |
 | `agent_timeout_ms` | `3600000` | Per-agent timeout (1h). |