diff --git a/docs/knowledge-base.md b/docs/knowledge-base.md
index ec46594e6..88d1fa8ee 100644
--- a/docs/knowledge-base.md
+++ b/docs/knowledge-base.md
@@ -14,7 +14,7 @@ separate work, not an implemented capability of this knowledge base.
 ```text
 src/inferrsi/data/skill_library/
   README.md                         Library navigation
-  skill-index.json                   Stable IDs → categorized paths/backend (60 skills)
+  skill-index.json                   Stable IDs → categorized paths/backend (61 skills)
   hardware/{ascend,cuda,host}/<id>/   Hardware/platform-specific complete skills
   frameworks/<framework>/<id>/      Framework-owned complete skills
   repositories/<repo>/              Engine entry, sub-skills and focused complete skills
@@ -171,18 +171,18 @@ Complete procedures, sidecars, references and cases are packaged, with 13 Ascend
 repositories, methods or workflows. The 21 engine and research-router skills
 use the same ownership layout under `repositories/<engine>/` and
 `workflows/inference-engine-research/`, including their sub-skills and references.
-The library contains 81 `SKILL.md` files, all in English.
+The library contains 85 `SKILL.md` files, all in English.
 The collection overview, import notes and source manifest belong to the research
 workflow's references; Strata design-space analysis belongs to Strata. The library
 root contains only navigation and the registry.
 
 The [generated catalog](../src/inferrsi/data/repo_packs/inferrsi/knowledge/README.md)
-projects the 60 registered skills into the contribution catalog. Its records
+projects the 61 registered skills into the contribution catalog. Its records
 use `inferrsi.<backend>.skills.<name_with_underscores>` IDs, with `ascend` and
 `cuda` for scopes that depend on those backends and `common` for the remaining skills.
 Each `skill_documents.original` points to the full packaged procedure. The
 `skills` CLI retains the original hyphenated skill IDs and defaults to this
-60-ID registry; engine procedures remain available through library navigation
+61-ID registry; engine procedures remain available through library navigation
 and catalog document pointers.
 
 The registered [h3.c runtime audit](../src/inferrsi/data/skill_library/repositories/h3-c/audit-h3-runtime-residency/SKILL.md)
diff --git a/docs/reference-matrix.md b/docs/reference-matrix.md
index ee72605f6..79629a166 100644
--- a/docs/reference-matrix.md
+++ b/docs/reference-matrix.md
@@ -402,3 +402,9 @@ readset. Original contribution access (2026-10-14), HiCache access (2026-10-15)
 and diffusion archive inspection (2026-10-22) remain separate. This navigation
 adds no upstream reinspection, common dependency tuple or validation result;
 source maps, notices and imported/candidate states stay with their owners.
+
+## DwarfStar portable AREX extraction
+
+| Source / revision | Inspected scope | Reuse, extraction and limits |
+| --- | --- | --- |
+| antirez/ds4 @ `ee7e8405232c8971000cc7ba2b945b00bf6d20d3`; AREX-Skill method @ `ac3fe1afa80fb9a09775ecfb2b6cc3ba850a2db6` | Selected SSD budgeting, shared prefill workspace, session/token/image identity, disk KV/tool replay, speculation and topology/test contracts | [DwarfStar operating graph](../src/inferrsi/data/skill_library/repositories/dwarfstar/SKILL.md) with bounded source hashes, bundled request helper and owner-local references. Actual portable create/verify method, not an unexecuted CLI run. Native C adaptation and partial drafting were user-approved. Accelerated/model/physical-distributed/quality/performance verification remains blocked; CPU checks do not establish those claims. Source MIT notice retained; `NO_LICENSE` is the unauthenticated resolver fallback, not a legal conclusion. |
diff --git a/docs/searchspace-matrix.md b/docs/searchspace-matrix.md
index d18f98128..32f09e131 100644
--- a/docs/searchspace-matrix.md
+++ b/docs/searchspace-matrix.md
@@ -439,3 +439,10 @@ Network coverage. Existing resource cells and maturity boundaries remain unchang
 This adds conditional Hopper guidance without new measured or green coverage.
 Registers, SMEM and HBM remain device resources; local CTA/cluster coordination
 and tile scheduling add no RAM/Host, SSD, Network or OS-control coverage.
+
+## DwarfStar source-derived execution/state graph
+
+| Resource | Stage / named mechanism | Canonical owner | Evidence boundary |
+| --- | --- | --- | --- |
+| RAM / Accelerator / SSD | Diagnose / Configure: routed-expert cache versus non-routed weights, context, shared prefill workspace and per-session recurrent/KV state | [DwarfStar runtime-memory](../src/inferrsi/data/skill_library/repositories/dwarfstar/sub-skills/runtime-memory/SKILL.md) | Bounded source-derived budgeting guidance; native GPU/model SSD I/O and actual peak memory are not validated. |
+| Host / Accelerator / service boundary | Diagnose / Verify: token and image prefix identity, disk payload compatibility, tool replay, speculation rollback and actual batching/fallback path | [DwarfStar serving-state](../src/inferrsi/data/skill_library/repositories/dwarfstar/sub-skills/serving-state/SKILL.md) and [qualification](../src/inferrsi/data/skill_library/repositories/dwarfstar/sub-skills/qualification/SKILL.md) | Model-free CPU bookkeeping and source contracts are distinct from accelerated correctness, physical TCP/RDMA or service performance. This adds no measured Network coverage. |
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/README.md b/src/inferrsi/data/repo_packs/inferrsi/knowledge/README.md
index 9a36368af..b6e9fe0cf 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/README.md
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/README.md
@@ -1,6 +1,6 @@
 # InferRSI skill discovery index
 
-Generated abstracts for all 60 registered skills in `data/skill_library/`.
+Generated abstracts for all 61 registered skills in `data/skill_library/`.
 Their complete procedures, references and cases ship in the same package.
 `skill-index.json` maps stable skill IDs to categorized directories. The
 reference and searchspace matrices
@@ -15,7 +15,7 @@ These are knowledge abstracts (`--kind knowledge`), with namespaced package
 IDs distinct from explicitly selected `checkout_ref` skills. Ascend/CUDA
 procedures use their corresponding backend; shared methods use `common`.
 
-Source snapshot: [f321699b42a490276ffa4ea52350591683b575cc](https://github.com/InferMatrixLab/InferRSI/tree/f321699b42a490276ffa4ea52350591683b575cc/src/inferrsi/data/skill_library).
+Source snapshot: [3291ffbeab2fe72b217d21c8cac282a4f5348267](https://github.com/InferMatrixLab/InferRSI/tree/3291ffbeab2fe72b217d21c8cac282a4f5348267/src/inferrsi/data/skill_library).
 
 ```bash
 inferrsi knowledge search 'profiling' --repo-type inferrsi --limit 50
@@ -29,7 +29,7 @@ contain the current input paths and bytes (Git checkout filters are respected).
 
 ```bash
 python tools/build_skill_index.py --source-revision <committed-source-revision>
-python tools/build_skill_index.py --source-revision f321699b42a490276ffa4ea52350591683b575cc --check
+python tools/build_skill_index.py --source-revision 3291ffbeab2fe72b217d21c8cac282a4f5348267 --check
 ```
 
 `--check` detects missing, stale and unexpected generated records without writing.
@@ -72,6 +72,7 @@ References and cases ship with their owning skill, not as duplicate knowledge ab
 | [diagnose-vllm-admission-pressure](common/skills/diagnose-vllm-admission-pressure.json) | common | candidate |
 | [diffusers-invocation-state](common/skills/diffusers-invocation-state.json) | common | candidate |
 | [draft-acceptance-issue](common/skills/draft-acceptance-issue.json) | common | unspecified |
+| [dwarfstar](common/skills/dwarfstar.json) | common | candidate |
 | [evaluate-elastic-serving-placement](common/skills/evaluate-elastic-serving-placement.json) | common | candidate |
 | [generation-cache-lifecycle](common/skills/generation-cache-lifecycle.json) | common | candidate |
 | [host-memory-management](common/skills/host-memory-management.json) | common | candidate |
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/ascend-aclgraph-capture-replay.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/ascend-aclgraph-capture-replay.json
index b9e2555b2..26a2a7c4a 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/ascend-aclgraph-capture-replay.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/ascend-aclgraph-capture-replay.json
@@ -21,8 +21,8 @@
   "sources": [
     {
       "type": "git",
-      "url": "https://github.com/InferMatrixLab/InferRSI/blob/f321699b42a490276ffa4ea52350591683b575cc/src/inferrsi/data/skill_library/hardware/ascend/ascend-aclgraph-capture-replay/SKILL.md",
-      "commit": "f321699b42a490276ffa4ea52350591683b575cc",
+      "url": "https://github.com/InferMatrixLab/InferRSI/blob/3291ffbeab2fe72b217d21c8cac282a4f5348267/src/inferrsi/data/skill_library/hardware/ascend/ascend-aclgraph-capture-replay/SKILL.md",
+      "commit": "3291ffbeab2fe72b217d21c8cac282a4f5348267",
       "paths": [
         "src/inferrsi/data/skill_library/hardware/ascend/ascend-aclgraph-capture-replay/SKILL.md",
         "src/inferrsi/data/skill_library/hardware/ascend/ascend-aclgraph-capture-replay/metadata.json",
@@ -41,7 +41,7 @@
     "src/inferrsi/data/skill_library/hardware/ascend/ascend-aclgraph-capture-replay/SKILL.md": "d494fc7e539cbb7388189bd7b3f04bb5c4983fdd7a82484348aad1cee7f741f2",
     "src/inferrsi/data/skill_library/hardware/ascend/ascend-aclgraph-capture-replay/metadata.json": "71c68431663ed2336cccc985a354adab6b13b4cfd6123a18a5d8d53c28268509",
     "src/inferrsi/data/skill_library/hardware/ascend/ascend-aclgraph-capture-replay/references/source-notes.md": "6ffbfb68cc7e04923b4fbca2a73a7555b7c3b45a6f40819bb5f3213ef05cf88d",
-    "src/inferrsi/data/skill_library/skill-index.json": "8ce35e18f7629260d67a95617b9918eecff73ed8d9370dd3d4595c725ada9aa1"
+    "src/inferrsi/data/skill_library/skill-index.json": "ea80d57e29a2949cffceefcc1715db375937499144864b9df41c3e7fcae61c16"
   },
   "skill_metadata": {
     "name": "ascend-aclgraph-capture-replay",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/ascend-device-health.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/ascend-device-health.json
index 9c1ee5b64..7947c9ff7 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/ascend-device-health.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/ascend-device-health.json
@@ -23,8 +23,8 @@
   "sources": [
     {
       "type": "git",
-      "url": "https://github.com/InferMatrixLab/InferRSI/blob/f321699b42a490276ffa4ea52350591683b575cc/src/inferrsi/data/skill_library/hardware/ascend/ascend-device-health/SKILL.md",
-      "commit": "f321699b42a490276ffa4ea52350591683b575cc",
+      "url": "https://github.com/InferMatrixLab/InferRSI/blob/3291ffbeab2fe72b217d21c8cac282a4f5348267/src/inferrsi/data/skill_library/hardware/ascend/ascend-device-health/SKILL.md",
+      "commit": "3291ffbeab2fe72b217d21c8cac282a4f5348267",
       "paths": [
         "src/inferrsi/data/skill_library/hardware/ascend/ascend-device-health/SKILL.md",
         "src/inferrsi/data/skill_library/hardware/ascend/ascend-device-health/metadata.json",
@@ -41,7 +41,7 @@
   "source_files": {
     "src/inferrsi/data/skill_library/hardware/ascend/ascend-device-health/SKILL.md": "03829039860ece35519f2e4c339a48b5014ec25e07a2c150a3efb6a3a4e35b4a",
     "src/inferrsi/data/skill_library/hardware/ascend/ascend-device-health/metadata.json": "80d80c15514919826fa72d6f6cf26b76db8accd85d21187fc239232ca454e7bf",
-    "src/inferrsi/data/skill_library/skill-index.json": "8ce35e18f7629260d67a95617b9918eecff73ed8d9370dd3d4595c725ada9aa1"
+    "src/inferrsi/data/skill_library/skill-index.json": "ea80d57e29a2949cffceefcc1715db375937499144864b9df41c3e7fcae61c16"
   },
   "skill_metadata": {
     "name": "ascend-device-health",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/ascend-distributed-parallel.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/ascend-distributed-parallel.json
index 0347b550a..2791af40e 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/ascend-distributed-parallel.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/ascend-distributed-parallel.json
@@ -24,8 +24,8 @@
   "sources": [
     {
       "type": "git",
-      "url": "https://github.com/InferMatrixLab/InferRSI/blob/f321699b42a490276ffa4ea52350591683b575cc/src/inferrsi/data/skill_library/hardware/ascend/ascend-distributed-parallel/SKILL.md",
-      "commit": "f321699b42a490276ffa4ea52350591683b575cc",
+      "url": "https://github.com/InferMatrixLab/InferRSI/blob/3291ffbeab2fe72b217d21c8cac282a4f5348267/src/inferrsi/data/skill_library/hardware/ascend/ascend-distributed-parallel/SKILL.md",
+      "commit": "3291ffbeab2fe72b217d21c8cac282a4f5348267",
       "paths": [
         "src/inferrsi/data/skill_library/hardware/ascend/ascend-distributed-parallel/SKILL.md",
         "src/inferrsi/data/skill_library/hardware/ascend/ascend-distributed-parallel/metadata.json",
@@ -44,7 +44,7 @@
     "src/inferrsi/data/skill_library/hardware/ascend/ascend-distributed-parallel/SKILL.md": "4ce9dcadd1ae3c7a8cc633336760d0a3c382f354f1b7028e3a10d28fa0d0c9fa",
     "src/inferrsi/data/skill_library/hardware/ascend/ascend-distributed-parallel/metadata.json": "b6cb25879b1bca5f5af08b48776f2e664cf5dcd11a0efb1c4f56879e2afa9006",
     "src/inferrsi/data/skill_library/hardware/ascend/ascend-distributed-parallel/references/awesome-ascend-communication-tests.md": "0cb27a3d24409171bb414d7ebf0c49ad62be5f46482f29624695496ba9eb3e63",
-    "src/inferrsi/data/skill_library/skill-index.json": "8ce35e18f7629260d67a95617b9918eecff73ed8d9370dd3d4595c725ada9aa1"
+    "src/inferrsi/data/skill_library/skill-index.json": "ea80d57e29a2949cffceefcc1715db375937499144864b9df41c3e7fcae61c16"
   },
   "skill_metadata": {
     "name": "ascend-distributed-parallel",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/ascend-graph-mode.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/ascend-graph-mode.json
index 0dcde5201..d06b587cf 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/ascend-graph-mode.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/ascend-graph-mode.json
@@ -22,8 +22,8 @@
   "sources": [
     {
       "type": "git",
-      "url": "https://github.com/InferMatrixLab/InferRSI/blob/f321699b42a490276ffa4ea52350591683b575cc/src/inferrsi/data/skill_library/hardware/ascend/ascend-graph-mode/SKILL.md",
-      "commit": "f321699b42a490276ffa4ea52350591683b575cc",
+      "url": "https://github.com/InferMatrixLab/InferRSI/blob/3291ffbeab2fe72b217d21c8cac282a4f5348267/src/inferrsi/data/skill_library/hardware/ascend/ascend-graph-mode/SKILL.md",
+      "commit": "3291ffbeab2fe72b217d21c8cac282a4f5348267",
       "paths": [
         "src/inferrsi/data/skill_library/hardware/ascend/ascend-graph-mode/SKILL.md",
         "src/inferrsi/data/skill_library/hardware/ascend/ascend-graph-mode/metadata.json",
@@ -40,7 +40,7 @@
   "source_files": {
     "src/inferrsi/data/skill_library/hardware/ascend/ascend-graph-mode/SKILL.md": "9203cc5a14a1d65df5fbf275734b93be023e06527459b2f79864f28a338cb44a",
     "src/inferrsi/data/skill_library/hardware/ascend/ascend-graph-mode/metadata.json": "72b41e7845915d34b09219624e82976386f3125816293540c2c38143fe6b477f",
-    "src/inferrsi/data/skill_library/skill-index.json": "8ce35e18f7629260d67a95617b9918eecff73ed8d9370dd3d4595c725ada9aa1"
+    "src/inferrsi/data/skill_library/skill-index.json": "ea80d57e29a2949cffceefcc1715db375937499144864b9df41c3e7fcae61c16"
   },
   "skill_metadata": {
     "name": "ascend-graph-mode",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/ascend-model-adaptation.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/ascend-model-adaptation.json
index 622f3db24..24616eca0 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/ascend-model-adaptation.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/ascend-model-adaptation.json
@@ -22,8 +22,8 @@
   "sources": [
     {
       "type": "git",
-      "url": "https://github.com/InferMatrixLab/InferRSI/blob/f321699b42a490276ffa4ea52350591683b575cc/src/inferrsi/data/skill_library/hardware/ascend/ascend-model-adaptation/SKILL.md",
-      "commit": "f321699b42a490276ffa4ea52350591683b575cc",
+      "url": "https://github.com/InferMatrixLab/InferRSI/blob/3291ffbeab2fe72b217d21c8cac282a4f5348267/src/inferrsi/data/skill_library/hardware/ascend/ascend-model-adaptation/SKILL.md",
+      "commit": "3291ffbeab2fe72b217d21c8cac282a4f5348267",
       "paths": [
         "src/inferrsi/data/skill_library/hardware/ascend/ascend-model-adaptation/SKILL.md",
         "src/inferrsi/data/skill_library/hardware/ascend/ascend-model-adaptation/metadata.json",
@@ -40,7 +40,7 @@
   "source_files": {
     "src/inferrsi/data/skill_library/hardware/ascend/ascend-model-adaptation/SKILL.md": "8fdf46fb2743b67a66bc069d22a45305a222debfb606777bc451b4b8f31d9cc6",
     "src/inferrsi/data/skill_library/hardware/ascend/ascend-model-adaptation/metadata.json": "be681a3206e5e183b4c9670c9d7784bcffce18c891449e5b15f94ca5f1324b9a",
-    "src/inferrsi/data/skill_library/skill-index.json": "8ce35e18f7629260d67a95617b9918eecff73ed8d9370dd3d4595c725ada9aa1"
+    "src/inferrsi/data/skill_library/skill-index.json": "ea80d57e29a2949cffceefcc1715db375937499144864b9df41c3e7fcae61c16"
   },
   "skill_metadata": {
     "name": "ascend-model-adaptation",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/ascend-operator-profiling.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/ascend-operator-profiling.json
index 903de41c8..a4fe7aed8 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/ascend-operator-profiling.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/ascend-operator-profiling.json
@@ -23,8 +23,8 @@
   "sources": [
     {
       "type": "git",
-      "url": "https://github.com/InferMatrixLab/InferRSI/blob/f321699b42a490276ffa4ea52350591683b575cc/src/inferrsi/data/skill_library/hardware/ascend/ascend-operator-profiling/SKILL.md",
-      "commit": "f321699b42a490276ffa4ea52350591683b575cc",
+      "url": "https://github.com/InferMatrixLab/InferRSI/blob/3291ffbeab2fe72b217d21c8cac282a4f5348267/src/inferrsi/data/skill_library/hardware/ascend/ascend-operator-profiling/SKILL.md",
+      "commit": "3291ffbeab2fe72b217d21c8cac282a4f5348267",
       "paths": [
         "src/inferrsi/data/skill_library/hardware/ascend/ascend-operator-profiling/SKILL.md",
         "src/inferrsi/data/skill_library/hardware/ascend/ascend-operator-profiling/metadata.json",
@@ -41,7 +41,7 @@
   "source_files": {
     "src/inferrsi/data/skill_library/hardware/ascend/ascend-operator-profiling/SKILL.md": "401981cf7828edcab0c9ea882f7ecd83badfa3c2a95006e912cf99d67b3d6a12",
     "src/inferrsi/data/skill_library/hardware/ascend/ascend-operator-profiling/metadata.json": "efd37bb23acb500dad4ec8ebe7ddfe5c3f196374052825f923ac85da8445f06b",
-    "src/inferrsi/data/skill_library/skill-index.json": "8ce35e18f7629260d67a95617b9918eecff73ed8d9370dd3d4595c725ada9aa1"
+    "src/inferrsi/data/skill_library/skill-index.json": "ea80d57e29a2949cffceefcc1715db375937499144864b9df41c3e7fcae61c16"
   },
   "skill_metadata": {
     "name": "ascend-operator-profiling",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/ascend-operator-tuning.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/ascend-operator-tuning.json
index 955eebbb1..d406f93b0 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/ascend-operator-tuning.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/ascend-operator-tuning.json
@@ -24,8 +24,8 @@
   "sources": [
     {
       "type": "git",
-      "url": "https://github.com/InferMatrixLab/InferRSI/blob/f321699b42a490276ffa4ea52350591683b575cc/src/inferrsi/data/skill_library/hardware/ascend/ascend-operator-tuning/SKILL.md",
-      "commit": "f321699b42a490276ffa4ea52350591683b575cc",
+      "url": "https://github.com/InferMatrixLab/InferRSI/blob/3291ffbeab2fe72b217d21c8cac282a4f5348267/src/inferrsi/data/skill_library/hardware/ascend/ascend-operator-tuning/SKILL.md",
+      "commit": "3291ffbeab2fe72b217d21c8cac282a4f5348267",
       "paths": [
         "src/inferrsi/data/skill_library/hardware/ascend/ascend-operator-tuning/SKILL.md",
         "src/inferrsi/data/skill_library/hardware/ascend/ascend-operator-tuning/metadata.json",
@@ -42,7 +42,7 @@
   "source_files": {
     "src/inferrsi/data/skill_library/hardware/ascend/ascend-operator-tuning/SKILL.md": "38b151bc6b4e9734c6fa41efb7a30343559400c0c75bbdd26bd8a968b8ea4592",
     "src/inferrsi/data/skill_library/hardware/ascend/ascend-operator-tuning/metadata.json": "04d2002fc8a0a7becf86b0b73daaf2da8716ab466d7f988da43d41a10f26680e",
-    "src/inferrsi/data/skill_library/skill-index.json": "8ce35e18f7629260d67a95617b9918eecff73ed8d9370dd3d4595c725ada9aa1"
+    "src/inferrsi/data/skill_library/skill-index.json": "ea80d57e29a2949cffceefcc1715db375937499144864b9df41c3e7fcae61c16"
   },
   "skill_metadata": {
     "name": "ascend-operator-tuning",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/ascend-optimization-casebook.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/ascend-optimization-casebook.json
index 9b988efdd..4c5328017 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/ascend-optimization-casebook.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/ascend-optimization-casebook.json
@@ -22,8 +22,8 @@
   "sources": [
     {
       "type": "git",
-      "url": "https://github.com/InferMatrixLab/InferRSI/blob/f321699b42a490276ffa4ea52350591683b575cc/src/inferrsi/data/skill_library/hardware/ascend/ascend-optimization-casebook/SKILL.md",
-      "commit": "f321699b42a490276ffa4ea52350591683b575cc",
+      "url": "https://github.com/InferMatrixLab/InferRSI/blob/3291ffbeab2fe72b217d21c8cac282a4f5348267/src/inferrsi/data/skill_library/hardware/ascend/ascend-optimization-casebook/SKILL.md",
+      "commit": "3291ffbeab2fe72b217d21c8cac282a4f5348267",
       "paths": [
         "src/inferrsi/data/skill_library/hardware/ascend/ascend-optimization-casebook/SKILL.md",
         "src/inferrsi/data/skill_library/hardware/ascend/ascend-optimization-casebook/metadata.json",
@@ -40,7 +40,7 @@
   "source_files": {
     "src/inferrsi/data/skill_library/hardware/ascend/ascend-optimization-casebook/SKILL.md": "647f5e9a288f2c4e1617ba53df59d9d827df749eb4fe86182d21a9c41175caf8",
     "src/inferrsi/data/skill_library/hardware/ascend/ascend-optimization-casebook/metadata.json": "f425f7c1abbd616387273e8932c6bd02f4b1da012f099286ad4171e0f6037b22",
-    "src/inferrsi/data/skill_library/skill-index.json": "8ce35e18f7629260d67a95617b9918eecff73ed8d9370dd3d4595c725ada9aa1"
+    "src/inferrsi/data/skill_library/skill-index.json": "ea80d57e29a2949cffceefcc1715db375937499144864b9df41c3e7fcae61c16"
   },
   "skill_metadata": {
     "name": "ascend-optimization-casebook",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/ascend-performance-analysis.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/ascend-performance-analysis.json
index b823bd746..5e2771a4a 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/ascend-performance-analysis.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/ascend-performance-analysis.json
@@ -22,8 +22,8 @@
   "sources": [
     {
       "type": "git",
-      "url": "https://github.com/InferMatrixLab/InferRSI/blob/f321699b42a490276ffa4ea52350591683b575cc/src/inferrsi/data/skill_library/hardware/ascend/ascend-performance-analysis/SKILL.md",
-      "commit": "f321699b42a490276ffa4ea52350591683b575cc",
+      "url": "https://github.com/InferMatrixLab/InferRSI/blob/3291ffbeab2fe72b217d21c8cac282a4f5348267/src/inferrsi/data/skill_library/hardware/ascend/ascend-performance-analysis/SKILL.md",
+      "commit": "3291ffbeab2fe72b217d21c8cac282a4f5348267",
       "paths": [
         "src/inferrsi/data/skill_library/hardware/ascend/ascend-performance-analysis/SKILL.md",
         "src/inferrsi/data/skill_library/hardware/ascend/ascend-performance-analysis/metadata.json",
@@ -44,7 +44,7 @@
     "src/inferrsi/data/skill_library/hardware/ascend/ascend-performance-analysis/metadata.json": "9e705625c8308013c287576b0247da6f6116d817a4a535cbecff80abd6ea35ee",
     "src/inferrsi/data/skill_library/hardware/ascend/ascend-performance-analysis/references/cannbot-CANN-2.0-LICENSE.md": "89b1c114df62cfd183004f38d18d07f186436c3b37203dd69c3e84becf9d2282",
     "src/inferrsi/data/skill_library/hardware/ascend/ascend-performance-analysis/references/cannbot-stage-evidence.md": "f56e4125325c9c1b9235a1d64fd732861c7686444492422daff44ba8dd648be4",
-    "src/inferrsi/data/skill_library/skill-index.json": "8ce35e18f7629260d67a95617b9918eecff73ed8d9370dd3d4595c725ada9aa1"
+    "src/inferrsi/data/skill_library/skill-index.json": "ea80d57e29a2949cffceefcc1715db375937499144864b9df41c3e7fcae61c16"
   },
   "skill_metadata": {
     "name": "ascend-performance-analysis",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/ascendc-tiling-design.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/ascendc-tiling-design.json
index e108ed0b1..88ae38bb8 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/ascendc-tiling-design.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/ascendc-tiling-design.json
@@ -24,8 +24,8 @@
   "sources": [
     {
       "type": "git",
-      "url": "https://github.com/InferMatrixLab/InferRSI/blob/f321699b42a490276ffa4ea52350591683b575cc/src/inferrsi/data/skill_library/hardware/ascend/ascendc-tiling-design/SKILL.md",
-      "commit": "f321699b42a490276ffa4ea52350591683b575cc",
+      "url": "https://github.com/InferMatrixLab/InferRSI/blob/3291ffbeab2fe72b217d21c8cac282a4f5348267/src/inferrsi/data/skill_library/hardware/ascend/ascendc-tiling-design/SKILL.md",
+      "commit": "3291ffbeab2fe72b217d21c8cac282a4f5348267",
       "paths": [
         "src/inferrsi/data/skill_library/hardware/ascend/ascendc-tiling-design/SKILL.md",
         "src/inferrsi/data/skill_library/hardware/ascend/ascendc-tiling-design/metadata.json",
@@ -42,7 +42,7 @@
   "source_files": {
     "src/inferrsi/data/skill_library/hardware/ascend/ascendc-tiling-design/SKILL.md": "93aa1b775ced304cebfe6c4dec0503ddb2260ecb3cb01edc17ce329446c5ce2b",
     "src/inferrsi/data/skill_library/hardware/ascend/ascendc-tiling-design/metadata.json": "f6d94f104cb70969024b2f4542078a240e200712c66279eca8b94f60bee1dd71",
-    "src/inferrsi/data/skill_library/skill-index.json": "8ce35e18f7629260d67a95617b9918eecff73ed8d9370dd3d4595c725ada9aa1"
+    "src/inferrsi/data/skill_library/skill-index.json": "ea80d57e29a2949cffceefcc1715db375937499144864b9df41c3e7fcae61c16"
   },
   "skill_metadata": {
     "name": "ascendc-tiling-design",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/cann-attribution-overlap.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/cann-attribution-overlap.json
index a4b5922fb..01e73a8fa 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/cann-attribution-overlap.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/cann-attribution-overlap.json
@@ -24,8 +24,8 @@
   "sources": [
     {
       "type": "git",
-      "url": "https://github.com/InferMatrixLab/InferRSI/blob/f321699b42a490276ffa4ea52350591683b575cc/src/inferrsi/data/skill_library/hardware/ascend/cann-attribution-overlap/SKILL.md",
-      "commit": "f321699b42a490276ffa4ea52350591683b575cc",
+      "url": "https://github.com/InferMatrixLab/InferRSI/blob/3291ffbeab2fe72b217d21c8cac282a4f5348267/src/inferrsi/data/skill_library/hardware/ascend/cann-attribution-overlap/SKILL.md",
+      "commit": "3291ffbeab2fe72b217d21c8cac282a4f5348267",
       "paths": [
         "src/inferrsi/data/skill_library/hardware/ascend/cann-attribution-overlap/SKILL.md",
         "src/inferrsi/data/skill_library/hardware/ascend/cann-attribution-overlap/metadata.json",
@@ -42,7 +42,7 @@
   "source_files": {
     "src/inferrsi/data/skill_library/hardware/ascend/cann-attribution-overlap/SKILL.md": "0e6f094dc554abea12e07def50f20b233b47127779a8aefad595b4ab1efd24c2",
     "src/inferrsi/data/skill_library/hardware/ascend/cann-attribution-overlap/metadata.json": "acccc8eb86de0df6caf78a9b3c6db76a2a686adc4a5ba0029a4d62d50ff9a8f8",
-    "src/inferrsi/data/skill_library/skill-index.json": "8ce35e18f7629260d67a95617b9918eecff73ed8d9370dd3d4595c725ada9aa1"
+    "src/inferrsi/data/skill_library/skill-index.json": "ea80d57e29a2949cffceefcc1715db375937499144864b9df41c3e7fcae61c16"
   },
   "skill_metadata": {
     "name": "cann-attribution-overlap",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/torchnpu-device-memory-lifecycle.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/torchnpu-device-memory-lifecycle.json
index 9a44c490b..52e78c172 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/torchnpu-device-memory-lifecycle.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/torchnpu-device-memory-lifecycle.json
@@ -22,8 +22,8 @@
   "sources": [
     {
       "type": "git",
-      "url": "https://github.com/InferMatrixLab/InferRSI/blob/f321699b42a490276ffa4ea52350591683b575cc/src/inferrsi/data/skill_library/hardware/ascend/torchnpu-device-memory-lifecycle/SKILL.md",
-      "commit": "f321699b42a490276ffa4ea52350591683b575cc",
+      "url": "https://github.com/InferMatrixLab/InferRSI/blob/3291ffbeab2fe72b217d21c8cac282a4f5348267/src/inferrsi/data/skill_library/hardware/ascend/torchnpu-device-memory-lifecycle/SKILL.md",
+      "commit": "3291ffbeab2fe72b217d21c8cac282a4f5348267",
       "paths": [
         "src/inferrsi/data/skill_library/hardware/ascend/torchnpu-device-memory-lifecycle/SKILL.md",
         "src/inferrsi/data/skill_library/hardware/ascend/torchnpu-device-memory-lifecycle/metadata.json",
@@ -42,7 +42,7 @@
     "src/inferrsi/data/skill_library/hardware/ascend/torchnpu-device-memory-lifecycle/SKILL.md": "8428df7157af024fa17cfe0b3818ed54bd9f468306dda46eb8d4f0edf4f3f230",
     "src/inferrsi/data/skill_library/hardware/ascend/torchnpu-device-memory-lifecycle/metadata.json": "c650dce88c07b636d68e063dbda4d319f6f7f3ecbcdd5e47f393cbfc3af7c71d",
     "src/inferrsi/data/skill_library/hardware/ascend/torchnpu-device-memory-lifecycle/references/source-notes.md": "80cba4d3a884e7ee2618326cf74b9f9c3a34e4d9cf443707f6bbe673f2d5ba65",
-    "src/inferrsi/data/skill_library/skill-index.json": "8ce35e18f7629260d67a95617b9918eecff73ed8d9370dd3d4595c725ada9aa1"
+    "src/inferrsi/data/skill_library/skill-index.json": "ea80d57e29a2949cffceefcc1715db375937499144864b9df41c3e7fcae61c16"
   },
   "skill_metadata": {
     "name": "torchnpu-device-memory-lifecycle",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/torchnpu-pinned-host-memory-lifecycle.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/torchnpu-pinned-host-memory-lifecycle.json
index c3601b49b..e9d7a4009 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/torchnpu-pinned-host-memory-lifecycle.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/torchnpu-pinned-host-memory-lifecycle.json
@@ -22,8 +22,8 @@
   "sources": [
     {
       "type": "git",
-      "url": "https://github.com/InferMatrixLab/InferRSI/blob/f321699b42a490276ffa4ea52350591683b575cc/src/inferrsi/data/skill_library/hardware/ascend/torchnpu-pinned-host-memory-lifecycle/SKILL.md",
-      "commit": "f321699b42a490276ffa4ea52350591683b575cc",
+      "url": "https://github.com/InferMatrixLab/InferRSI/blob/3291ffbeab2fe72b217d21c8cac282a4f5348267/src/inferrsi/data/skill_library/hardware/ascend/torchnpu-pinned-host-memory-lifecycle/SKILL.md",
+      "commit": "3291ffbeab2fe72b217d21c8cac282a4f5348267",
       "paths": [
         "src/inferrsi/data/skill_library/hardware/ascend/torchnpu-pinned-host-memory-lifecycle/SKILL.md",
         "src/inferrsi/data/skill_library/hardware/ascend/torchnpu-pinned-host-memory-lifecycle/metadata.json",
@@ -42,7 +42,7 @@
     "src/inferrsi/data/skill_library/hardware/ascend/torchnpu-pinned-host-memory-lifecycle/SKILL.md": "23e93ea7c98f99743509fdc04139d78ed2e7123c63e80b366c6beaba7eb70d0d",
     "src/inferrsi/data/skill_library/hardware/ascend/torchnpu-pinned-host-memory-lifecycle/metadata.json": "97836a38465106d816a9ace894a7b277d6d330f9accade8f8493ce6f0fb57ef9",
     "src/inferrsi/data/skill_library/hardware/ascend/torchnpu-pinned-host-memory-lifecycle/references/source-notes.md": "435193590b4a2ba32e2e4c6fa9d7e8f889528d412cbbb48e35ec5433f7dab384",
-    "src/inferrsi/data/skill_library/skill-index.json": "8ce35e18f7629260d67a95617b9918eecff73ed8d9370dd3d4595c725ada9aa1"
+    "src/inferrsi/data/skill_library/skill-index.json": "ea80d57e29a2949cffceefcc1715db375937499144864b9df41c3e7fcae61c16"
   },
   "skill_metadata": {
     "name": "torchnpu-pinned-host-memory-lifecycle",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/assess-serving-preemption.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/assess-serving-preemption.json
index ae17b35fc..71caf568d 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/assess-serving-preemption.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/assess-serving-preemption.json
@@ -16,8 +16,8 @@
   "sources": [
     {
       "type": "git",
-      "url": "https://github.com/InferMatrixLab/InferRSI/blob/f321699b42a490276ffa4ea52350591683b575cc/src/inferrsi/data/skill_library/methods/assess-serving-preemption/SKILL.md",
-      "commit": "f321699b42a490276ffa4ea52350591683b575cc",
+      "url": "https://github.com/InferMatrixLab/InferRSI/blob/3291ffbeab2fe72b217d21c8cac282a4f5348267/src/inferrsi/data/skill_library/methods/assess-serving-preemption/SKILL.md",
+      "commit": "3291ffbeab2fe72b217d21c8cac282a4f5348267",
       "paths": [
         "src/inferrsi/data/skill_library/methods/assess-serving-preemption/SKILL.md",
         "src/inferrsi/data/skill_library/methods/assess-serving-preemption/metadata.json",
@@ -36,7 +36,7 @@
     "src/inferrsi/data/skill_library/methods/assess-serving-preemption/SKILL.md": "0a5f48d9352c5bf460775ef47f9136829978811f4a16e84983e299782cae305d",
     "src/inferrsi/data/skill_library/methods/assess-serving-preemption/metadata.json": "77ce5ba16245cbd80d46d0abdf80aaf26fc0924fa2d2a5624ff430de68a1daf0",
     "src/inferrsi/data/skill_library/methods/assess-serving-preemption/references/source-notes.md": "f29fcecfba0490008a354c11b42403983fe165e56ff907ad61cc6040d2b6407f",
-    "src/inferrsi/data/skill_library/skill-index.json": "8ce35e18f7629260d67a95617b9918eecff73ed8d9370dd3d4595c725ada9aa1"
+    "src/inferrsi/data/skill_library/skill-index.json": "ea80d57e29a2949cffceefcc1715db375937499144864b9df41c3e7fcae61c16"
   },
   "skill_metadata": {
     "name": "assess-serving-preemption",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/assess-split-backward-schedule.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/assess-split-backward-schedule.json
index e179e127d..67f1b295a 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/assess-split-backward-schedule.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/assess-split-backward-schedule.json
@@ -16,8 +16,8 @@
   "sources": [
     {
       "type": "git",
-      "url": "https://github.com/InferMatrixLab/InferRSI/blob/f321699b42a490276ffa4ea52350591683b575cc/src/inferrsi/data/skill_library/methods/assess-split-backward-schedule/SKILL.md",
-      "commit": "f321699b42a490276ffa4ea52350591683b575cc",
+      "url": "https://github.com/InferMatrixLab/InferRSI/blob/3291ffbeab2fe72b217d21c8cac282a4f5348267/src/inferrsi/data/skill_library/methods/assess-split-backward-schedule/SKILL.md",
+      "commit": "3291ffbeab2fe72b217d21c8cac282a4f5348267",
       "paths": [
         "src/inferrsi/data/skill_library/methods/assess-split-backward-schedule/SKILL.md",
         "src/inferrsi/data/skill_library/methods/assess-split-backward-schedule/metadata.json",
@@ -36,7 +36,7 @@
     "src/inferrsi/data/skill_library/methods/assess-split-backward-schedule/SKILL.md": "1e64f28ea71672fb53c56f6108b87864738c6d03d6dcd32eedaef9e0625e96e3",
     "src/inferrsi/data/skill_library/methods/assess-split-backward-schedule/metadata.json": "c6acae57d70190cea0e18c71c79127962181e58169e011df75d8635b3db4daa7",
     "src/inferrsi/data/skill_library/methods/assess-split-backward-schedule/references/source-notes.md": "61b1609f71c37b61d136b8c4d39f793de87cffb1cf31c8ec737a1c3fa6fb1b4e",
-    "src/inferrsi/data/skill_library/skill-index.json": "8ce35e18f7629260d67a95617b9918eecff73ed8d9370dd3d4595c725ada9aa1"
+    "src/inferrsi/data/skill_library/skill-index.json": "ea80d57e29a2949cffceefcc1715db375937499144864b9df41c3e7fcae61c16"
   },
   "skill_metadata": {
     "name": "assess-split-backward-schedule",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/audit-h3-runtime-residency.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/audit-h3-runtime-residency.json
index 53a2e6004..2afd8d801 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/audit-h3-runtime-residency.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/audit-h3-runtime-residency.json
@@ -30,8 +30,8 @@
   "sources": [
     {
       "type": "git",
-      "url": "https://github.com/InferMatrixLab/InferRSI/blob/f321699b42a490276ffa4ea52350591683b575cc/src/inferrsi/data/skill_library/repositories/h3-c/audit-h3-runtime-residency/SKILL.md",
-      "commit": "f321699b42a490276ffa4ea52350591683b575cc",
+      "url": "https://github.com/InferMatrixLab/InferRSI/blob/3291ffbeab2fe72b217d21c8cac282a4f5348267/src/inferrsi/data/skill_library/repositories/h3-c/audit-h3-runtime-residency/SKILL.md",
+      "commit": "3291ffbeab2fe72b217d21c8cac282a4f5348267",
       "paths": [
         "src/inferrsi/data/skill_library/repositories/h3-c/audit-h3-runtime-residency/LICENSE",
         "src/inferrsi/data/skill_library/repositories/h3-c/audit-h3-runtime-residency/SKILL.md",
@@ -66,7 +66,7 @@
     "src/inferrsi/data/skill_library/repositories/h3-c/audit-h3-runtime-residency/references/source-identity.json": "730f8dd23648a3f8f47b9726cc613d893a79d1817d515eec609a9b06d8a9ff90",
     "src/inferrsi/data/skill_library/repositories/h3-c/audit-h3-runtime-residency/references/source-notes.md": "7304bf9369efc157dc7aa40686dd828cc14db45c49442e908b3d3cc4b329eb86",
     "src/inferrsi/data/skill_library/repositories/h3-c/audit-h3-runtime-residency/references/validation.md": "0754873a758e048480938314894e25b504ae0c269549312e7f8e6020d83fe294",
-    "src/inferrsi/data/skill_library/skill-index.json": "8ce35e18f7629260d67a95617b9918eecff73ed8d9370dd3d4595c725ada9aa1"
+    "src/inferrsi/data/skill_library/skill-index.json": "ea80d57e29a2949cffceefcc1715db375937499144864b9df41c3e7fcae61c16"
   },
   "skill_metadata": {
     "name": "audit-h3-runtime-residency",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/audit-llama-runtime-state.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/audit-llama-runtime-state.json
index 68294a98b..02696941b 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/audit-llama-runtime-state.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/audit-llama-runtime-state.json
@@ -23,8 +23,8 @@
   "sources": [
     {
       "type": "git",
-      "url": "https://github.com/InferMatrixLab/InferRSI/blob/f321699b42a490276ffa4ea52350591683b575cc/src/inferrsi/data/skill_library/repositories/llama-cpp/audit-llama-runtime-state/SKILL.md",
-      "commit": "f321699b42a490276ffa4ea52350591683b575cc",
+      "url": "https://github.com/InferMatrixLab/InferRSI/blob/3291ffbeab2fe72b217d21c8cac282a4f5348267/src/inferrsi/data/skill_library/repositories/llama-cpp/audit-llama-runtime-state/SKILL.md",
+      "commit": "3291ffbeab2fe72b217d21c8cac282a4f5348267",
       "paths": [
         "src/inferrsi/data/skill_library/repositories/llama-cpp/audit-llama-runtime-state/SKILL.md",
         "src/inferrsi/data/skill_library/repositories/llama-cpp/audit-llama-runtime-state/metadata.json",
@@ -43,7 +43,7 @@
     "src/inferrsi/data/skill_library/repositories/llama-cpp/audit-llama-runtime-state/SKILL.md": "95e81b6bc096a954fa73139ea5bc001d1efa6c99747161da62f012b494e5c013",
     "src/inferrsi/data/skill_library/repositories/llama-cpp/audit-llama-runtime-state/metadata.json": "1d30b31dc8352fc6a8d7b8abbef8118c840e47e5b90795dae0c758bf9717278c",
     "src/inferrsi/data/skill_library/repositories/llama-cpp/audit-llama-runtime-state/references/source-notes.md": "3d12a401e9decb55a2f4f3b2bfa0d360ef5110f63a0469dfe285654e78054190",
-    "src/inferrsi/data/skill_library/skill-index.json": "8ce35e18f7629260d67a95617b9918eecff73ed8d9370dd3d4595c725ada9aa1"
+    "src/inferrsi/data/skill_library/skill-index.json": "ea80d57e29a2949cffceefcc1715db375937499144864b9df41c3e7fcae61c16"
   },
   "skill_metadata": {
     "name": "audit-llama-runtime-state",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/audit-sd-cpp-residency.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/audit-sd-cpp-residency.json
index 045a6af3b..96d973075 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/audit-sd-cpp-residency.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/audit-sd-cpp-residency.json
@@ -23,8 +23,8 @@
   "sources": [
     {
       "type": "git",
-      "url": "https://github.com/InferMatrixLab/InferRSI/blob/f321699b42a490276ffa4ea52350591683b575cc/src/inferrsi/data/skill_library/repositories/stable-diffusion-cpp/audit-sd-cpp-residency/SKILL.md",
-      "commit": "f321699b42a490276ffa4ea52350591683b575cc",
+      "url": "https://github.com/InferMatrixLab/InferRSI/blob/3291ffbeab2fe72b217d21c8cac282a4f5348267/src/inferrsi/data/skill_library/repositories/stable-diffusion-cpp/audit-sd-cpp-residency/SKILL.md",
+      "commit": "3291ffbeab2fe72b217d21c8cac282a4f5348267",
       "paths": [
         "src/inferrsi/data/skill_library/repositories/stable-diffusion-cpp/audit-sd-cpp-residency/LICENSE",
         "src/inferrsi/data/skill_library/repositories/stable-diffusion-cpp/audit-sd-cpp-residency/SKILL.md",
@@ -51,7 +51,7 @@
     "src/inferrsi/data/skill_library/repositories/stable-diffusion-cpp/audit-sd-cpp-residency/references/source-identity.json": "af4830370b6eddc2d176a9ae14da4ab56a90b0fe048a01a89ee051df4365d8e9",
     "src/inferrsi/data/skill_library/repositories/stable-diffusion-cpp/audit-sd-cpp-residency/references/source-notes.md": "b67affbd7d237d82b7f51be49f1537e9820fe56ec70a71ee76809f9e599ca693",
     "src/inferrsi/data/skill_library/repositories/stable-diffusion-cpp/audit-sd-cpp-residency/references/validation-boundaries.md": "5dcf5fa7bf734fbe13214f57d2dfb40ff8215d9cd34e072a9e615bf103b82fc2",
-    "src/inferrsi/data/skill_library/skill-index.json": "8ce35e18f7629260d67a95617b9918eecff73ed8d9370dd3d4595c725ada9aa1"
+    "src/inferrsi/data/skill_library/skill-index.json": "ea80d57e29a2949cffceefcc1715db375937499144864b9df41c3e7fcae61c16"
   },
   "skill_metadata": {
     "name": "audit-sd-cpp-residency",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/audit-sglang-diffusion-lifetimes.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/audit-sglang-diffusion-lifetimes.json
index ac57aab19..f384d59e0 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/audit-sglang-diffusion-lifetimes.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/audit-sglang-diffusion-lifetimes.json
@@ -23,8 +23,8 @@
   "sources": [
     {
       "type": "git",
-      "url": "https://github.com/InferMatrixLab/InferRSI/blob/f321699b42a490276ffa4ea52350591683b575cc/src/inferrsi/data/skill_library/repositories/sglang/audit-sglang-diffusion-lifetimes/SKILL.md",
-      "commit": "f321699b42a490276ffa4ea52350591683b575cc",
+      "url": "https://github.com/InferMatrixLab/InferRSI/blob/3291ffbeab2fe72b217d21c8cac282a4f5348267/src/inferrsi/data/skill_library/repositories/sglang/audit-sglang-diffusion-lifetimes/SKILL.md",
+      "commit": "3291ffbeab2fe72b217d21c8cac282a4f5348267",
       "paths": [
         "src/inferrsi/data/skill_library/repositories/sglang/audit-sglang-diffusion-lifetimes/LICENSE",
         "src/inferrsi/data/skill_library/repositories/sglang/audit-sglang-diffusion-lifetimes/SKILL.md",
@@ -45,7 +45,7 @@
     "src/inferrsi/data/skill_library/repositories/sglang/audit-sglang-diffusion-lifetimes/SKILL.md": "b0a4ba4751087126143ef86983798deb0ed00e24b6ff06bf6ef76a9df6e274f2",
     "src/inferrsi/data/skill_library/repositories/sglang/audit-sglang-diffusion-lifetimes/metadata.json": "6ed57dfb4f25ecb34b81f2188b055e05d172ba2581e6a3b605b814b2f733a4d6",
     "src/inferrsi/data/skill_library/repositories/sglang/audit-sglang-diffusion-lifetimes/references/source-notes.md": "d0fcfd1177ae2f351041c5b2abba7f5121cec133006d9e4dca8f74ddb567332d",
-    "src/inferrsi/data/skill_library/skill-index.json": "8ce35e18f7629260d67a95617b9918eecff73ed8d9370dd3d4595c725ada9aa1"
+    "src/inferrsi/data/skill_library/skill-index.json": "ea80d57e29a2949cffceefcc1715db375937499144864b9df41c3e7fcae61c16"
   },
   "skill_metadata": {
     "name": "audit-sglang-diffusion-lifetimes",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/audit-vllm-kv-lifetimes.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/audit-vllm-kv-lifetimes.json
index 1e192cc1a..4efa3ff76 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/audit-vllm-kv-lifetimes.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/audit-vllm-kv-lifetimes.json
@@ -21,8 +21,8 @@
   "sources": [
     {
       "type": "git",
-      "url": "https://github.com/InferMatrixLab/InferRSI/blob/f321699b42a490276ffa4ea52350591683b575cc/src/inferrsi/data/skill_library/repositories/vllm/audit-vllm-kv-lifetimes/SKILL.md",
-      "commit": "f321699b42a490276ffa4ea52350591683b575cc",
+      "url": "https://github.com/InferMatrixLab/InferRSI/blob/3291ffbeab2fe72b217d21c8cac282a4f5348267/src/inferrsi/data/skill_library/repositories/vllm/audit-vllm-kv-lifetimes/SKILL.md",
+      "commit": "3291ffbeab2fe72b217d21c8cac282a4f5348267",
       "paths": [
         "src/inferrsi/data/skill_library/repositories/vllm/audit-vllm-kv-lifetimes/SKILL.md",
         "src/inferrsi/data/skill_library/repositories/vllm/audit-vllm-kv-lifetimes/metadata.json",
@@ -41,7 +41,7 @@
     "src/inferrsi/data/skill_library/repositories/vllm/audit-vllm-kv-lifetimes/SKILL.md": "313733a65cc48442b18dda1c4665139a689963a90e1032a2235bc6f40e3a8e06",
     "src/inferrsi/data/skill_library/repositories/vllm/audit-vllm-kv-lifetimes/metadata.json": "b43fc3a7ccae59c664fcdf9ff77962edcc07c2a3df8b59324db551642bd23e23",
     "src/inferrsi/data/skill_library/repositories/vllm/audit-vllm-kv-lifetimes/references/source-notes.md": "3665b034594fc8cc767c408002f4e16f8803369242405027c7f37ce6c9ecebe9",
-    "src/inferrsi/data/skill_library/skill-index.json": "8ce35e18f7629260d67a95617b9918eecff73ed8d9370dd3d4595c725ada9aa1"
+    "src/inferrsi/data/skill_library/skill-index.json": "ea80d57e29a2949cffceefcc1715db375937499144864b9df41c3e7fcae61c16"
   },
   "skill_metadata": {
     "name": "audit-vllm-kv-lifetimes",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/bound-windowed-attention-replay.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/bound-windowed-attention-replay.json
index 5fe50ff4b..7ce9453f6 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/bound-windowed-attention-replay.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/bound-windowed-attention-replay.json
@@ -23,8 +23,8 @@
   "sources": [
     {
       "type": "git",
-      "url": "https://github.com/InferMatrixLab/InferRSI/blob/f321699b42a490276ffa4ea52350591683b575cc/src/inferrsi/data/skill_library/repositories/deepinfer/bound-windowed-attention-replay/SKILL.md",
-      "commit": "f321699b42a490276ffa4ea52350591683b575cc",
+      "url": "https://github.com/InferMatrixLab/InferRSI/blob/3291ffbeab2fe72b217d21c8cac282a4f5348267/src/inferrsi/data/skill_library/repositories/deepinfer/bound-windowed-attention-replay/SKILL.md",
+      "commit": "3291ffbeab2fe72b217d21c8cac282a4f5348267",
       "paths": [
         "src/inferrsi/data/skill_library/repositories/deepinfer/bound-windowed-attention-replay/SKILL.md",
         "src/inferrsi/data/skill_library/repositories/deepinfer/bound-windowed-attention-replay/metadata.json",
@@ -45,7 +45,7 @@
     "src/inferrsi/data/skill_library/repositories/deepinfer/bound-windowed-attention-replay/metadata.json": "014be24d0e669f4acbc0717dc323eec784026cf185186619d5a19d6d84496131",
     "src/inferrsi/data/skill_library/repositories/deepinfer/bound-windowed-attention-replay/references/source-notes.md": "fd917bd5ac80bee9c9f4525f8e0300f6d67dc144e2a16632af42a67f3d53d098",
     "src/inferrsi/data/skill_library/repositories/deepinfer/bound-windowed-attention-replay/references/validation.md": "60e71e70c5519bcdf6c171eec2be2f342e8801bedd746d6005f11adc646dfaff",
-    "src/inferrsi/data/skill_library/skill-index.json": "8ce35e18f7629260d67a95617b9918eecff73ed8d9370dd3d4595c725ada9aa1"
+    "src/inferrsi/data/skill_library/skill-index.json": "ea80d57e29a2949cffceefcc1715db375937499144864b9df41c3e7fcae61c16"
   },
   "skill_metadata": {
     "name": "bound-windowed-attention-replay",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/budget-speculative-verification.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/budget-speculative-verification.json
index 6fa758e83..3172adff8 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/budget-speculative-verification.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/budget-speculative-verification.json
@@ -23,8 +23,8 @@
   "sources": [
     {
       "type": "git",
-      "url": "https://github.com/InferMatrixLab/InferRSI/blob/f321699b42a490276ffa4ea52350591683b575cc/src/inferrsi/data/skill_library/repositories/deepinfer/budget-speculative-verification/SKILL.md",
-      "commit": "f321699b42a490276ffa4ea52350591683b575cc",
+      "url": "https://github.com/InferMatrixLab/InferRSI/blob/3291ffbeab2fe72b217d21c8cac282a4f5348267/src/inferrsi/data/skill_library/repositories/deepinfer/budget-speculative-verification/SKILL.md",
+      "commit": "3291ffbeab2fe72b217d21c8cac282a4f5348267",
       "paths": [
         "src/inferrsi/data/skill_library/repositories/deepinfer/budget-speculative-verification/SKILL.md",
         "src/inferrsi/data/skill_library/repositories/deepinfer/budget-speculative-verification/metadata.json",
@@ -45,7 +45,7 @@
     "src/inferrsi/data/skill_library/repositories/deepinfer/budget-speculative-verification/metadata.json": "4d68bf4c1767eaee12ade963bbd551d0e61d1bcd81a984bb02fd3aae46965919",
     "src/inferrsi/data/skill_library/repositories/deepinfer/budget-speculative-verification/references/source-notes.md": "6fa955022f16cd598f44fd6521901d1cb868b06cec545cabe2b82769a641b5d6",
     "src/inferrsi/data/skill_library/repositories/deepinfer/budget-speculative-verification/references/validation.md": "673d93f6700052974b340fab25b3438e0e6344960569ac9abec7d7720f662288",
-    "src/inferrsi/data/skill_library/skill-index.json": "8ce35e18f7629260d67a95617b9918eecff73ed8d9370dd3d4595c725ada9aa1"
+    "src/inferrsi/data/skill_library/skill-index.json": "ea80d57e29a2949cffceefcc1715db375937499144864b9df41c3e7fcae61c16"
   },
   "skill_metadata": {
     "name": "budget-speculative-verification",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/calibrate-serving-simulation.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/calibrate-serving-simulation.json
index a0e3f13bc..4ec7e64b6 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/calibrate-serving-simulation.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/calibrate-serving-simulation.json
@@ -16,8 +16,8 @@
   "sources": [
     {
       "type": "git",
-      "url": "https://github.com/InferMatrixLab/InferRSI/blob/f321699b42a490276ffa4ea52350591683b575cc/src/inferrsi/data/skill_library/methods/calibrate-serving-simulation/SKILL.md",
-      "commit": "f321699b42a490276ffa4ea52350591683b575cc",
+      "url": "https://github.com/InferMatrixLab/InferRSI/blob/3291ffbeab2fe72b217d21c8cac282a4f5348267/src/inferrsi/data/skill_library/methods/calibrate-serving-simulation/SKILL.md",
+      "commit": "3291ffbeab2fe72b217d21c8cac282a4f5348267",
       "paths": [
         "src/inferrsi/data/skill_library/methods/calibrate-serving-simulation/SKILL.md",
         "src/inferrsi/data/skill_library/methods/calibrate-serving-simulation/metadata.json",
@@ -36,7 +36,7 @@
     "src/inferrsi/data/skill_library/methods/calibrate-serving-simulation/SKILL.md": "ac773c6f1c110a664dfca28a3ae9129b10c07614e60ca3640dea776220e7b352",
     "src/inferrsi/data/skill_library/methods/calibrate-serving-simulation/metadata.json": "261ce9f644ea6eebb90c830f41fb8b320d2b401456e5bd8d494095a44b8620d0",
     "src/inferrsi/data/skill_library/methods/calibrate-serving-simulation/references/source-notes.md": "9696d3c0ef96b6611787446bd6f96d4d6255ca8d107105de8c1d9af83edb9fa5",
-    "src/inferrsi/data/skill_library/skill-index.json": "8ce35e18f7629260d67a95617b9918eecff73ed8d9370dd3d4595c725ada9aa1"
+    "src/inferrsi/data/skill_library/skill-index.json": "ea80d57e29a2949cffceefcc1715db375937499144864b9df41c3e7fcae61c16"
   },
   "skill_metadata": {
     "name": "calibrate-serving-simulation",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/cgroup-resource-controls.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/cgroup-resource-controls.json
index 229d9c8fc..be4050e56 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/cgroup-resource-controls.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/cgroup-resource-controls.json
@@ -25,8 +25,8 @@
   "sources": [
     {
       "type": "git",
-      "url": "https://github.com/InferMatrixLab/InferRSI/blob/f321699b42a490276ffa4ea52350591683b575cc/src/inferrsi/data/skill_library/hardware/host/cgroup-resource-controls/SKILL.md",
-      "commit": "f321699b42a490276ffa4ea52350591683b575cc",
+      "url": "https://github.com/InferMatrixLab/InferRSI/blob/3291ffbeab2fe72b217d21c8cac282a4f5348267/src/inferrsi/data/skill_library/hardware/host/cgroup-resource-controls/SKILL.md",
+      "commit": "3291ffbeab2fe72b217d21c8cac282a4f5348267",
       "paths": [
         "src/inferrsi/data/skill_library/hardware/host/cgroup-resource-controls/SKILL.md",
         "src/inferrsi/data/skill_library/hardware/host/cgroup-resource-controls/metadata.json",
@@ -43,7 +43,7 @@
   "source_files": {
     "src/inferrsi/data/skill_library/hardware/host/cgroup-resource-controls/SKILL.md": "fd1f5a7e6a55714730044155f5b38c1d700383d4e16c3e91963bbfbe1c773af5",
     "src/inferrsi/data/skill_library/hardware/host/cgroup-resource-controls/metadata.json": "19e40763fedc20990651bb989c32a7693306c5c54af46c3d4be77fd4f87b6c03",
-    "src/inferrsi/data/skill_library/skill-index.json": "8ce35e18f7629260d67a95617b9918eecff73ed8d9370dd3d4595c725ada9aa1"
+    "src/inferrsi/data/skill_library/skill-index.json": "ea80d57e29a2949cffceefcc1715db375937499144864b9df41c3e7fcae61c16"
   },
   "skill_metadata": {
     "name": "cgroup-resource-controls",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/checkpoint-recovery-consistency.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/checkpoint-recovery-consistency.json
index 12cd3793b..9ce0e8b4e 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/checkpoint-recovery-consistency.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/checkpoint-recovery-consistency.json
@@ -22,8 +22,8 @@
   "sources": [
     {
       "type": "git",
-      "url": "https://github.com/InferMatrixLab/InferRSI/blob/f321699b42a490276ffa4ea52350591683b575cc/src/inferrsi/data/skill_library/methods/checkpoint-recovery-consistency/SKILL.md",
-      "commit": "f321699b42a490276ffa4ea52350591683b575cc",
+      "url": "https://github.com/InferMatrixLab/InferRSI/blob/3291ffbeab2fe72b217d21c8cac282a4f5348267/src/inferrsi/data/skill_library/methods/checkpoint-recovery-consistency/SKILL.md",
+      "commit": "3291ffbeab2fe72b217d21c8cac282a4f5348267",
       "paths": [
         "src/inferrsi/data/skill_library/methods/checkpoint-recovery-consistency/SKILL.md",
         "src/inferrsi/data/skill_library/methods/checkpoint-recovery-consistency/metadata.json",
@@ -44,7 +44,7 @@
     "src/inferrsi/data/skill_library/methods/checkpoint-recovery-consistency/metadata.json": "4dc8659806539b4784ea9348a01d4194ac70d0bb297a916f90ff6419efa5c328",
     "src/inferrsi/data/skill_library/methods/checkpoint-recovery-consistency/references/dlrover-persistence-replicas.md": "055f4357f79026fcbcc766bbe77db610376ec353623f8898f5d14a30ee69e7bc",
     "src/inferrsi/data/skill_library/methods/checkpoint-recovery-consistency/references/sources.md": "9425b1aab963a2c8d35f3db897bd15ee69699bf5b0d0c8425c3c8bba3e9a0233",
-    "src/inferrsi/data/skill_library/skill-index.json": "8ce35e18f7629260d67a95617b9918eecff73ed8d9370dd3d4595c725ada9aa1"
+    "src/inferrsi/data/skill_library/skill-index.json": "ea80d57e29a2949cffceefcc1715db375937499144864b9df41c3e7fcae61c16"
   },
   "skill_metadata": {
     "name": "checkpoint-recovery-consistency",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/choose-parallel-training-plan.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/choose-parallel-training-plan.json
index 16a826e4c..056465acc 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/choose-parallel-training-plan.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/choose-parallel-training-plan.json
@@ -16,8 +16,8 @@
   "sources": [
     {
       "type": "git",
-      "url": "https://github.com/InferMatrixLab/InferRSI/blob/f321699b42a490276ffa4ea52350591683b575cc/src/inferrsi/data/skill_library/methods/choose-parallel-training-plan/SKILL.md",
-      "commit": "f321699b42a490276ffa4ea52350591683b575cc",
+      "url": "https://github.com/InferMatrixLab/InferRSI/blob/3291ffbeab2fe72b217d21c8cac282a4f5348267/src/inferrsi/data/skill_library/methods/choose-parallel-training-plan/SKILL.md",
+      "commit": "3291ffbeab2fe72b217d21c8cac282a4f5348267",
       "paths": [
         "src/inferrsi/data/skill_library/methods/choose-parallel-training-plan/SKILL.md",
         "src/inferrsi/data/skill_library/methods/choose-parallel-training-plan/metadata.json",
@@ -36,7 +36,7 @@
     "src/inferrsi/data/skill_library/methods/choose-parallel-training-plan/SKILL.md": "25d04aed9782b2a8ff1186828becdc7fce93c4418039ceaef9b8dcc5ec9bcd24",
     "src/inferrsi/data/skill_library/methods/choose-parallel-training-plan/metadata.json": "c6acae57d70190cea0e18c71c79127962181e58169e011df75d8635b3db4daa7",
     "src/inferrsi/data/skill_library/methods/choose-parallel-training-plan/references/source-notes.md": "9f9184d46b19312138a6e805e439e43dc0594a0cc6b46675255a582dc946ca5b",
-    "src/inferrsi/data/skill_library/skill-index.json": "8ce35e18f7629260d67a95617b9918eecff73ed8d9370dd3d4595c725ada9aa1"
+    "src/inferrsi/data/skill_library/skill-index.json": "ea80d57e29a2949cffceefcc1715db375937499144864b9df41c3e7fcae61c16"
   },
   "skill_metadata": {
     "name": "choose-parallel-training-plan",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/cite-check-downstream-pr.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/cite-check-downstream-pr.json
index 17d0e43a4..9c601948f 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/cite-check-downstream-pr.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/cite-check-downstream-pr.json
@@ -16,8 +16,8 @@
   "sources": [
     {
       "type": "git",
-      "url": "https://github.com/InferMatrixLab/InferRSI/blob/f321699b42a490276ffa4ea52350591683b575cc/src/inferrsi/data/skill_library/workflows/cite-check-downstream-pr/SKILL.md",
-      "commit": "f321699b42a490276ffa4ea52350591683b575cc",
+      "url": "https://github.com/InferMatrixLab/InferRSI/blob/3291ffbeab2fe72b217d21c8cac282a4f5348267/src/inferrsi/data/skill_library/workflows/cite-check-downstream-pr/SKILL.md",
+      "commit": "3291ffbeab2fe72b217d21c8cac282a4f5348267",
       "paths": [
         "src/inferrsi/data/skill_library/skill-index.json",
         "src/inferrsi/data/skill_library/workflows/cite-check-downstream-pr/SKILL.md"
@@ -31,7 +31,7 @@
   },
   "checkout_ref": "checkout:cite-check-downstream-pr",
   "source_files": {
-    "src/inferrsi/data/skill_library/skill-index.json": "8ce35e18f7629260d67a95617b9918eecff73ed8d9370dd3d4595c725ada9aa1",
+    "src/inferrsi/data/skill_library/skill-index.json": "ea80d57e29a2949cffceefcc1715db375937499144864b9df41c3e7fcae61c16",
     "src/inferrsi/data/skill_library/workflows/cite-check-downstream-pr/SKILL.md": "485932cdf3b1fba5dc69be73e75d38f86a7b79442d26e82a296c963a7598d226"
   },
   "skill_metadata": {
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/claim-contract.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/claim-contract.json
index 0fa2bfaec..34247e67a 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/claim-contract.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/claim-contract.json
@@ -16,8 +16,8 @@
   "sources": [
     {
       "type": "git",
-      "url": "https://github.com/InferMatrixLab/InferRSI/blob/f321699b42a490276ffa4ea52350591683b575cc/src/inferrsi/data/skill_library/workflows/claim-contract/SKILL.md",
-      "commit": "f321699b42a490276ffa4ea52350591683b575cc",
+      "url": "https://github.com/InferMatrixLab/InferRSI/blob/3291ffbeab2fe72b217d21c8cac282a4f5348267/src/inferrsi/data/skill_library/workflows/claim-contract/SKILL.md",
+      "commit": "3291ffbeab2fe72b217d21c8cac282a4f5348267",
       "paths": [
         "src/inferrsi/data/skill_library/skill-index.json",
         "src/inferrsi/data/skill_library/workflows/claim-contract/SKILL.md"
@@ -31,7 +31,7 @@
   },
   "checkout_ref": "checkout:claim-contract",
   "source_files": {
-    "src/inferrsi/data/skill_library/skill-index.json": "8ce35e18f7629260d67a95617b9918eecff73ed8d9370dd3d4595c725ada9aa1",
+    "src/inferrsi/data/skill_library/skill-index.json": "ea80d57e29a2949cffceefcc1715db375937499144864b9df41c3e7fcae61c16",
     "src/inferrsi/data/skill_library/workflows/claim-contract/SKILL.md": "3af1dbbe81d91cb05fc2ef15a2ae25d7592a08558e1a5535f49c52f33a488c55"
   },
   "skill_metadata": {
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/comfy-ui.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/comfy-ui.json
index b883efaa4..12e3909d6 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/comfy-ui.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/comfy-ui.json
@@ -22,8 +22,8 @@
   "sources": [
     {
       "type": "git",
-      "url": "https://github.com/InferMatrixLab/InferRSI/blob/f321699b42a490276ffa4ea52350591683b575cc/src/inferrsi/data/skill_library/repositories/comfy-ui/SKILL.md",
-      "commit": "f321699b42a490276ffa4ea52350591683b575cc",
+      "url": "https://github.com/InferMatrixLab/InferRSI/blob/3291ffbeab2fe72b217d21c8cac282a4f5348267/src/inferrsi/data/skill_library/repositories/comfy-ui/SKILL.md",
+      "commit": "3291ffbeab2fe72b217d21c8cac282a4f5348267",
       "paths": [
         "src/inferrsi/data/skill_library/repositories/comfy-ui/LICENSE",
         "src/inferrsi/data/skill_library/repositories/comfy-ui/SKILL.md",
@@ -50,7 +50,7 @@
     "src/inferrsi/data/skill_library/repositories/comfy-ui/references/model-residency.md": "014c68386e7cdfdb7b4f566836e8b2ff61bd5a5a506e7b4723f83fe80c5d4de7",
     "src/inferrsi/data/skill_library/repositories/comfy-ui/references/repo-provenance.md": "ca996ef515af42a3ce304747fa3f12a31c00533deaace5ba2ce21aefa0a62e2e",
     "src/inferrsi/data/skill_library/repositories/comfy-ui/references/validation-boundaries.md": "1913dedbeb558d872d6a0e015cc11d822910c3cb259a236db29fb60ad8a64263",
-    "src/inferrsi/data/skill_library/skill-index.json": "8ce35e18f7629260d67a95617b9918eecff73ed8d9370dd3d4595c725ada9aa1"
+    "src/inferrsi/data/skill_library/skill-index.json": "ea80d57e29a2949cffceefcc1715db375937499144864b9df41c3e7fcae61c16"
   },
   "skill_metadata": {
     "name": "comfy-ui",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/competitor-scan.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/competitor-scan.json
index 20051ece7..e646112f0 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/competitor-scan.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/competitor-scan.json
@@ -16,8 +16,8 @@
   "sources": [
     {
       "type": "git",
-      "url": "https://github.com/InferMatrixLab/InferRSI/blob/f321699b42a490276ffa4ea52350591683b575cc/src/inferrsi/data/skill_library/workflows/competitor-scan/SKILL.md",
-      "commit": "f321699b42a490276ffa4ea52350591683b575cc",
+      "url": "https://github.com/InferMatrixLab/InferRSI/blob/3291ffbeab2fe72b217d21c8cac282a4f5348267/src/inferrsi/data/skill_library/workflows/competitor-scan/SKILL.md",
+      "commit": "3291ffbeab2fe72b217d21c8cac282a4f5348267",
       "paths": [
         "src/inferrsi/data/skill_library/skill-index.json",
         "src/inferrsi/data/skill_library/workflows/competitor-scan/SKILL.md"
@@ -31,7 +31,7 @@
   },
   "checkout_ref": "checkout:competitor-scan",
   "source_files": {
-    "src/inferrsi/data/skill_library/skill-index.json": "8ce35e18f7629260d67a95617b9918eecff73ed8d9370dd3d4595c725ada9aa1",
+    "src/inferrsi/data/skill_library/skill-index.json": "ea80d57e29a2949cffceefcc1715db375937499144864b9df41c3e7fcae61c16",
     "src/inferrsi/data/skill_library/workflows/competitor-scan/SKILL.md": "0ede58ace65a9324a538b4816add1684f230ba536d4364fe5a5dd2c0bbed3776"
   },
   "skill_metadata": {
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/dcp-async-save-performance.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/dcp-async-save-performance.json
index 8d9f0dfcf..7010f588b 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/dcp-async-save-performance.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/dcp-async-save-performance.json
@@ -23,8 +23,8 @@
   "sources": [
     {
       "type": "git",
-      "url": "https://github.com/InferMatrixLab/InferRSI/blob/f321699b42a490276ffa4ea52350591683b575cc/src/inferrsi/data/skill_library/frameworks/pytorch/dcp-async-save-performance/SKILL.md",
-      "commit": "f321699b42a490276ffa4ea52350591683b575cc",
+      "url": "https://github.com/InferMatrixLab/InferRSI/blob/3291ffbeab2fe72b217d21c8cac282a4f5348267/src/inferrsi/data/skill_library/frameworks/pytorch/dcp-async-save-performance/SKILL.md",
+      "commit": "3291ffbeab2fe72b217d21c8cac282a4f5348267",
       "paths": [
         "src/inferrsi/data/skill_library/frameworks/pytorch/dcp-async-save-performance/SKILL.md",
         "src/inferrsi/data/skill_library/frameworks/pytorch/dcp-async-save-performance/metadata.json",
@@ -43,7 +43,7 @@
     "src/inferrsi/data/skill_library/frameworks/pytorch/dcp-async-save-performance/SKILL.md": "645139ee0c88ec865cad952e24cda35f4648d4b777b732c9b7d274b6c7830845",
     "src/inferrsi/data/skill_library/frameworks/pytorch/dcp-async-save-performance/metadata.json": "db7ad368394ccbfdeff6a16e3423c300d51305b1f95ff5c3936df91738a8cc35",
     "src/inferrsi/data/skill_library/frameworks/pytorch/dcp-async-save-performance/references/sources.md": "22afe4f4d77f441096cc657d55d763473bfcbea8fb531c8e493fe49ab84955f7",
-    "src/inferrsi/data/skill_library/skill-index.json": "8ce35e18f7629260d67a95617b9918eecff73ed8d9370dd3d4595c725ada9aa1"
+    "src/inferrsi/data/skill_library/skill-index.json": "ea80d57e29a2949cffceefcc1715db375937499144864b9df41c3e7fcae61c16"
   },
   "skill_metadata": {
     "name": "dcp-async-save-performance",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/dense-performance-evidence.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/dense-performance-evidence.json
index 92fa7f4f8..8cf0a25fc 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/dense-performance-evidence.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/dense-performance-evidence.json
@@ -21,8 +21,8 @@
   "sources": [
     {
       "type": "git",
-      "url": "https://github.com/InferMatrixLab/InferRSI/blob/f321699b42a490276ffa4ea52350591683b575cc/src/inferrsi/data/skill_library/methods/dense-performance-evidence/SKILL.md",
-      "commit": "f321699b42a490276ffa4ea52350591683b575cc",
+      "url": "https://github.com/InferMatrixLab/InferRSI/blob/3291ffbeab2fe72b217d21c8cac282a4f5348267/src/inferrsi/data/skill_library/methods/dense-performance-evidence/SKILL.md",
+      "commit": "3291ffbeab2fe72b217d21c8cac282a4f5348267",
       "paths": [
         "src/inferrsi/data/skill_library/methods/dense-performance-evidence/SKILL.md",
         "src/inferrsi/data/skill_library/methods/dense-performance-evidence/metadata.json",
@@ -39,7 +39,7 @@
   "source_files": {
     "src/inferrsi/data/skill_library/methods/dense-performance-evidence/SKILL.md": "53cc1867a13f54d1006d50517dea8848869ca50df3f0cd610b2424cc8d18c933",
     "src/inferrsi/data/skill_library/methods/dense-performance-evidence/metadata.json": "a0d1e3b17140295510a764045d1b222e19be3a3406e6d1e7dddaad49ff474490",
-    "src/inferrsi/data/skill_library/skill-index.json": "8ce35e18f7629260d67a95617b9918eecff73ed8d9370dd3d4595c725ada9aa1"
+    "src/inferrsi/data/skill_library/skill-index.json": "ea80d57e29a2949cffceefcc1715db375937499144864b9df41c3e7fcae61c16"
   },
   "skill_metadata": {
     "name": "dense-performance-evidence",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/diagnose-vllm-admission-pressure.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/diagnose-vllm-admission-pressure.json
index c1297aca9..15adc5852 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/diagnose-vllm-admission-pressure.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/diagnose-vllm-admission-pressure.json
@@ -21,8 +21,8 @@
   "sources": [
     {
       "type": "git",
-      "url": "https://github.com/InferMatrixLab/InferRSI/blob/f321699b42a490276ffa4ea52350591683b575cc/src/inferrsi/data/skill_library/repositories/vllm/diagnose-vllm-admission-pressure/SKILL.md",
-      "commit": "f321699b42a490276ffa4ea52350591683b575cc",
+      "url": "https://github.com/InferMatrixLab/InferRSI/blob/3291ffbeab2fe72b217d21c8cac282a4f5348267/src/inferrsi/data/skill_library/repositories/vllm/diagnose-vllm-admission-pressure/SKILL.md",
+      "commit": "3291ffbeab2fe72b217d21c8cac282a4f5348267",
       "paths": [
         "src/inferrsi/data/skill_library/repositories/vllm/diagnose-vllm-admission-pressure/SKILL.md",
         "src/inferrsi/data/skill_library/repositories/vllm/diagnose-vllm-admission-pressure/metadata.json",
@@ -67,7 +67,7 @@
     "src/inferrsi/data/skill_library/repositories/vllm/diagnose-vllm-admission-pressure/references/papers-20261015/references/loongserve-readset.md": "f1a383e7bdbbd5a5a472d7f6e1292c8a17d4910ff736ca5a6aa422e1bee74f8c",
     "src/inferrsi/data/skill_library/repositories/vllm/diagnose-vllm-admission-pressure/references/papers-20261015/references/orca-readset.md": "a497ef23b7e60c0fbc15d95b7e07cd20547cfe56f37c24dfe6ed70d79ef7a3cf",
     "src/inferrsi/data/skill_library/repositories/vllm/diagnose-vllm-admission-pressure/references/source-notes.md": "84033fd514773f10c0bddd9914fabb980497ba2b308f226a922c3e1888413827",
-    "src/inferrsi/data/skill_library/skill-index.json": "8ce35e18f7629260d67a95617b9918eecff73ed8d9370dd3d4595c725ada9aa1"
+    "src/inferrsi/data/skill_library/skill-index.json": "ea80d57e29a2949cffceefcc1715db375937499144864b9df41c3e7fcae61c16"
   },
   "skill_metadata": {
     "name": "diagnose-vllm-admission-pressure",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/diffusers-invocation-state.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/diffusers-invocation-state.json
index 3150a8890..de5a6ebb3 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/diffusers-invocation-state.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/diffusers-invocation-state.json
@@ -22,8 +22,8 @@
   "sources": [
     {
       "type": "git",
-      "url": "https://github.com/InferMatrixLab/InferRSI/blob/f321699b42a490276ffa4ea52350591683b575cc/src/inferrsi/data/skill_library/frameworks/diffusers/diffusers-invocation-state/SKILL.md",
-      "commit": "f321699b42a490276ffa4ea52350591683b575cc",
+      "url": "https://github.com/InferMatrixLab/InferRSI/blob/3291ffbeab2fe72b217d21c8cac282a4f5348267/src/inferrsi/data/skill_library/frameworks/diffusers/diffusers-invocation-state/SKILL.md",
+      "commit": "3291ffbeab2fe72b217d21c8cac282a4f5348267",
       "paths": [
         "src/inferrsi/data/skill_library/frameworks/diffusers/diffusers-invocation-state/SKILL.md",
         "src/inferrsi/data/skill_library/frameworks/diffusers/diffusers-invocation-state/metadata.json",
@@ -46,7 +46,7 @@
     "src/inferrsi/data/skill_library/frameworks/diffusers/diffusers-invocation-state/references/ownership.md": "40688cc44183027efbf2fca8d3548c299cdf2e33fef54f76ab3549353eb9bd1b",
     "src/inferrsi/data/skill_library/frameworks/diffusers/diffusers-invocation-state/references/source-map.md": "10562a75683bd9bd07ec2e37a1cbd0f93f2534d6f85b55a566adb2d97485215f",
     "src/inferrsi/data/skill_library/frameworks/diffusers/diffusers-invocation-state/references/validation.md": "5a0b854b87d091d7d505b0bdd5c03330052507a851ab6edf35b93a6675f2d4f7",
-    "src/inferrsi/data/skill_library/skill-index.json": "8ce35e18f7629260d67a95617b9918eecff73ed8d9370dd3d4595c725ada9aa1"
+    "src/inferrsi/data/skill_library/skill-index.json": "ea80d57e29a2949cffceefcc1715db375937499144864b9df41c3e7fcae61c16"
   },
   "skill_metadata": {
     "name": "diffusers-invocation-state",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/draft-acceptance-issue.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/draft-acceptance-issue.json
index cbcd7bd93..372fcb02f 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/draft-acceptance-issue.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/draft-acceptance-issue.json
@@ -16,8 +16,8 @@
   "sources": [
     {
       "type": "git",
-      "url": "https://github.com/InferMatrixLab/InferRSI/blob/f321699b42a490276ffa4ea52350591683b575cc/src/inferrsi/data/skill_library/workflows/draft-acceptance-issue/SKILL.md",
-      "commit": "f321699b42a490276ffa4ea52350591683b575cc",
+      "url": "https://github.com/InferMatrixLab/InferRSI/blob/3291ffbeab2fe72b217d21c8cac282a4f5348267/src/inferrsi/data/skill_library/workflows/draft-acceptance-issue/SKILL.md",
+      "commit": "3291ffbeab2fe72b217d21c8cac282a4f5348267",
       "paths": [
         "src/inferrsi/data/skill_library/skill-index.json",
         "src/inferrsi/data/skill_library/workflows/draft-acceptance-issue/SKILL.md"
@@ -31,7 +31,7 @@
   },
   "checkout_ref": "checkout:draft-acceptance-issue",
   "source_files": {
-    "src/inferrsi/data/skill_library/skill-index.json": "8ce35e18f7629260d67a95617b9918eecff73ed8d9370dd3d4595c725ada9aa1",
+    "src/inferrsi/data/skill_library/skill-index.json": "ea80d57e29a2949cffceefcc1715db375937499144864b9df41c3e7fcae61c16",
     "src/inferrsi/data/skill_library/workflows/draft-acceptance-issue/SKILL.md": "ce983acfb7d7d3272775e7cf00c92e4ddadd97511b602575a76e5ab7ade01b56"
   },
   "skill_metadata": {
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/dwarfstar.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/dwarfstar.json
new file mode 100644
index 000000000..d8e8d0648
--- /dev/null
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/dwarfstar.json
@@ -0,0 +1,192 @@
+{
+  "schema_version": 1,
+  "id": "inferrsi.common.skills.dwarfstar",
+  "title": "dwarfstar",
+  "repo": "inferrsi",
+  "tier": "contribution",
+  "backend": "common",
+  "component": "skills",
+  "scenario": "skills.discovery",
+  "status": "imported",
+  "summary": "Investigate DwarfStar model fit, expert SSD streaming, session and vision cache correctness, speculative sampling and distributed execution using pinned source contracts and explicit native verification boundaries.",
+  "tags": [
+    "dwarfstar",
+    "ds4",
+    "moe",
+    "ssd-streaming",
+    "session-state",
+    "kv-cache",
+    "speculation",
+    "tensor-parallelism"
+  ],
+  "origin": [
+    "external"
+  ],
+  "sources": [
+    {
+      "type": "git",
+      "url": "https://github.com/InferMatrixLab/InferRSI/blob/3291ffbeab2fe72b217d21c8cac282a4f5348267/src/inferrsi/data/skill_library/repositories/dwarfstar/SKILL.md",
+      "commit": "3291ffbeab2fe72b217d21c8cac282a4f5348267",
+      "paths": [
+        "src/inferrsi/data/skill_library/repositories/dwarfstar/LICENSE",
+        "src/inferrsi/data/skill_library/repositories/dwarfstar/SKILL.md",
+        "src/inferrsi/data/skill_library/repositories/dwarfstar/metadata.json",
+        "src/inferrsi/data/skill_library/repositories/dwarfstar/references/repo-provenance.md",
+        "src/inferrsi/data/skill_library/repositories/dwarfstar/references/repo-routing-metadata.json",
+        "src/inferrsi/data/skill_library/repositories/dwarfstar/references/source-identity.json",
+        "src/inferrsi/data/skill_library/repositories/dwarfstar/scripts/request_template.py",
+        "src/inferrsi/data/skill_library/repositories/dwarfstar/sub-skills/qualification/SKILL.md",
+        "src/inferrsi/data/skill_library/repositories/dwarfstar/sub-skills/qualification/references/troubleshooting.md",
+        "src/inferrsi/data/skill_library/repositories/dwarfstar/sub-skills/qualification/references/verification-contract.md",
+        "src/inferrsi/data/skill_library/repositories/dwarfstar/sub-skills/runtime-memory/SKILL.md",
+        "src/inferrsi/data/skill_library/repositories/dwarfstar/sub-skills/runtime-memory/references/memory-ledger.md",
+        "src/inferrsi/data/skill_library/repositories/dwarfstar/sub-skills/runtime-memory/references/troubleshooting.md",
+        "src/inferrsi/data/skill_library/repositories/dwarfstar/sub-skills/serving-state/SKILL.md",
+        "src/inferrsi/data/skill_library/repositories/dwarfstar/sub-skills/serving-state/references/state-contracts.md",
+        "src/inferrsi/data/skill_library/repositories/dwarfstar/sub-skills/serving-state/references/troubleshooting.md",
+        "src/inferrsi/data/skill_library/skill-index.json"
+      ],
+      "content_sha256": "a04a252ecdcd4710a3273595f4bd1f3eb8e24d5bb1bfe005f058a78388254dc9"
+    }
+  ],
+  "index_type": "packaged-skill",
+  "skill_documents": {
+    "original": "data/skill_library/repositories/dwarfstar/SKILL.md"
+  },
+  "checkout_ref": "checkout:dwarfstar",
+  "source_files": {
+    "src/inferrsi/data/skill_library/repositories/dwarfstar/LICENSE": "4a8a1358f5b048627d71e4988de0bafd386f4988991929629e6cb9e29b7b9b2f",
+    "src/inferrsi/data/skill_library/repositories/dwarfstar/SKILL.md": "a04a252ecdcd4710a3273595f4bd1f3eb8e24d5bb1bfe005f058a78388254dc9",
+    "src/inferrsi/data/skill_library/repositories/dwarfstar/metadata.json": "e1f0c2dc3d7ea9278039c2e8c3e8ca4f8e3aba33504b7784d628e903f0b5fc8c",
+    "src/inferrsi/data/skill_library/repositories/dwarfstar/references/repo-provenance.md": "7d72f9c85ee60fe003f5f5a585b52e9989ebcedca625b584fa39a976d4433998",
+    "src/inferrsi/data/skill_library/repositories/dwarfstar/references/repo-routing-metadata.json": "8e195b420b67151932af5f435c284f49677cc7863e650c83cfa9e762fd00db0b",
+    "src/inferrsi/data/skill_library/repositories/dwarfstar/references/source-identity.json": "a38fe07d003accd83b7626d7e189f7b49b4b73e7a223e4d441b0e33add5d9c91",
+    "src/inferrsi/data/skill_library/repositories/dwarfstar/scripts/request_template.py": "62e4cde002f2aebeaddd63e6e1f04fc727c271d2cbdaa00856158baadee5340b",
+    "src/inferrsi/data/skill_library/repositories/dwarfstar/sub-skills/qualification/SKILL.md": "3b8eb25f6b8e4729b48cce5619daa24422bee251f119ea7f9cc1faf0097ec6bb",
+    "src/inferrsi/data/skill_library/repositories/dwarfstar/sub-skills/qualification/references/troubleshooting.md": "cd832ab2a420490524251a042c0c83c2f75bbfdc12e914b91da8f91edf6860e9",
+    "src/inferrsi/data/skill_library/repositories/dwarfstar/sub-skills/qualification/references/verification-contract.md": "2801f6683c2974d1cfa5b86c6268fd33f175f1cbcc9ed85c932e9b3504fac3e1",
+    "src/inferrsi/data/skill_library/repositories/dwarfstar/sub-skills/runtime-memory/SKILL.md": "ad1d34d429baaa41fb72c94ff4e554b26b2f5b701cc4deed03f1d066c7d3fc52",
+    "src/inferrsi/data/skill_library/repositories/dwarfstar/sub-skills/runtime-memory/references/memory-ledger.md": "765915ecd3a318cc7908658d5a79bbdff615fc24409501febcb7ff8af65a1050",
+    "src/inferrsi/data/skill_library/repositories/dwarfstar/sub-skills/runtime-memory/references/troubleshooting.md": "42e4785b7d3c1375efa324583897327f28ef76c0b7788092da92095bddeb1440",
+    "src/inferrsi/data/skill_library/repositories/dwarfstar/sub-skills/serving-state/SKILL.md": "71c27212847146376d72e6cc6d22abe2f7eed56402c5917cd7dfcda0b2aa5982",
+    "src/inferrsi/data/skill_library/repositories/dwarfstar/sub-skills/serving-state/references/state-contracts.md": "8f74a07c947021e3069468268f6b5aa2548a17db518c5c0505b84dde73a1f31c",
+    "src/inferrsi/data/skill_library/repositories/dwarfstar/sub-skills/serving-state/references/troubleshooting.md": "c9eebe4109d418c4dfb9bb2425c145ebc1a52da9b665e0556149bc31a442e2fc",
+    "src/inferrsi/data/skill_library/skill-index.json": "ea80d57e29a2949cffceefcc1715db375937499144864b9df41c3e7fcae61c16"
+  },
+  "skill_metadata": {
+    "name": "dwarfstar",
+    "description": "Investigate DwarfStar model fit, expert SSD streaming, session and vision cache correctness, speculative sampling and distributed execution using pinned source contracts and explicit native verification boundaries.",
+    "license": "NO_LICENSE",
+    "disable-model-invocation": true,
+    "metadata": {
+      "disco-role": "operating"
+    },
+    "schema_version": 1,
+    "kind": "skill",
+    "origin": [
+      "external"
+    ],
+    "status": "candidate",
+    "empirical_validation": "partial_cpu_only",
+    "tags": [
+      "dwarfstar",
+      "ds4",
+      "moe",
+      "ssd-streaming",
+      "session-state",
+      "kv-cache",
+      "speculation",
+      "tensor-parallelism"
+    ],
+    "sources": [
+      {
+        "type": "git",
+        "url": "https://github.com/antirez/ds4",
+        "commit": "ee7e8405232c8971000cc7ba2b945b00bf6d20d3",
+        "paths": [
+          "README.md",
+          "LICENSE",
+          "Makefile",
+          "ds4.h",
+          "ds4.c",
+          "ds4_ssd.c",
+          "ds4_ssd.h",
+          "ds4_kvstore.c",
+          "ds4_kvstore.h",
+          "ds4_server.c",
+          "ds4_tp.c",
+          "ds4_tp.h",
+          "docs/SSD_STREAMING.md",
+          "docs/SERVER.md",
+          "docs/SPECULATIVE_DECODING.md",
+          "docs/DISTRIBUTED.md",
+          "docs/CUDA_MULTI_GPU.md",
+          "docs/TESTING.md",
+          "tests/test_session_state.c",
+          "tests/test_sampling.c",
+          "tests/test_tp_commands.c",
+          "tests/test_engine_mgpu_placement.c",
+          "ds4_cuda.cu",
+          "QA_BEFORE_RELEASES.md"
+        ],
+        "readset": "references/source-identity.json",
+        "license": "MIT"
+      },
+      {
+        "type": "git",
+        "url": "https://github.com/VectorSpaceLab/AREX-Skill",
+        "commit": "ac3fe1afa80fb9a09775ecfb2b6cc3ba850a2db6",
+        "paths": [
+          "cli/packages/coding-agent/src/disco/skills/create-repo-skill",
+          "cli/packages/coding-agent/src/disco/skills/prepare-repo-skill-env",
+          "cli/packages/coding-agent/src/disco/skills/workflow-authoring",
+          "cli/packages/coding-agent/src/disco/skills/verify-repo-skill"
+        ],
+        "scope": "Portable instructions executed by compatible agent with native C adaptation; no DisCo CLI/model run or managed import."
+      }
+    ],
+    "construction": {
+      "anchor_kind": "source",
+      "distillation_form": "repository-operating-graph",
+      "strategy": "arex-portable-meta-skills",
+      "status": "partial-candidate",
+      "native_import": "not-run",
+      "method": "User-approved native C adaptation and partial drafting; manual stable-ID subskill orchestration and integrated create/verify stages. Required-backend blocks disable automatic admission."
+    },
+    "applicability": {
+      "scope": "Pinned DwarfStar source investigation and experiment planning; not arbitrary GGUF support or cross-backend qualification."
+    },
+    "limitations": [
+      "Metal/CUDA/ROCm inference, actual model output, SSD runtime I/O, GPU recurrent rollback, physical distributed behavior, service quality/performance and skill efficacy remain unverified.",
+      "CPU build/help and selected bookkeeping tests do not clear BLOCKED_REQUIRED_BACKEND.",
+      "NO_LICENSE is the actual unauthenticated GitHub CLI resolver fallback, not a legal conclusion; pinned source MIT notice is retained.",
+      "Only selected source bodies were inspected; full kernels, conversion/imatrix tools and every workflow are not covered."
+    ],
+    "provenance_readsets": [
+      "references/repo-provenance.md",
+      "references/source-identity.json"
+    ],
+    "consumption_mode": "packaged"
+  },
+  "applicability": {
+    "scope": "Pinned DwarfStar source investigation and experiment planning; not arbitrary GGUF support or cross-backend qualification."
+  },
+  "recommendations": [
+    "Read the pinned SKILL.md and its prerequisites before choosing this procedure.",
+    "Read the packaged references and owner-local cases alongside the procedure."
+  ],
+  "limitations": [
+    "This generated abstract points to the complete packaged skill; external evidence and upstream repositories remain source links.",
+    "The imported status describes this index, not the original skill; its declared status is candidate.",
+    "Reading or indexing does not authorize execution, validate efficacy, promote a candidate or grant a license.",
+    "The backend categorizes the procedure's hardware dependency; it does not certify compatibility, and skills.discovery is a discovery scenario.",
+    "Metal/CUDA/ROCm inference, actual model output, SSD runtime I/O, GPU recurrent rollback, physical distributed behavior, service quality/performance and skill efficacy remain unverified.",
+    "CPU build/help and selected bookkeeping tests do not clear BLOCKED_REQUIRED_BACKEND.",
+    "NO_LICENSE is the actual unauthenticated GitHub CLI resolver fallback, not a legal conclusion; pinned source MIT notice is retained.",
+    "Only selected source bodies were inspected; full kernels, conversion/imatrix tools and every workflow are not covered."
+  ],
+  "verification": {
+    "locally_verified": false,
+    "evidence_level": "guidance"
+  }
+}
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/evaluate-elastic-serving-placement.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/evaluate-elastic-serving-placement.json
index c7f364017..daac86b18 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/evaluate-elastic-serving-placement.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/evaluate-elastic-serving-placement.json
@@ -16,8 +16,8 @@
   "sources": [
     {
       "type": "git",
-      "url": "https://github.com/InferMatrixLab/InferRSI/blob/f321699b42a490276ffa4ea52350591683b575cc/src/inferrsi/data/skill_library/methods/evaluate-elastic-serving-placement/SKILL.md",
-      "commit": "f321699b42a490276ffa4ea52350591683b575cc",
+      "url": "https://github.com/InferMatrixLab/InferRSI/blob/3291ffbeab2fe72b217d21c8cac282a4f5348267/src/inferrsi/data/skill_library/methods/evaluate-elastic-serving-placement/SKILL.md",
+      "commit": "3291ffbeab2fe72b217d21c8cac282a4f5348267",
       "paths": [
         "src/inferrsi/data/skill_library/methods/evaluate-elastic-serving-placement/SKILL.md",
         "src/inferrsi/data/skill_library/methods/evaluate-elastic-serving-placement/metadata.json",
@@ -36,7 +36,7 @@
     "src/inferrsi/data/skill_library/methods/evaluate-elastic-serving-placement/SKILL.md": "82d0099d3c33706f46c657d4fa7f76d3d486e019b36760d1de001921328a60cc",
     "src/inferrsi/data/skill_library/methods/evaluate-elastic-serving-placement/metadata.json": "37379a4d4a058ca46da2b20a920dc01be9ec11e557c32b2332702ec76027f4b5",
     "src/inferrsi/data/skill_library/methods/evaluate-elastic-serving-placement/references/source-notes.md": "ea416831fd7d0874e06d0a76fb0c7196df53072b717673a0cdf635db6bfd917b",
-    "src/inferrsi/data/skill_library/skill-index.json": "8ce35e18f7629260d67a95617b9918eecff73ed8d9370dd3d4595c725ada9aa1"
+    "src/inferrsi/data/skill_library/skill-index.json": "ea80d57e29a2949cffceefcc1715db375937499144864b9df41c3e7fcae61c16"
   },
   "skill_metadata": {
     "name": "evaluate-elastic-serving-placement",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/generation-cache-lifecycle.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/generation-cache-lifecycle.json
index eb11dbe8a..6b9a4f657 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/generation-cache-lifecycle.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/generation-cache-lifecycle.json
@@ -22,8 +22,8 @@
   "sources": [
     {
       "type": "git",
-      "url": "https://github.com/InferMatrixLab/InferRSI/blob/f321699b42a490276ffa4ea52350591683b575cc/src/inferrsi/data/skill_library/repositories/mlx-lm/generation-cache-lifecycle/SKILL.md",
-      "commit": "f321699b42a490276ffa4ea52350591683b575cc",
+      "url": "https://github.com/InferMatrixLab/InferRSI/blob/3291ffbeab2fe72b217d21c8cac282a4f5348267/src/inferrsi/data/skill_library/repositories/mlx-lm/generation-cache-lifecycle/SKILL.md",
+      "commit": "3291ffbeab2fe72b217d21c8cac282a4f5348267",
       "paths": [
         "src/inferrsi/data/skill_library/repositories/mlx-lm/generation-cache-lifecycle/SKILL.md",
         "src/inferrsi/data/skill_library/repositories/mlx-lm/generation-cache-lifecycle/metadata.json",
@@ -48,7 +48,7 @@
     "src/inferrsi/data/skill_library/repositories/mlx-lm/generation-cache-lifecycle/references/generation-batch.md": "f16d3c57d45e63bcaffd3871aa41c76972ee11f7dbf1e12b45d9b170303cad9b",
     "src/inferrsi/data/skill_library/repositories/mlx-lm/generation-cache-lifecycle/references/source-map.md": "91d26181f246307f2113f94b12ac9587d689b654dcd216e9f14b11cf5ebb650c",
     "src/inferrsi/data/skill_library/repositories/mlx-lm/generation-cache-lifecycle/references/validation.md": "73b5dd347017baba519ecc85415aead9993c1751d116b9ba4f07689f58c5f783",
-    "src/inferrsi/data/skill_library/skill-index.json": "8ce35e18f7629260d67a95617b9918eecff73ed8d9370dd3d4595c725ada9aa1"
+    "src/inferrsi/data/skill_library/skill-index.json": "ea80d57e29a2949cffceefcc1715db375937499144864b9df41c3e7fcae61c16"
   },
   "skill_metadata": {
     "name": "generation-cache-lifecycle",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/host-memory-management.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/host-memory-management.json
index b9c9c8fe6..ae60bb5f6 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/host-memory-management.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/host-memory-management.json
@@ -27,8 +27,8 @@
   "sources": [
     {
       "type": "git",
-      "url": "https://github.com/InferMatrixLab/InferRSI/blob/f321699b42a490276ffa4ea52350591683b575cc/src/inferrsi/data/skill_library/hardware/host/host-memory-management/SKILL.md",
-      "commit": "f321699b42a490276ffa4ea52350591683b575cc",
+      "url": "https://github.com/InferMatrixLab/InferRSI/blob/3291ffbeab2fe72b217d21c8cac282a4f5348267/src/inferrsi/data/skill_library/hardware/host/host-memory-management/SKILL.md",
+      "commit": "3291ffbeab2fe72b217d21c8cac282a4f5348267",
       "paths": [
         "src/inferrsi/data/skill_library/hardware/host/host-memory-management/SKILL.md",
         "src/inferrsi/data/skill_library/hardware/host/host-memory-management/metadata.json",
@@ -51,7 +51,7 @@
     "src/inferrsi/data/skill_library/hardware/host/host-memory-management/references/vllm-ascend-LICENSE.md": "c71d239df91726fc519c6eb72d318ec65820627232b2f796219e87dcf35d0ab4",
     "src/inferrsi/data/skill_library/hardware/host/host-memory-management/references/vllm-ascend-native-offload-readset.json": "cc6c2011be746002ba3ff96ccefab1809863748ddd17eb102b75e7b0e134fe55",
     "src/inferrsi/data/skill_library/hardware/host/host-memory-management/references/vllm-ascend-native-offload.md": "6c2a72b5ba3708a01d229e0576570837a617e6a42dbb0b5312a26a65a6c31542",
-    "src/inferrsi/data/skill_library/skill-index.json": "8ce35e18f7629260d67a95617b9918eecff73ed8d9370dd3d4595c725ada9aa1"
+    "src/inferrsi/data/skill_library/skill-index.json": "ea80d57e29a2949cffceefcc1715db375937499144864b9df41c3e7fcae61c16"
   },
   "skill_metadata": {
     "name": "host-memory-management",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/host-overhead-optimization.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/host-overhead-optimization.json
index 413037ff2..b9ab5f03c 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/host-overhead-optimization.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/host-overhead-optimization.json
@@ -22,8 +22,8 @@
   "sources": [
     {
       "type": "git",
-      "url": "https://github.com/InferMatrixLab/InferRSI/blob/f321699b42a490276ffa4ea52350591683b575cc/src/inferrsi/data/skill_library/hardware/host/host-overhead-optimization/SKILL.md",
-      "commit": "f321699b42a490276ffa4ea52350591683b575cc",
+      "url": "https://github.com/InferMatrixLab/InferRSI/blob/3291ffbeab2fe72b217d21c8cac282a4f5348267/src/inferrsi/data/skill_library/hardware/host/host-overhead-optimization/SKILL.md",
+      "commit": "3291ffbeab2fe72b217d21c8cac282a4f5348267",
       "paths": [
         "src/inferrsi/data/skill_library/hardware/host/host-overhead-optimization/SKILL.md",
         "src/inferrsi/data/skill_library/hardware/host/host-overhead-optimization/metadata.json",
@@ -42,7 +42,7 @@
     "src/inferrsi/data/skill_library/hardware/host/host-overhead-optimization/SKILL.md": "36ef88826a32ac3147b869001004ded17a8d5a491e25bfe105005ba39ea9d1eb",
     "src/inferrsi/data/skill_library/hardware/host/host-overhead-optimization/metadata.json": "ae71c389dfdb1a19d267feecb4c6f1fbb4c9afad501ff6d8fab7e38b29bcc37a",
     "src/inferrsi/data/skill_library/hardware/host/host-overhead-optimization/references/deepinfer-zero-overhead-scheduling.md": "f452bf40132c9195d739f1c86b3aa18857138c7ed0f1861af5853d93b9955924",
-    "src/inferrsi/data/skill_library/skill-index.json": "8ce35e18f7629260d67a95617b9918eecff73ed8d9370dd3d4595c725ada9aa1"
+    "src/inferrsi/data/skill_library/skill-index.json": "ea80d57e29a2949cffceefcc1715db375937499144864b9df41c3e7fcae61c16"
   },
   "skill_metadata": {
     "name": "host-overhead-optimization",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/inference-benchmark-design.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/inference-benchmark-design.json
index 7ab9fa931..b9cf44a0c 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/inference-benchmark-design.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/inference-benchmark-design.json
@@ -30,8 +30,8 @@
   "sources": [
     {
       "type": "git",
-      "url": "https://github.com/InferMatrixLab/InferRSI/blob/f321699b42a490276ffa4ea52350591683b575cc/src/inferrsi/data/skill_library/methods/inference-benchmark-design/SKILL.md",
-      "commit": "f321699b42a490276ffa4ea52350591683b575cc",
+      "url": "https://github.com/InferMatrixLab/InferRSI/blob/3291ffbeab2fe72b217d21c8cac282a4f5348267/src/inferrsi/data/skill_library/methods/inference-benchmark-design/SKILL.md",
+      "commit": "3291ffbeab2fe72b217d21c8cac282a4f5348267",
       "paths": [
         "src/inferrsi/data/skill_library/methods/inference-benchmark-design/SKILL.md",
         "src/inferrsi/data/skill_library/methods/inference-benchmark-design/metadata.json",
@@ -52,7 +52,7 @@
     "src/inferrsi/data/skill_library/methods/inference-benchmark-design/metadata.json": "18310402a8ade03fd48ac687f79ba9d447871e90337de84c0e1bc15c0b827306",
     "src/inferrsi/data/skill_library/methods/inference-benchmark-design/references/benchmark-commands.md": "0898424e0bff3a6246e1eeeda6abd6c5bd974c047fe32f3bbd5ab9d31ce17cd3",
     "src/inferrsi/data/skill_library/methods/inference-benchmark-design/references/task-benchmark-selection.md": "0d4bb080a754e26846e81680023db0dff5e6a4c96be4327c411c3b72ac4ef8e2",
-    "src/inferrsi/data/skill_library/skill-index.json": "8ce35e18f7629260d67a95617b9918eecff73ed8d9370dd3d4595c725ada9aa1"
+    "src/inferrsi/data/skill_library/skill-index.json": "ea80d57e29a2949cffceefcc1715db375937499144864b9df41c3e7fcae61c16"
   },
   "skill_metadata": {
     "name": "inference-benchmark-design",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/inference-optimization-casebook.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/inference-optimization-casebook.json
index e070e2f0e..817284014 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/inference-optimization-casebook.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/inference-optimization-casebook.json
@@ -21,8 +21,8 @@
   "sources": [
     {
       "type": "git",
-      "url": "https://github.com/InferMatrixLab/InferRSI/blob/f321699b42a490276ffa4ea52350591683b575cc/src/inferrsi/data/skill_library/methods/inference-optimization-casebook/SKILL.md",
-      "commit": "f321699b42a490276ffa4ea52350591683b575cc",
+      "url": "https://github.com/InferMatrixLab/InferRSI/blob/3291ffbeab2fe72b217d21c8cac282a4f5348267/src/inferrsi/data/skill_library/methods/inference-optimization-casebook/SKILL.md",
+      "commit": "3291ffbeab2fe72b217d21c8cac282a4f5348267",
       "paths": [
         "src/inferrsi/data/skill_library/methods/inference-optimization-casebook/SKILL.md",
         "src/inferrsi/data/skill_library/methods/inference-optimization-casebook/cases/admission-drain-occupancy.json",
@@ -55,7 +55,7 @@
     "src/inferrsi/data/skill_library/methods/inference-optimization-casebook/cases/qwen-b300-rejected-directions.json": "146ea1564a46ed03092c7dc5796b318fc05b04b1821105ee00eb7fa7fa1333e2",
     "src/inferrsi/data/skill_library/methods/inference-optimization-casebook/cases/qwen-b300-rejected-directions.md": "cce2930f544a0f78f21797fba4d8ded298e4a81f76c34d7977b0e31d361bc738",
     "src/inferrsi/data/skill_library/methods/inference-optimization-casebook/metadata.json": "fb55067fb45970319787068a3206110641cceaf0ccbabd706be6d0b3d2afc2ed",
-    "src/inferrsi/data/skill_library/skill-index.json": "8ce35e18f7629260d67a95617b9918eecff73ed8d9370dd3d4595c725ada9aa1"
+    "src/inferrsi/data/skill_library/skill-index.json": "ea80d57e29a2949cffceefcc1715db375937499144864b9df41c3e7fcae61c16"
   },
   "skill_metadata": {
     "name": "inference-optimization-casebook",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/kernel-writing-guide.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/kernel-writing-guide.json
index 46f71cc89..7c0f95577 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/kernel-writing-guide.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/kernel-writing-guide.json
@@ -27,8 +27,8 @@
   "sources": [
     {
       "type": "git",
-      "url": "https://github.com/InferMatrixLab/InferRSI/blob/f321699b42a490276ffa4ea52350591683b575cc/src/inferrsi/data/skill_library/methods/kernel-writing-guide/SKILL.md",
-      "commit": "f321699b42a490276ffa4ea52350591683b575cc",
+      "url": "https://github.com/InferMatrixLab/InferRSI/blob/3291ffbeab2fe72b217d21c8cac282a4f5348267/src/inferrsi/data/skill_library/methods/kernel-writing-guide/SKILL.md",
+      "commit": "3291ffbeab2fe72b217d21c8cac282a4f5348267",
       "paths": [
         "src/inferrsi/data/skill_library/methods/kernel-writing-guide/SKILL.md",
         "src/inferrsi/data/skill_library/methods/kernel-writing-guide/metadata.json",
@@ -47,7 +47,7 @@
     "src/inferrsi/data/skill_library/methods/kernel-writing-guide/SKILL.md": "01fb1712d7eb988c5e412f02340d2e48b4f8564fcf4ab232dc5897869d0d50a0",
     "src/inferrsi/data/skill_library/methods/kernel-writing-guide/metadata.json": "0189d184337662056fc6952c7beffc1bf6154c4fa173ae3e3e463bd22d70f36c",
     "src/inferrsi/data/skill_library/methods/kernel-writing-guide/references/akg-affine-tail-guards.md": "113deb2db10b171db3425bb3f1792bee763706b794c8cc2e222b4cc0d0114118",
-    "src/inferrsi/data/skill_library/skill-index.json": "8ce35e18f7629260d67a95617b9918eecff73ed8d9370dd3d4595c725ada9aa1"
+    "src/inferrsi/data/skill_library/skill-index.json": "ea80d57e29a2949cffceefcc1715db375937499144864b9df41c3e7fcae61c16"
   },
   "skill_metadata": {
     "name": "kernel-writing-guide",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/kv-tier-management.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/kv-tier-management.json
index 22dd509c4..bffc4318c 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/kv-tier-management.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/kv-tier-management.json
@@ -24,8 +24,8 @@
   "sources": [
     {
       "type": "git",
-      "url": "https://github.com/InferMatrixLab/InferRSI/blob/f321699b42a490276ffa4ea52350591683b575cc/src/inferrsi/data/skill_library/repositories/lmcache/kv-tier-management/SKILL.md",
-      "commit": "f321699b42a490276ffa4ea52350591683b575cc",
+      "url": "https://github.com/InferMatrixLab/InferRSI/blob/3291ffbeab2fe72b217d21c8cac282a4f5348267/src/inferrsi/data/skill_library/repositories/lmcache/kv-tier-management/SKILL.md",
+      "commit": "3291ffbeab2fe72b217d21c8cac282a4f5348267",
       "paths": [
         "src/inferrsi/data/skill_library/repositories/lmcache/kv-tier-management/SKILL.md",
         "src/inferrsi/data/skill_library/repositories/lmcache/kv-tier-management/metadata.json",
@@ -42,7 +42,7 @@
   "source_files": {
     "src/inferrsi/data/skill_library/repositories/lmcache/kv-tier-management/SKILL.md": "e4b78f368d70c8abb74a935acb910414015f1aeec1b50f7a295e836822ec4b27",
     "src/inferrsi/data/skill_library/repositories/lmcache/kv-tier-management/metadata.json": "c221189b8d85fac3bd41ef1c7cd560f1aebc8cfa0381c182d1d571a7169be829",
-    "src/inferrsi/data/skill_library/skill-index.json": "8ce35e18f7629260d67a95617b9918eecff73ed8d9370dd3d4595c725ada9aa1"
+    "src/inferrsi/data/skill_library/skill-index.json": "ea80d57e29a2949cffceefcc1715db375937499144864b9df41c3e7fcae61c16"
   },
   "skill_metadata": {
     "name": "kv-tier-management",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/lazy-evaluation-stream-lifetimes.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/lazy-evaluation-stream-lifetimes.json
index 96bdafa58..ecb366816 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/lazy-evaluation-stream-lifetimes.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/lazy-evaluation-stream-lifetimes.json
@@ -22,8 +22,8 @@
   "sources": [
     {
       "type": "git",
-      "url": "https://github.com/InferMatrixLab/InferRSI/blob/f321699b42a490276ffa4ea52350591683b575cc/src/inferrsi/data/skill_library/frameworks/mlx/lazy-evaluation-stream-lifetimes/SKILL.md",
-      "commit": "f321699b42a490276ffa4ea52350591683b575cc",
+      "url": "https://github.com/InferMatrixLab/InferRSI/blob/3291ffbeab2fe72b217d21c8cac282a4f5348267/src/inferrsi/data/skill_library/frameworks/mlx/lazy-evaluation-stream-lifetimes/SKILL.md",
+      "commit": "3291ffbeab2fe72b217d21c8cac282a4f5348267",
       "paths": [
         "src/inferrsi/data/skill_library/frameworks/mlx/lazy-evaluation-stream-lifetimes/SKILL.md",
         "src/inferrsi/data/skill_library/frameworks/mlx/lazy-evaluation-stream-lifetimes/metadata.json",
@@ -46,7 +46,7 @@
     "src/inferrsi/data/skill_library/frameworks/mlx/lazy-evaluation-stream-lifetimes/references/execution-lifetime.md": "c4bc41e08aca68f661325f61e6f0db1760a079dc362f295c847bdc374fee3bf6",
     "src/inferrsi/data/skill_library/frameworks/mlx/lazy-evaluation-stream-lifetimes/references/source-map.md": "798b9cb89ed6b885eb47464ca9f137d2802a54571b9327e631d7742d3e5abd5d",
     "src/inferrsi/data/skill_library/frameworks/mlx/lazy-evaluation-stream-lifetimes/references/validation.md": "ebfc00bf7d6ab0a1dc415953ae84d56ccdc713d5614aff85e433c4a843793bff",
-    "src/inferrsi/data/skill_library/skill-index.json": "8ce35e18f7629260d67a95617b9918eecff73ed8d9370dd3d4595c725ada9aa1"
+    "src/inferrsi/data/skill_library/skill-index.json": "ea80d57e29a2949cffceefcc1715db375937499144864b9df41c3e7fcae61c16"
   },
   "skill_metadata": {
     "name": "lazy-evaluation-stream-lifetimes",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/npu-skill-factory.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/npu-skill-factory.json
index 5a1045141..01e752168 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/npu-skill-factory.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/npu-skill-factory.json
@@ -22,8 +22,8 @@
   "sources": [
     {
       "type": "git",
-      "url": "https://github.com/InferMatrixLab/InferRSI/blob/f321699b42a490276ffa4ea52350591683b575cc/src/inferrsi/data/skill_library/workflows/npu-skill-factory/SKILL.md",
-      "commit": "f321699b42a490276ffa4ea52350591683b575cc",
+      "url": "https://github.com/InferMatrixLab/InferRSI/blob/3291ffbeab2fe72b217d21c8cac282a4f5348267/src/inferrsi/data/skill_library/workflows/npu-skill-factory/SKILL.md",
+      "commit": "3291ffbeab2fe72b217d21c8cac282a4f5348267",
       "paths": [
         "src/inferrsi/data/skill_library/skill-index.json",
         "src/inferrsi/data/skill_library/workflows/npu-skill-factory/SKILL.md",
@@ -54,7 +54,7 @@
   },
   "checkout_ref": "checkout:npu-skill-factory",
   "source_files": {
-    "src/inferrsi/data/skill_library/skill-index.json": "8ce35e18f7629260d67a95617b9918eecff73ed8d9370dd3d4595c725ada9aa1",
+    "src/inferrsi/data/skill_library/skill-index.json": "ea80d57e29a2949cffceefcc1715db375937499144864b9df41c3e7fcae61c16",
     "src/inferrsi/data/skill_library/workflows/npu-skill-factory/SKILL.md": "74cd863df2e80dc2bda1713d24207809173a3116f30325e3487e1e3517cfec1e",
     "src/inferrsi/data/skill_library/workflows/npu-skill-factory/metadata.json": "12fc9272bd5fbf02dd9b897d8ca40984b72ae54f20492ace5304711bbc4d71c3",
     "src/inferrsi/data/skill_library/workflows/npu-skill-factory/references/papers-20261015/COMPARISON.md": "38587678a96581f5c2dc697d982127343565504bffa191352811dffcc165aa23",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/optimization-search-policy.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/optimization-search-policy.json
index e53af73a2..ae2ace94e 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/optimization-search-policy.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/optimization-search-policy.json
@@ -21,8 +21,8 @@
   "sources": [
     {
       "type": "git",
-      "url": "https://github.com/InferMatrixLab/InferRSI/blob/f321699b42a490276ffa4ea52350591683b575cc/src/inferrsi/data/skill_library/methods/optimization-search-policy/SKILL.md",
-      "commit": "f321699b42a490276ffa4ea52350591683b575cc",
+      "url": "https://github.com/InferMatrixLab/InferRSI/blob/3291ffbeab2fe72b217d21c8cac282a4f5348267/src/inferrsi/data/skill_library/methods/optimization-search-policy/SKILL.md",
+      "commit": "3291ffbeab2fe72b217d21c8cac282a4f5348267",
       "paths": [
         "src/inferrsi/data/skill_library/methods/optimization-search-policy/SKILL.md",
         "src/inferrsi/data/skill_library/methods/optimization-search-policy/metadata.json",
@@ -51,7 +51,7 @@
     "src/inferrsi/data/skill_library/methods/optimization-search-policy/references/papers-20261015/knowledge/adaptation-interfaces.md": "5e761957cb5ecd93926e6ffc13deee7c830c6e8f0ee3c3846e9dd3b50dc74446",
     "src/inferrsi/data/skill_library/methods/optimization-search-policy/references/papers-20261015/references/source-notes.md": "54d83c3713045d59a48821bcf93598374ba4a630e3891ad39498f919317c4fb8",
     "src/inferrsi/data/skill_library/methods/optimization-search-policy/references/papers-20261015/updates/optimization-search-policy.md": "d3db44fa9a6831b86a02c638b323ed943e69fe256866213f10d68c72d762879e",
-    "src/inferrsi/data/skill_library/skill-index.json": "8ce35e18f7629260d67a95617b9918eecff73ed8d9370dd3d4595c725ada9aa1"
+    "src/inferrsi/data/skill_library/skill-index.json": "ea80d57e29a2949cffceefcc1715db375937499144864b9df41c3e7fcae61c16"
   },
   "skill_metadata": {
     "name": "optimization-search-policy",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/os-process-placement.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/os-process-placement.json
index 080d9f9fd..da1e6f6d3 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/os-process-placement.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/os-process-placement.json
@@ -24,8 +24,8 @@
   "sources": [
     {
       "type": "git",
-      "url": "https://github.com/InferMatrixLab/InferRSI/blob/f321699b42a490276ffa4ea52350591683b575cc/src/inferrsi/data/skill_library/hardware/host/os-process-placement/SKILL.md",
-      "commit": "f321699b42a490276ffa4ea52350591683b575cc",
+      "url": "https://github.com/InferMatrixLab/InferRSI/blob/3291ffbeab2fe72b217d21c8cac282a4f5348267/src/inferrsi/data/skill_library/hardware/host/os-process-placement/SKILL.md",
+      "commit": "3291ffbeab2fe72b217d21c8cac282a4f5348267",
       "paths": [
         "src/inferrsi/data/skill_library/hardware/host/os-process-placement/SKILL.md",
         "src/inferrsi/data/skill_library/hardware/host/os-process-placement/metadata.json",
@@ -44,7 +44,7 @@
     "src/inferrsi/data/skill_library/hardware/host/os-process-placement/SKILL.md": "2003492e4d89fde0443b180d74933e764bd1979435e109c8d26e20eb8f6e5053",
     "src/inferrsi/data/skill_library/hardware/host/os-process-placement/metadata.json": "ec02995c075f0a540921f57b4846f2cafe334eb90f44389f17c3c50457c3e9c5",
     "src/inferrsi/data/skill_library/hardware/host/os-process-placement/references/llama-numa-thread-readset.md": "748b254b3b043a5ddb0f480eb4a566854d27eedd7b111499863e9bf0a5a6951e",
-    "src/inferrsi/data/skill_library/skill-index.json": "8ce35e18f7629260d67a95617b9918eecff73ed8d9370dd3d4595c725ada9aa1"
+    "src/inferrsi/data/skill_library/skill-index.json": "ea80d57e29a2949cffceefcc1715db375937499144864b9df41c3e7fcae61c16"
   },
   "skill_metadata": {
     "name": "os-process-placement",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/performance-ceiling-analysis.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/performance-ceiling-analysis.json
index 51d4d4a39..9a5fe5d2d 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/performance-ceiling-analysis.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/performance-ceiling-analysis.json
@@ -22,8 +22,8 @@
   "sources": [
     {
       "type": "git",
-      "url": "https://github.com/InferMatrixLab/InferRSI/blob/f321699b42a490276ffa4ea52350591683b575cc/src/inferrsi/data/skill_library/methods/performance-ceiling-analysis/SKILL.md",
-      "commit": "f321699b42a490276ffa4ea52350591683b575cc",
+      "url": "https://github.com/InferMatrixLab/InferRSI/blob/3291ffbeab2fe72b217d21c8cac282a4f5348267/src/inferrsi/data/skill_library/methods/performance-ceiling-analysis/SKILL.md",
+      "commit": "3291ffbeab2fe72b217d21c8cac282a4f5348267",
       "paths": [
         "src/inferrsi/data/skill_library/methods/performance-ceiling-analysis/SKILL.md",
         "src/inferrsi/data/skill_library/methods/performance-ceiling-analysis/metadata.json",
@@ -78,7 +78,7 @@
     "src/inferrsi/data/skill_library/methods/performance-ceiling-analysis/references/roofline-contract.md": "419ec4bcaec45ac579a133bce083d4f48779f26e664d7a86f720fa65f343b1d1",
     "src/inferrsi/data/skill_library/methods/performance-ceiling-analysis/references/roofline-example.json": "597cd3205df043cff3ae9d15924c77934e8d8eb300c00c4840f6bb2921f90faf",
     "src/inferrsi/data/skill_library/methods/performance-ceiling-analysis/scripts/roofline.py": "524de48a6764401695fe1444f1fa3d38f363069b0cd2fba4715a86ca4b28a5d4",
-    "src/inferrsi/data/skill_library/skill-index.json": "8ce35e18f7629260d67a95617b9918eecff73ed8d9370dd3d4595c725ada9aa1"
+    "src/inferrsi/data/skill_library/skill-index.json": "ea80d57e29a2949cffceefcc1715db375937499144864b9df41c3e7fcae61c16"
   },
   "skill_metadata": {
     "name": "performance-ceiling-analysis",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/plan-activation-residency.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/plan-activation-residency.json
index 0898c3cb9..a5e5c3e13 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/plan-activation-residency.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/plan-activation-residency.json
@@ -16,8 +16,8 @@
   "sources": [
     {
       "type": "git",
-      "url": "https://github.com/InferMatrixLab/InferRSI/blob/f321699b42a490276ffa4ea52350591683b575cc/src/inferrsi/data/skill_library/methods/plan-activation-residency/SKILL.md",
-      "commit": "f321699b42a490276ffa4ea52350591683b575cc",
+      "url": "https://github.com/InferMatrixLab/InferRSI/blob/3291ffbeab2fe72b217d21c8cac282a4f5348267/src/inferrsi/data/skill_library/methods/plan-activation-residency/SKILL.md",
+      "commit": "3291ffbeab2fe72b217d21c8cac282a4f5348267",
       "paths": [
         "src/inferrsi/data/skill_library/methods/plan-activation-residency/SKILL.md",
         "src/inferrsi/data/skill_library/methods/plan-activation-residency/metadata.json",
@@ -36,7 +36,7 @@
     "src/inferrsi/data/skill_library/methods/plan-activation-residency/SKILL.md": "542d8a1e6963f5019986e1fbd79d3a247a045a51487c09313026087ac457c15e",
     "src/inferrsi/data/skill_library/methods/plan-activation-residency/metadata.json": "c6acae57d70190cea0e18c71c79127962181e58169e011df75d8635b3db4daa7",
     "src/inferrsi/data/skill_library/methods/plan-activation-residency/references/source-notes.md": "5ff52655cfd06fcf970ed2a7d4d4fe1c31115b102af2d41497de0c74848ed2b3",
-    "src/inferrsi/data/skill_library/skill-index.json": "8ce35e18f7629260d67a95617b9918eecff73ed8d9370dd3d4595c725ada9aa1"
+    "src/inferrsi/data/skill_library/skill-index.json": "ea80d57e29a2949cffceefcc1715db375937499144864b9df41c3e7fcae61c16"
   },
   "skill_metadata": {
     "name": "plan-activation-residency",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/repository-evolution-analysis.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/repository-evolution-analysis.json
index a1a7eb39b..028892325 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/repository-evolution-analysis.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/repository-evolution-analysis.json
@@ -21,8 +21,8 @@
   "sources": [
     {
       "type": "git",
-      "url": "https://github.com/InferMatrixLab/InferRSI/blob/f321699b42a490276ffa4ea52350591683b575cc/src/inferrsi/data/skill_library/methods/repository-evolution-analysis/SKILL.md",
-      "commit": "f321699b42a490276ffa4ea52350591683b575cc",
+      "url": "https://github.com/InferMatrixLab/InferRSI/blob/3291ffbeab2fe72b217d21c8cac282a4f5348267/src/inferrsi/data/skill_library/methods/repository-evolution-analysis/SKILL.md",
+      "commit": "3291ffbeab2fe72b217d21c8cac282a4f5348267",
       "paths": [
         "src/inferrsi/data/skill_library/methods/repository-evolution-analysis/SKILL.md",
         "src/inferrsi/data/skill_library/methods/repository-evolution-analysis/metadata.json",
@@ -41,7 +41,7 @@
     "src/inferrsi/data/skill_library/methods/repository-evolution-analysis/SKILL.md": "7e20d3390f2501263e5e298e026c47cac16c234f2397833131e15ae5d34c0b97",
     "src/inferrsi/data/skill_library/methods/repository-evolution-analysis/metadata.json": "d0e4d6950250e4cec43d82e23324c9301a925d4c22be30b4d2713d790cb6504b",
     "src/inferrsi/data/skill_library/methods/repository-evolution-analysis/references/capability-presence-audit.md": "c21eb4a6aabe811a49a6540a34c54dbcfdcf7efae3052a79b54034c7838026a5",
-    "src/inferrsi/data/skill_library/skill-index.json": "8ce35e18f7629260d67a95617b9918eecff73ed8d9370dd3d4595c725ada9aa1"
+    "src/inferrsi/data/skill_library/skill-index.json": "ea80d57e29a2949cffceefcc1715db375937499144864b9df41c3e7fcae61c16"
   },
   "skill_metadata": {
     "name": "repository-evolution-analysis",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/rsi-evidence.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/rsi-evidence.json
index af6d0ea37..84a929439 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/rsi-evidence.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/rsi-evidence.json
@@ -16,8 +16,8 @@
   "sources": [
     {
       "type": "git",
-      "url": "https://github.com/InferMatrixLab/InferRSI/blob/f321699b42a490276ffa4ea52350591683b575cc/src/inferrsi/data/skill_library/workflows/rsi-evidence/SKILL.md",
-      "commit": "f321699b42a490276ffa4ea52350591683b575cc",
+      "url": "https://github.com/InferMatrixLab/InferRSI/blob/3291ffbeab2fe72b217d21c8cac282a4f5348267/src/inferrsi/data/skill_library/workflows/rsi-evidence/SKILL.md",
+      "commit": "3291ffbeab2fe72b217d21c8cac282a4f5348267",
       "paths": [
         "src/inferrsi/data/skill_library/skill-index.json",
         "src/inferrsi/data/skill_library/workflows/rsi-evidence/SKILL.md"
@@ -31,7 +31,7 @@
   },
   "checkout_ref": "checkout:rsi-evidence",
   "source_files": {
-    "src/inferrsi/data/skill_library/skill-index.json": "8ce35e18f7629260d67a95617b9918eecff73ed8d9370dd3d4595c725ada9aa1",
+    "src/inferrsi/data/skill_library/skill-index.json": "ea80d57e29a2949cffceefcc1715db375937499144864b9df41c3e7fcae61c16",
     "src/inferrsi/data/skill_library/workflows/rsi-evidence/SKILL.md": "a55610874b658b86a39a1a049efb4e4eab0140f6579ed74dd7718a4ca8cedaca"
   },
   "skill_metadata": {
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/skill-evolution-playbook.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/skill-evolution-playbook.json
index 8b46a9dec..8ea519a18 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/skill-evolution-playbook.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/skill-evolution-playbook.json
@@ -21,8 +21,8 @@
   "sources": [
     {
       "type": "git",
-      "url": "https://github.com/InferMatrixLab/InferRSI/blob/f321699b42a490276ffa4ea52350591683b575cc/src/inferrsi/data/skill_library/methods/skill-evolution-playbook/SKILL.md",
-      "commit": "f321699b42a490276ffa4ea52350591683b575cc",
+      "url": "https://github.com/InferMatrixLab/InferRSI/blob/3291ffbeab2fe72b217d21c8cac282a4f5348267/src/inferrsi/data/skill_library/methods/skill-evolution-playbook/SKILL.md",
+      "commit": "3291ffbeab2fe72b217d21c8cac282a4f5348267",
       "paths": [
         "src/inferrsi/data/skill_library/methods/skill-evolution-playbook/SKILL.md",
         "src/inferrsi/data/skill_library/methods/skill-evolution-playbook/metadata.json",
@@ -39,7 +39,7 @@
   "source_files": {
     "src/inferrsi/data/skill_library/methods/skill-evolution-playbook/SKILL.md": "9cd37f21fe0d862069e4798edca8a04d59f8b840e55d521901594a63dddd9ba3",
     "src/inferrsi/data/skill_library/methods/skill-evolution-playbook/metadata.json": "92b36338b9257e2e41b95e4370694d7de52ed84bc7971056b3bfced07b0993d8",
-    "src/inferrsi/data/skill_library/skill-index.json": "8ce35e18f7629260d67a95617b9918eecff73ed8d9370dd3d4595c725ada9aa1"
+    "src/inferrsi/data/skill_library/skill-index.json": "ea80d57e29a2949cffceefcc1715db375937499144864b9df41c3e7fcae61c16"
   },
   "skill_metadata": {
     "name": "skill-evolution-playbook",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/storage-io-measurement.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/storage-io-measurement.json
index 076a3c9f7..380d97123 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/storage-io-measurement.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/storage-io-measurement.json
@@ -24,8 +24,8 @@
   "sources": [
     {
       "type": "git",
-      "url": "https://github.com/InferMatrixLab/InferRSI/blob/f321699b42a490276ffa4ea52350591683b575cc/src/inferrsi/data/skill_library/hardware/host/storage-io-measurement/SKILL.md",
-      "commit": "f321699b42a490276ffa4ea52350591683b575cc",
+      "url": "https://github.com/InferMatrixLab/InferRSI/blob/3291ffbeab2fe72b217d21c8cac282a4f5348267/src/inferrsi/data/skill_library/hardware/host/storage-io-measurement/SKILL.md",
+      "commit": "3291ffbeab2fe72b217d21c8cac282a4f5348267",
       "paths": [
         "src/inferrsi/data/skill_library/hardware/host/storage-io-measurement/SKILL.md",
         "src/inferrsi/data/skill_library/hardware/host/storage-io-measurement/metadata.json",
@@ -42,7 +42,7 @@
   "source_files": {
     "src/inferrsi/data/skill_library/hardware/host/storage-io-measurement/SKILL.md": "63b883c10a8d3e7d918abfc7175b72e2b401ff1e02697f94b8382ec1c8f369b7",
     "src/inferrsi/data/skill_library/hardware/host/storage-io-measurement/metadata.json": "0a41c7b00978332b54ee1e4a7c1ade88ecde3e961128f24d1b4f7708caff5ff5",
-    "src/inferrsi/data/skill_library/skill-index.json": "8ce35e18f7629260d67a95617b9918eecff73ed8d9370dd3d4595c725ada9aa1"
+    "src/inferrsi/data/skill_library/skill-index.json": "ea80d57e29a2949cffceefcc1715db375937499144864b9df41c3e7fcae61c16"
   },
   "skill_metadata": {
     "name": "storage-io-measurement",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/training-input-pipeline.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/training-input-pipeline.json
index ffe45aa8e..4af1db19e 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/training-input-pipeline.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/training-input-pipeline.json
@@ -23,8 +23,8 @@
   "sources": [
     {
       "type": "git",
-      "url": "https://github.com/InferMatrixLab/InferRSI/blob/f321699b42a490276ffa4ea52350591683b575cc/src/inferrsi/data/skill_library/methods/training-input-pipeline/SKILL.md",
-      "commit": "f321699b42a490276ffa4ea52350591683b575cc",
+      "url": "https://github.com/InferMatrixLab/InferRSI/blob/3291ffbeab2fe72b217d21c8cac282a4f5348267/src/inferrsi/data/skill_library/methods/training-input-pipeline/SKILL.md",
+      "commit": "3291ffbeab2fe72b217d21c8cac282a4f5348267",
       "paths": [
         "src/inferrsi/data/skill_library/methods/training-input-pipeline/SKILL.md",
         "src/inferrsi/data/skill_library/methods/training-input-pipeline/metadata.json",
@@ -43,7 +43,7 @@
     "src/inferrsi/data/skill_library/methods/training-input-pipeline/SKILL.md": "64bd2c6b795d966045d1d3cca55976d3d49b08b06df7c7e06a70c2824cf55b7e",
     "src/inferrsi/data/skill_library/methods/training-input-pipeline/metadata.json": "80659099fce6a6ec05e9084b75452c877a5812e30f60a78f31ef8c7efcdc1bac",
     "src/inferrsi/data/skill_library/methods/training-input-pipeline/references/sources.md": "60e0f3e88655d8e1d5eb4797f39ca4370deda01614f265b8d2ea4559d8ac020f",
-    "src/inferrsi/data/skill_library/skill-index.json": "8ce35e18f7629260d67a95617b9918eecff73ed8d9370dd3d4595c725ada9aa1"
+    "src/inferrsi/data/skill_library/skill-index.json": "ea80d57e29a2949cffceefcc1715db375937499144864b9df41c3e7fcae61c16"
   },
   "skill_metadata": {
     "name": "training-input-pipeline",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/training-resource-budget.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/training-resource-budget.json
index 8d02f9a53..c321a7ecf 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/training-resource-budget.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/training-resource-budget.json
@@ -23,8 +23,8 @@
   "sources": [
     {
       "type": "git",
-      "url": "https://github.com/InferMatrixLab/InferRSI/blob/f321699b42a490276ffa4ea52350591683b575cc/src/inferrsi/data/skill_library/methods/training-resource-budget/SKILL.md",
-      "commit": "f321699b42a490276ffa4ea52350591683b575cc",
+      "url": "https://github.com/InferMatrixLab/InferRSI/blob/3291ffbeab2fe72b217d21c8cac282a4f5348267/src/inferrsi/data/skill_library/methods/training-resource-budget/SKILL.md",
+      "commit": "3291ffbeab2fe72b217d21c8cac282a4f5348267",
       "paths": [
         "src/inferrsi/data/skill_library/methods/training-resource-budget/SKILL.md",
         "src/inferrsi/data/skill_library/methods/training-resource-budget/metadata.json",
@@ -57,7 +57,7 @@
     "src/inferrsi/data/skill_library/methods/training-resource-budget/references/papers-20261015/references/source-readset.md": "2a5c166767f56fa47e7eba7386f1f36da8cb73702e5b5bbe14d6a456ca9ebfae",
     "src/inferrsi/data/skill_library/methods/training-resource-budget/references/sources.md": "6d248f58af8b350fc66c22898988d6f451ac60d03bcd805cd6e84df5e55174ac",
     "src/inferrsi/data/skill_library/methods/training-resource-budget/references/stas-decision-notes.md": "f1a993e6c4ca9a035a2effcaf5500081f227b0cc56cc885a6b22fda0041fc744",
-    "src/inferrsi/data/skill_library/skill-index.json": "8ce35e18f7629260d67a95617b9918eecff73ed8d9370dd3d4595c725ada9aa1"
+    "src/inferrsi/data/skill_library/skill-index.json": "ea80d57e29a2949cffceefcc1715db375937499144864b9df41c3e7fcae61c16"
   },
   "skill_metadata": {
     "name": "training-resource-budget",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/weight-loading-optimization.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/weight-loading-optimization.json
index e7910bf4d..3feebc266 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/weight-loading-optimization.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/weight-loading-optimization.json
@@ -26,8 +26,8 @@
   "sources": [
     {
       "type": "git",
-      "url": "https://github.com/InferMatrixLab/InferRSI/blob/f321699b42a490276ffa4ea52350591683b575cc/src/inferrsi/data/skill_library/methods/weight-loading-optimization/SKILL.md",
-      "commit": "f321699b42a490276ffa4ea52350591683b575cc",
+      "url": "https://github.com/InferMatrixLab/InferRSI/blob/3291ffbeab2fe72b217d21c8cac282a4f5348267/src/inferrsi/data/skill_library/methods/weight-loading-optimization/SKILL.md",
+      "commit": "3291ffbeab2fe72b217d21c8cac282a4f5348267",
       "paths": [
         "src/inferrsi/data/skill_library/methods/weight-loading-optimization/SKILL.md",
         "src/inferrsi/data/skill_library/methods/weight-loading-optimization/metadata.json",
@@ -48,7 +48,7 @@
     "src/inferrsi/data/skill_library/methods/weight-loading-optimization/metadata.json": "798cbff6f10b7c24b188080ebf233b188b0e98ef2c6fdb5e00e1fe3eff727c0f",
     "src/inferrsi/data/skill_library/methods/weight-loading-optimization/references/fastsafetensors-transfer-readset.md": "5a39f14b102b0cee304541e011aa31e44e6b13527d1bee9987c3f3b66c064d3d",
     "src/inferrsi/data/skill_library/methods/weight-loading-optimization/references/llama-loading-readset.md": "6c640f28ee3f0253a17b81461c865cb9f947367002cb75bf72550c4572d02e97",
-    "src/inferrsi/data/skill_library/skill-index.json": "8ce35e18f7629260d67a95617b9918eecff73ed8d9370dd3d4595c725ada9aa1"
+    "src/inferrsi/data/skill_library/skill-index.json": "ea80d57e29a2949cffceefcc1715db375937499144864b9df41c3e7fcae61c16"
   },
   "skill_metadata": {
     "name": "weight-loading-optimization",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/cuda/skills/audit-afd-p2p-exchanges.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/cuda/skills/audit-afd-p2p-exchanges.json
index 7dfe3e179..1e4b9bb33 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/cuda/skills/audit-afd-p2p-exchanges.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/cuda/skills/audit-afd-p2p-exchanges.json
@@ -23,8 +23,8 @@
   "sources": [
     {
       "type": "git",
-      "url": "https://github.com/InferMatrixLab/InferRSI/blob/f321699b42a490276ffa4ea52350591683b575cc/src/inferrsi/data/skill_library/repositories/afd-plugin/audit-afd-p2p-exchanges/SKILL.md",
-      "commit": "f321699b42a490276ffa4ea52350591683b575cc",
+      "url": "https://github.com/InferMatrixLab/InferRSI/blob/3291ffbeab2fe72b217d21c8cac282a4f5348267/src/inferrsi/data/skill_library/repositories/afd-plugin/audit-afd-p2p-exchanges/SKILL.md",
+      "commit": "3291ffbeab2fe72b217d21c8cac282a4f5348267",
       "paths": [
         "src/inferrsi/data/skill_library/repositories/afd-plugin/audit-afd-p2p-exchanges/SKILL.md",
         "src/inferrsi/data/skill_library/repositories/afd-plugin/audit-afd-p2p-exchanges/metadata.json",
@@ -45,7 +45,7 @@
     "src/inferrsi/data/skill_library/repositories/afd-plugin/audit-afd-p2p-exchanges/metadata.json": "4e6100dfb9fc7bc00b19915f4fd38ad509704f2a609c658adf4d632d4e05b6c1",
     "src/inferrsi/data/skill_library/repositories/afd-plugin/audit-afd-p2p-exchanges/references/afd-plugin-LICENSE.md": "0e8566ea362276878e1f38f97da49964dbea301c15d4aa9d727f97a4c39fd943",
     "src/inferrsi/data/skill_library/repositories/afd-plugin/audit-afd-p2p-exchanges/references/source-notes.md": "5073d87450cb60d1b245308dff8807e73749c07de976d50f2fa26694fad81a72",
-    "src/inferrsi/data/skill_library/skill-index.json": "8ce35e18f7629260d67a95617b9918eecff73ed8d9370dd3d4595c725ada9aa1"
+    "src/inferrsi/data/skill_library/skill-index.json": "ea80d57e29a2949cffceefcc1715db375937499144864b9df41c3e7fcae61c16"
   },
   "skill_metadata": {
     "name": "audit-afd-p2p-exchanges",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/cuda/skills/diagnose-attention-bottlenecks.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/cuda/skills/diagnose-attention-bottlenecks.json
index 496d7c545..19f027ab8 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/cuda/skills/diagnose-attention-bottlenecks.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/cuda/skills/diagnose-attention-bottlenecks.json
@@ -16,8 +16,8 @@
   "sources": [
     {
       "type": "git",
-      "url": "https://github.com/InferMatrixLab/InferRSI/blob/f321699b42a490276ffa4ea52350591683b575cc/src/inferrsi/data/skill_library/hardware/cuda/diagnose-attention-bottlenecks/SKILL.md",
-      "commit": "f321699b42a490276ffa4ea52350591683b575cc",
+      "url": "https://github.com/InferMatrixLab/InferRSI/blob/3291ffbeab2fe72b217d21c8cac282a4f5348267/src/inferrsi/data/skill_library/hardware/cuda/diagnose-attention-bottlenecks/SKILL.md",
+      "commit": "3291ffbeab2fe72b217d21c8cac282a4f5348267",
       "paths": [
         "src/inferrsi/data/skill_library/hardware/cuda/diagnose-attention-bottlenecks/SKILL.md",
         "src/inferrsi/data/skill_library/hardware/cuda/diagnose-attention-bottlenecks/metadata.json",
@@ -52,7 +52,7 @@
     "src/inferrsi/data/skill_library/hardware/cuda/diagnose-attention-bottlenecks/references/papers-20261015/cases/partitioning-after-fusion.md": "4d4cabe0991eab554dcfbd4fa9c8199e61fa4767709b0c4cd646e2588bb398a2",
     "src/inferrsi/data/skill_library/hardware/cuda/diagnose-attention-bottlenecks/references/papers-20261015/knowledge/attention-cost-model-boundaries.md": "5ed8f888443e38ccb72545b6025447aab7b92936f78c151fed680e582126239b",
     "src/inferrsi/data/skill_library/hardware/cuda/diagnose-attention-bottlenecks/references/sources.md": "53436e0de4e6dda2d29426591bf61f8526378cdc6f9ec7967b88b1bcd4c61e6f",
-    "src/inferrsi/data/skill_library/skill-index.json": "8ce35e18f7629260d67a95617b9918eecff73ed8d9370dd3d4595c725ada9aa1"
+    "src/inferrsi/data/skill_library/skill-index.json": "ea80d57e29a2949cffceefcc1715db375937499144864b9df41c3e7fcae61c16"
   },
   "skill_metadata": {
     "name": "diagnose-attention-bottlenecks",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/cuda/skills/host-overhead-profiling.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/cuda/skills/host-overhead-profiling.json
index ccf4f80ec..f590679ce 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/cuda/skills/host-overhead-profiling.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/cuda/skills/host-overhead-profiling.json
@@ -22,8 +22,8 @@
   "sources": [
     {
       "type": "git",
-      "url": "https://github.com/InferMatrixLab/InferRSI/blob/f321699b42a490276ffa4ea52350591683b575cc/src/inferrsi/data/skill_library/hardware/cuda/host-overhead-profiling/SKILL.md",
-      "commit": "f321699b42a490276ffa4ea52350591683b575cc",
+      "url": "https://github.com/InferMatrixLab/InferRSI/blob/3291ffbeab2fe72b217d21c8cac282a4f5348267/src/inferrsi/data/skill_library/hardware/cuda/host-overhead-profiling/SKILL.md",
+      "commit": "3291ffbeab2fe72b217d21c8cac282a4f5348267",
       "paths": [
         "src/inferrsi/data/skill_library/hardware/cuda/host-overhead-profiling/SKILL.md",
         "src/inferrsi/data/skill_library/hardware/cuda/host-overhead-profiling/metadata.json",
@@ -40,7 +40,7 @@
   "source_files": {
     "src/inferrsi/data/skill_library/hardware/cuda/host-overhead-profiling/SKILL.md": "326e566fbe30d3ba04e4b50c3e87aad9df037f0ac4ad6fe129165ae97dc5f91d",
     "src/inferrsi/data/skill_library/hardware/cuda/host-overhead-profiling/metadata.json": "76e34c1c64aba473a9e126b032a776417caa0f913f6ac3cf4dba3a5d123b1ee0",
-    "src/inferrsi/data/skill_library/skill-index.json": "8ce35e18f7629260d67a95617b9918eecff73ed8d9370dd3d4595c725ada9aa1"
+    "src/inferrsi/data/skill_library/skill-index.json": "ea80d57e29a2949cffceefcc1715db375937499144864b9df41c3e7fcae61c16"
   },
   "skill_metadata": {
     "name": "host-overhead-profiling",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/cuda/skills/reason-hopper-kernels.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/cuda/skills/reason-hopper-kernels.json
index 06c5752f1..198d57561 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/cuda/skills/reason-hopper-kernels.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/cuda/skills/reason-hopper-kernels.json
@@ -36,8 +36,8 @@
   "sources": [
     {
       "type": "git",
-      "url": "https://github.com/InferMatrixLab/InferRSI/blob/f321699b42a490276ffa4ea52350591683b575cc/src/inferrsi/data/skill_library/hardware/cuda/reason-hopper-kernels/SKILL.md",
-      "commit": "f321699b42a490276ffa4ea52350591683b575cc",
+      "url": "https://github.com/InferMatrixLab/InferRSI/blob/3291ffbeab2fe72b217d21c8cac282a4f5348267/src/inferrsi/data/skill_library/hardware/cuda/reason-hopper-kernels/SKILL.md",
+      "commit": "3291ffbeab2fe72b217d21c8cac282a4f5348267",
       "paths": [
         "src/inferrsi/data/skill_library/hardware/cuda/reason-hopper-kernels/SKILL.md",
         "src/inferrsi/data/skill_library/hardware/cuda/reason-hopper-kernels/metadata.json",
@@ -72,7 +72,7 @@
     "src/inferrsi/data/skill_library/hardware/cuda/reason-hopper-kernels/references/source-inventory.json": "1dc09c895c970a0c9708158bcc6dd65cdb736274fd2c35d44542c70407d3849d",
     "src/inferrsi/data/skill_library/hardware/cuda/reason-hopper-kernels/references/sources.md": "6b74b9e774f2a80db0860510d9e6940d36284430f29c1e124f115a9eaf25833d",
     "src/inferrsi/data/skill_library/hardware/cuda/reason-hopper-kernels/references/synchronization.md": "4322f6b3c0e4368bf8c192b4ec752529767f50d19c0c38dad8fdf9cd781e41d9",
-    "src/inferrsi/data/skill_library/skill-index.json": "8ce35e18f7629260d67a95617b9918eecff73ed8d9370dd3d4595c725ada9aa1"
+    "src/inferrsi/data/skill_library/skill-index.json": "ea80d57e29a2949cffceefcc1715db375937499144864b9df41c3e7fcae61c16"
   },
   "skill_metadata": {
     "name": "reason-hopper-kernels",
diff --git a/src/inferrsi/data/skill_library/README.md b/src/inferrsi/data/skill_library/README.md
index 1e41f2308..32f41456b 100644
--- a/src/inferrsi/data/skill_library/README.md
+++ b/src/inferrsi/data/skill_library/README.md
@@ -1,7 +1,7 @@
 # InferRSI skill library
 
-The English-only library contains **81 skill documents**: 60 registered skills
-and 21 engine and research-router skills. Complete procedures, metadata,
+The English-only library contains **85 skill documents**: 61 registered roots (including the DwarfStar graph),
+21 separately indexed engine/router documents and three DwarfStar sub-skills. Complete procedures, metadata,
 references, cases and source notices ship with the package.
 
 ## Browse by owner
@@ -20,7 +20,7 @@ existing resource directory.
 
 ## Read registered skills
 
-[`skill-index.json`](skill-index.json) maps 60 stable skill IDs to their canonical
+[`skill-index.json`](skill-index.json) maps 61 stable skill IDs to their canonical
 paths and backend scopes. Install the optional `skills` extra for these commands:
 
 ```bash
@@ -118,7 +118,7 @@ guidance; it establishes no cross-pin implementation compatibility.
 
 Five repository entry skills, 15 focused sub-skills and one research router
 provide source-grounded procedures. These 21 documents are available through
-the links and catalog records below, separately from the 60-ID registry.
+the links and catalog records below, separately from the 61-ID registry.
 The registered ComfyUI and stable-diffusion.cpp audits are listed alongside
 these imports and use generated abstracts rather than runtime repository packs.
 
@@ -227,3 +227,12 @@ separates mined mechanisms, unmined source headers and Blackwell-only recipes;
 [source/license qualifications](hardware/cuda/reason-hopper-kernels/references/sources.md)
 retain contradictions and distribution discrepancies. GPU execution and efficacy
 remain not_run.
+
+## DwarfStar portable AREX candidate
+
+[DwarfStar](repositories/dwarfstar/SKILL.md) routes memory/SSD, serving-state and
+qualification questions to one source-pinned operating graph. The portable AREX
+create/verify workflow was adapted to native C with the user-approved partial
+backend boundary. CPU checks do not qualify accelerated/model/service execution;
+required backend gaps remain explicit. Its generated catalog ID is
+`inferrsi.common.skills.dwarfstar`.
diff --git a/src/inferrsi/data/skill_library/repositories/dwarfstar/LICENSE b/src/inferrsi/data/skill_library/repositories/dwarfstar/LICENSE
new file mode 100644
index 000000000..c5b10d432
--- /dev/null
+++ b/src/inferrsi/data/skill_library/repositories/dwarfstar/LICENSE
@@ -0,0 +1,23 @@
+MIT License
+
+Copyright (c) 2026 The ds4.c authors
+Copyright (c) 2023-2026 The ggml authors
+Copyright (c) 2023 DeepSeek
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.
diff --git a/src/inferrsi/data/skill_library/repositories/dwarfstar/SKILL.md b/src/inferrsi/data/skill_library/repositories/dwarfstar/SKILL.md
new file mode 100644
index 000000000..e06a3425e
--- /dev/null
+++ b/src/inferrsi/data/skill_library/repositories/dwarfstar/SKILL.md
@@ -0,0 +1,86 @@
+---
+name: dwarfstar
+description: "Investigate DwarfStar model fit, expert SSD streaming, session and vision cache correctness, speculative sampling and distributed execution using pinned source contracts and explicit native verification boundaries."
+license: NO_LICENSE
+disable-model-invocation: true
+metadata:
+  disco-role: operating
+---
+
+# DwarfStar inference engineering
+
+Use this candidate to turn a DwarfStar capacity, state-correctness or execution
+question into a bounded hypothesis and discriminating check. It distills
+`antirez/ds4`, not an arbitrary GGUF runner. Start with the
+[source and method contract](references/repo-provenance.md).
+
+The operating guidance is source-derived. A local CPU build or bookkeeping test
+does not validate Metal/CUDA/ROCm kernels, model output, a real service, physical
+tensor parallelism or a performance gain. Required backend evidence remains
+**BLOCKED_REQUIRED_BACKEND**; this candidate is not fully backend-verified.
+
+## Route the question
+
+| Question | Read next | Expected output |
+| --- | --- | --- |
+| A model fits on disk but not in RAM/VRAM; cache grows; more sessions fail | [runtime-memory](sub-skills/runtime-memory/SKILL.md) | Representation and phase/owner memory ledger, one causal hypothesis |
+| A repeated prompt misses, reuses wrong image/tool history or corrupts a continuation | [serving-state](sub-skills/serving-state/SKILL.md) | Token/image/state identity and invalidation ledger |
+| A draft is faster but changes results; a TP parser test passes; a backend falls back | [qualification](sub-skills/qualification/SKILL.md) | Matching correctness test, actual execution path and remaining native gap |
+
+Use the bundled [request builder](scripts/request_template.py) when preparing a
+minimal Chat Completions request. It prints JSON only: no engine, network,
+credential, model download or service is started. Its options are deliberately
+smaller than the upstream API.
+
+## Establish the execution tuple
+
+Record the engine revision and patches, model family/checkpoint/quantization and
+artifact identity, selected native backend/build, device and transport topology,
+RAM/VRAM/storage budget, context/prefill chunk/session count, image inputs,
+sampling/seed and speculation policy. Add cold/warm/live/disk-cache state and
+the workload's quality and service requirements. Unknown fields remain unknown.
+
+Do not infer model support from a flag, shared quantization name, successful file
+mapping or API model alias. The actual loaded GGUF and model-specific graph
+select semantics. Flash, V4.1, PRO, GLM and Qwen routes differ.
+
+## Native preparation, not automatic installation
+
+DwarfStar is a native C engine, not a Python distribution. A prepared CPU
+inspection build can expose CLI help without weights; accelerated builds require
+their matching compiler/runtime. An operator with a compatible installed build
+can inspect `ds4 --help` and `ds4-server --help` before choosing options.
+For an authorized source installation, use the public repository at the
+reviewed source revision and the native `make cpu` target for CPU inspection;
+Metal/CUDA/ROCm require the matching native build described by this graph.
+Building, downloading weights, exposing a listener and changing system memory
+limits are separate actions requiring task authorization.
+
+This extraction adapted AREX's Python-only environment fields to native C build
+and CLI identity. It did not invent an installed Python package, run a DisCo
+model backend or install into a managed DisCo skill collection.
+
+## Work one bounded hypothesis
+
+1. Select the relevant route and read its bundled mechanism and failure notes.
+2. Separate source observations, upstream reports and proposed experiments.
+3. Identify the owner, lifetime, validity key and release/rollback boundary.
+4. Freeze controls before a comparison; change one mechanism or explicitly
+   declare a coupled change. Numeric gates and campaign revisions belong in
+   the acceptance issue, not in this reusable guidance.
+5. Verify the smallest invariant before measuring performance; retain failures,
+   refusals, fallback messages, cancellations and unavailable fixtures.
+6. For a service claim, run the accepted model/workload/quality/concurrency
+   protocol on the actual backend. Microtests are not service qualification.
+
+## Stop and report
+
+Return the executing tuple, observed signal, source-backed mechanism, alternate
+explanation, proposed falsifier, actual check outcome and unresolved gap. Stop
+at unavailable hardware/weights, unsafe exposure or the authorized budget.
+Do not disable a memory guard to turn a refusal into a capacity claim.
+
+Read [source notices](LICENSE) when redistributing adapted material.
+`NO_LICENSE` is the AREX resolver's unauthenticated-CLI warning fallback, not a
+claim that DwarfStar lacks a legal license; the pinned source license is MIT.
+Model weights and retained third-party code have independent obligations.
diff --git a/src/inferrsi/data/skill_library/repositories/dwarfstar/metadata.json b/src/inferrsi/data/skill_library/repositories/dwarfstar/metadata.json
new file mode 100644
index 000000000..1bebef173
--- /dev/null
+++ b/src/inferrsi/data/skill_library/repositories/dwarfstar/metadata.json
@@ -0,0 +1,89 @@
+{
+  "schema_version": 1,
+  "kind": "skill",
+  "origin": [
+    "external"
+  ],
+  "status": "candidate",
+  "empirical_validation": "partial_cpu_only",
+  "license": "NO_LICENSE",
+  "tags": [
+    "dwarfstar",
+    "ds4",
+    "moe",
+    "ssd-streaming",
+    "session-state",
+    "kv-cache",
+    "speculation",
+    "tensor-parallelism"
+  ],
+  "sources": [
+    {
+      "type": "git",
+      "url": "https://github.com/antirez/ds4",
+      "commit": "ee7e8405232c8971000cc7ba2b945b00bf6d20d3",
+      "paths": [
+        "README.md",
+        "LICENSE",
+        "Makefile",
+        "ds4.h",
+        "ds4.c",
+        "ds4_ssd.c",
+        "ds4_ssd.h",
+        "ds4_kvstore.c",
+        "ds4_kvstore.h",
+        "ds4_server.c",
+        "ds4_tp.c",
+        "ds4_tp.h",
+        "docs/SSD_STREAMING.md",
+        "docs/SERVER.md",
+        "docs/SPECULATIVE_DECODING.md",
+        "docs/DISTRIBUTED.md",
+        "docs/CUDA_MULTI_GPU.md",
+        "docs/TESTING.md",
+        "tests/test_session_state.c",
+        "tests/test_sampling.c",
+        "tests/test_tp_commands.c",
+        "tests/test_engine_mgpu_placement.c",
+        "ds4_cuda.cu",
+        "QA_BEFORE_RELEASES.md"
+      ],
+      "readset": "references/source-identity.json",
+      "license": "MIT"
+    },
+    {
+      "type": "git",
+      "url": "https://github.com/VectorSpaceLab/AREX-Skill",
+      "commit": "ac3fe1afa80fb9a09775ecfb2b6cc3ba850a2db6",
+      "paths": [
+        "cli/packages/coding-agent/src/disco/skills/create-repo-skill",
+        "cli/packages/coding-agent/src/disco/skills/prepare-repo-skill-env",
+        "cli/packages/coding-agent/src/disco/skills/workflow-authoring",
+        "cli/packages/coding-agent/src/disco/skills/verify-repo-skill"
+      ],
+      "scope": "Portable instructions executed by compatible agent with native C adaptation; no DisCo CLI/model run or managed import."
+    }
+  ],
+  "construction": {
+    "anchor_kind": "source",
+    "distillation_form": "repository-operating-graph",
+    "strategy": "arex-portable-meta-skills",
+    "status": "partial-candidate",
+    "native_import": "not-run",
+    "method": "User-approved native C adaptation and partial drafting; manual stable-ID subskill orchestration and integrated create/verify stages. Required-backend blocks disable automatic admission."
+  },
+  "applicability": {
+    "scope": "Pinned DwarfStar source investigation and experiment planning; not arbitrary GGUF support or cross-backend qualification."
+  },
+  "limitations": [
+    "Metal/CUDA/ROCm inference, actual model output, SSD runtime I/O, GPU recurrent rollback, physical distributed behavior, service quality/performance and skill efficacy remain unverified.",
+    "CPU build/help and selected bookkeeping tests do not clear BLOCKED_REQUIRED_BACKEND.",
+    "NO_LICENSE is the actual unauthenticated GitHub CLI resolver fallback, not a legal conclusion; pinned source MIT notice is retained.",
+    "Only selected source bodies were inspected; full kernels, conversion/imatrix tools and every workflow are not covered."
+  ],
+  "provenance_readsets": [
+    "references/repo-provenance.md",
+    "references/source-identity.json"
+  ],
+  "consumption_mode": "packaged"
+}
diff --git a/src/inferrsi/data/skill_library/repositories/dwarfstar/references/repo-provenance.md b/src/inferrsi/data/skill_library/repositories/dwarfstar/references/repo-provenance.md
new file mode 100644
index 000000000..6a74bf4cf
--- /dev/null
+++ b/src/inferrsi/data/skill_library/repositories/dwarfstar/references/repo-provenance.md
@@ -0,0 +1,56 @@
+# Source and method provenance
+
+Schema: `disco.repo-provenance.v1` (native-C adaptation).
+
+- Repository: `antirez/ds4` (public project name: DwarfStar).
+- Source commit: `ee7e8405232c8971000cc7ba2b945b00bf6d20d3`, branch `main`;
+  no exact tag asserted. The tracked source was clean at capture. Later local
+  CPU build products do not change the cited source inputs.
+- Package identity: native C/Make project; no Python distribution/version is
+  fabricated. Source revision is the version contract.
+- Method: AREX-Skill at `ac3fe1afa80fb9a09775ecfb2b6cc3ba850a2db6`, portable
+  `create-repo-skill`, `prepare-repo-skill-env`, `workflow-authoring` and
+  `verify-repo-skill` instructions executed by a compatible agent with manual
+  lane orchestration. This is not a DisCo CLI distillation/model-API run.
+- Destination: the user explicitly selected InferRSI candidate knowledge, not
+  a live DisCo managed collection. No managed importer or external agent export
+  was executed. InferRSI's own registry/catalog remain its discovery authority.
+
+## Bounded evidence
+
+The extraction covers README and selected model/runtime/SSD/server/speculation/
+distributed/testing guides; the public C header; SSD budget and disk-KV code;
+selected engine/session, prefill-workspace, sampling/rollback and server/replay
+bodies; and selected model-free test/build definitions. The
+[source identity inventory](source-identity.json) gives exact file hashes and
+the readset's bounded paths; a listed file does not mean every body was mined.
+
+Full conversion/imatrix tooling, model payloads, all kernel bodies, release
+automation, complete supported-platform QA and every failure path are outside
+this extraction. Speed tables are upstream reports, not reproduced results.
+The procedures preserve model/backend distinctions and known contradictory
+documentation instead of assuming a universal supported matrix.
+
+## License evidence
+
+The pinned source `LICENSE` is MIT and retains ds4.c, ggml and DeepSeek notices;
+the notice is bundled as `LICENSE` at the skill root. The AREX license resolver
+was actually run once for this commit but its GitHub CLI was unauthenticated.
+Its required fallback is `NO_LICENSE`, consistently used in root/sub-skill
+frontmatter. This warning is not a legal determination. Source-license evidence
+is retained separately from helper execution status. Model weights, external
+tools and source vendors do not acquire MIT licensing merely through this skill.
+
+## Validation boundary and refresh
+
+Source/structure assertions, native CPU build/help and selected model-free
+bookkeeping checks have narrower meanings than model/backend correctness.
+Accelerated inference, actual SSD I/O under a model, recurrent GPU rollback,
+quality, real service concurrency and physical multi-GPU/network operation
+remain blocked/unverified. Candidate knowledge does not constitute a campaign
+PASS, skill-efficacy result or permission to execute imported instructions.
+
+Before reusing on another revision, compare the selected paths and their
+consumers, model compatibility, flags, state serialization, sampling rules,
+backend eligibility and test definitions. Refresh affected guidance and checks;
+neither a changed hash alone nor a matching filename proves compatibility.
diff --git a/src/inferrsi/data/skill_library/repositories/dwarfstar/references/repo-routing-metadata.json b/src/inferrsi/data/skill_library/repositories/dwarfstar/references/repo-routing-metadata.json
new file mode 100644
index 000000000..b119b5f3d
--- /dev/null
+++ b/src/inferrsi/data/skill_library/repositories/dwarfstar/references/repo-routing-metadata.json
@@ -0,0 +1,13 @@
+{
+  "schema_version": "2.0",
+  "repo_id": "antirez/ds4",
+  "skill_id": "dwarfstar",
+  "taxonomy_sha256": "f8c306386015711634ddbb43a5eb95d1f58909c3513ce2063ba42efdd583a431",
+  "routing_status": "classified",
+  "assignments": [
+    {
+      "area": "Model Deployment and Optimization",
+      "family": "Inference Serving"
+    }
+  ]
+}
diff --git a/src/inferrsi/data/skill_library/repositories/dwarfstar/references/source-identity.json b/src/inferrsi/data/skill_library/repositories/dwarfstar/references/source-identity.json
new file mode 100644
index 000000000..69a6ef57f
--- /dev/null
+++ b/src/inferrsi/data/skill_library/repositories/dwarfstar/references/source-identity.json
@@ -0,0 +1,32 @@
+{
+  "schema_version": 1,
+  "repository": "antirez/ds4",
+  "commit": "ee7e8405232c8971000cc7ba2b945b00bf6d20d3",
+  "coverage": "Bounded selected bodies; file hashes identify source inputs, not whole-file mining or execution.",
+  "files": {
+    "README.md": "459c30f1fb1c994938c88520a129d6f0b7ff89594508c012a69b4ead17bd36c3",
+    "LICENSE": "4a8a1358f5b048627d71e4988de0bafd386f4988991929629e6cb9e29b7b9b2f",
+    "Makefile": "f2285f4b1069a31fd4db16b116cdf722eb793d98d57693176d3b6382ce49cf9c",
+    "ds4.h": "fd34c9485866cbf9b55e6e9cefeb0c19db4023de567d9c2196ddc3b91d6bc0d4",
+    "ds4.c": "bcc01e2d71f8324e31ffa798bea813ab45c583832558f476547f276ce57f94c4",
+    "ds4_ssd.c": "6f093186ee5cde476f08f808a993f46665ab25ac0c1b8c207cb404c7b94cd999",
+    "ds4_ssd.h": "61f4312a8b3686d5f5591f52365e6cb6301fd44d63775d5f490bbc2dd5c8e348",
+    "ds4_kvstore.c": "e8d9849f979470d500c5f3370a59bd5b10283eb694cae042f6103e8d1bebe2dc",
+    "ds4_kvstore.h": "e6d0e2836385285718436d89292d1052099b7cc9a2c8a1c79d3d59e214d81c3b",
+    "ds4_server.c": "eb23b37b957ffeb9bbe4c399842f02f7a9a7a771549fbdb99b919a2bf84f8ae6",
+    "ds4_tp.c": "306b23700135b9534a5da16f593a867c0611a5cb087e6189e0a47a88e9aa9d25",
+    "ds4_tp.h": "b6dde28c9fd414ad28d60392473b357bf66e41dee1afb0ee3a5c3c3dcb9ba2d2",
+    "docs/SSD_STREAMING.md": "cb45d910cc781fac5be403f4912b1c0a164151107970bd6df5665e0dc3a59496",
+    "docs/SERVER.md": "14e92c35f8e6c1db1a39f5e30676180ba753dfc1134a589e590a8307aba5e5dd",
+    "docs/SPECULATIVE_DECODING.md": "ce60aad0fca19557174615c663f50d431e42ea3d09cab1ee8c4c90950547c5cf",
+    "docs/DISTRIBUTED.md": "5e8f5d11863f7f0f0e22cb7dcbfc7c5f4d41dadbb8cac7552a3cfed8edb13bbd",
+    "docs/CUDA_MULTI_GPU.md": "8f72c6ec259167198a896c3ee2930b10ff7c99221f7e01b87f5b1292d530abee",
+    "docs/TESTING.md": "584d1bf616525403378f18107cbb2df64dfb75702ea042e04efd2084a60d16fe",
+    "tests/test_session_state.c": "59efd38f519d7cd75dd7ed77b30751df0228064b95db48c95bd08d509a6009b8",
+    "tests/test_sampling.c": "85d5e39c1239e673199c897742649381605fdd0793fca635940fcdb2fd2485e0",
+    "tests/test_tp_commands.c": "ca984c124a0f6f5f27016bac988f1caa5d763493c7f06650907f5dbc230202a4",
+    "tests/test_engine_mgpu_placement.c": "e90d775452a8622e67fd5b00d6d918123b3c2f4faba9c8fe7c269504a4dbdb6d",
+    "ds4_cuda.cu": "591ea3d389601666043147d4bcf726510abd1dd681eb84d082d01b0d6d68842a",
+    "QA_BEFORE_RELEASES.md": "0b4be818bf97252fba786770a85fca1f1d93b91b0843ee92cfdfadb1633c10ed"
+  }
+}
diff --git a/src/inferrsi/data/skill_library/repositories/dwarfstar/scripts/request_template.py b/src/inferrsi/data/skill_library/repositories/dwarfstar/scripts/request_template.py
new file mode 100644
index 000000000..ab8edc176
--- /dev/null
+++ b/src/inferrsi/data/skill_library/repositories/dwarfstar/scripts/request_template.py
@@ -0,0 +1,36 @@
+#!/usr/bin/env python3
+"""Build a small DwarfStar Chat Completions request without network activity.
+
+Adapted from the request shape in antirez/ds4 docs/SERVER.md at the source
+revision recorded by this skill. Prints JSON to stdout; never contacts a server.
+Example: python request_template.py --model local-model --prompt 'Hello'
+"""
+
+import argparse
+import json
+
+
+def main():
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("--model", required=True, help="API alias, not a model file selector")
+    parser.add_argument("--prompt", required=True)
+    parser.add_argument("--max-tokens", type=int, default=64)
+    parser.add_argument("--temperature", type=float, default=0.0)
+    parser.add_argument("--stream", action="store_true")
+    args = parser.parse_args()
+    if not args.model.strip() or not args.prompt.strip():
+        parser.error("model and prompt must be non-empty")
+    if args.max_tokens <= 0 or not 0 <= args.temperature <= 2:
+        parser.error("max-tokens must be positive; helper temperature range is [0, 2]")
+    request = {
+        "model": args.model,
+        "messages": [{"role": "user", "content": args.prompt}],
+        "max_tokens": args.max_tokens,
+        "temperature": args.temperature,
+        "stream": args.stream,
+    }
+    print(json.dumps(request, ensure_ascii=False, indent=2, allow_nan=False))
+
+
+if __name__ == "__main__":
+    main()
diff --git a/src/inferrsi/data/skill_library/repositories/dwarfstar/sub-skills/qualification/SKILL.md b/src/inferrsi/data/skill_library/repositories/dwarfstar/sub-skills/qualification/SKILL.md
new file mode 100644
index 000000000..12520198b
--- /dev/null
+++ b/src/inferrsi/data/skill_library/repositories/dwarfstar/sub-skills/qualification/SKILL.md
@@ -0,0 +1,85 @@
+---
+name: qualification
+description: "Qualify DwarfStar speculative decoding and distributed inference by separating sampling semantics, recurrent-state rollback, model-specific support and physical-backend evidence. Use for DSpark or MTP comparisons, exact sampling, TP failure recovery, correctness gates or performance claims."
+license: NO_LICENSE
+disable-model-invocation: true
+metadata:
+  disco-role: operating
+---
+
+# Qualify speculation and distributed execution
+
+Use this route before interpreting accepted-draft counts, local protocol tests
+or a faster decode as a qualified engine result. This is source-derived
+candidate guidance, not a backend certification or measured improvement.
+
+Read the [verification contract](references/verification-contract.md) for the
+model/mode matrix, proof ladder and experiment worksheet. Use
+[troubleshooting](references/troubleshooting.md) when a continuation diverges,
+a speculative block cannot be restored or peers fail to agree.
+
+## Inputs and output
+
+Record the engine revision/build flags, model family and exact weight hashes,
+support-model identity, backend/device/topology, context and prefill sizes,
+request concurrency, prompt/output limits, sampling parameters, seed policy,
+cache/restore state and proposed change. A project name, bit width or filename
+alone does not identify the computation.
+
+Return one bounded conclusion with its invariant, actual check, evidence class,
+counter-signal and remaining gap. Freeze numerical, distributional, quality,
+latency and resource gates in the authorized experiment, not in this skill.
+
+## Procedure
+
+1. **Choose the semantic control first.** Ordinary target-only decode, greedy
+   speculation, opportunistic nonzero-temperature speculation and exact sampled
+   speculation are different modes. Do not call default opportunistic acceptance
+   distribution-preserving. `--mtp-exact-sampling` selects the exact path where
+   implemented; it does not promise equal seeded output or equal floating-point
+   reductions. DeepSeek `--quality` or `--dspark-strict` can isolate a
+   target-only control; keep their other numerical effects visible.
+2. **Match the actual capability tuple.** Flash 0731, Flash Vision Experimental,
+   V4.1 Flash, GLM and Qwen have different draft layouts and eligibility gates.
+   A support file that loads or yields plausible acceptance is not proof that
+   it matches the target. Check the bundled model/mode matrix; preserve its
+   V4.1 network-DSpark documentation conflict rather than announcing support
+   from a generic flag or one source branch.
+3. **Verify state, not only returned tokens.** Accepted rows must commit the
+   corresponding target KV/compressor/indexer/recurrent frontier. Rejected rows
+   must not remain visible. Tool-mode boundaries, EOS, cancellation, partial
+   acceptance, restore and stale rewind need separate cases. A truncated token
+   vector is not proof that recurrent state was restored. Send session-prefix
+   and image-identity work to [serving-state](../serving-state/SKILL.md).
+4. **Keep the three distribution mechanisms separate.** In-process CUDA pairing
+   shards work across selected local devices; network TP coordinates two ranks;
+   pipeline mode assigns complete layer ranges. Their flags, capacity,
+   communication and failure contracts are not interchangeable. A local
+   socketpair test is useful protocol evidence, not physical TCP/RDMA or GPU
+   inference evidence. Send placement/budget work to
+   [runtime-memory](../runtime-memory/SKILL.md).
+5. **Climb the proof ladder without upgrading its result.** Start with the
+   selected model-free numerical/protocol checks, then actual backend kernels,
+   matched-model continuation and state tests, and finally the authorized
+   service workload. Missing required hardware, model or peer evidence remains
+   `BLOCKED_REQUIRED_BACKEND` for AREX verification and unverified/INCONCLUSIVE
+   for the affected campaign claim. A CPU check cannot clear it.
+6. **Measure net work after correctness.** Separate preparation, prefill,
+   drafter, verifier, accepted-prefix commit, replay/correction and transport
+   time. Record proposed/accepted lengths, direct/partial commits, replay
+   fallbacks, verifier errors, per-request latency and aggregate throughput.
+   Acceptance alone omits draft and verification cost; aggregate throughput is
+   not the speed of one request. Test both predictable and unpredictable inputs.
+
+## Report and stop
+
+Name the precise tested tuple and distinguish source inspection, a completed
+CPU/native check, an upstream report and an unexecuted proposal. Retain failed
+trials and invalidated sessions. A mismatched checkpoint, unsupported mode,
+unavailable accelerator/peer, missing frozen validator or exhausted authorized
+budget stops the affected qualification; it does not justify a silent fallback
+or a performance PASS.
+
+No reading of this route authorizes engine installation, model downloads,
+network reconfiguration, process termination or a new experiment. Final admission
+of a partially verified knowledge candidate is a separate informed decision.
diff --git a/src/inferrsi/data/skill_library/repositories/dwarfstar/sub-skills/qualification/references/troubleshooting.md b/src/inferrsi/data/skill_library/repositories/dwarfstar/sub-skills/qualification/references/troubleshooting.md
new file mode 100644
index 000000000..78ece5e36
--- /dev/null
+++ b/src/inferrsi/data/skill_library/repositories/dwarfstar/sub-skills/qualification/references/troubleshooting.md
@@ -0,0 +1,21 @@
+# Qualification troubleshooting
+
+| Observable symptom | Likely contract to inspect | Concrete next check/recovery | Stop boundary |
+| --- | --- | --- | --- |
+| Plausible draft acceptance but degraded continuation | Target/support checkpoint mismatch or a changed sampling mode | Record exact target/support hashes and family; compare target-only, opportunistic and exact modes separately; inspect first divergence and full quality output | Missing matched support file or quality validator; do not download/convert automatically |
+| Same seed gives different output under exact sampling | Extra accept/reject random draws; batched verifier arithmetic | Check the promised invariant: filtered distribution and state/quality, not identical RNG consumption across different algorithms | Do not redefine a frozen exact-byte gate after observing a difference |
+| High acceptance but no speedup | Draft/verifier/commit/replay cost or transport/cache pressure exceeds saved decode | Break out phase times, accepted length and replay/correction counts; compare unpredictable as well as predictable prompts under identical controls | No complete timing/quality evidence means no gain claim |
+| `speculative decode requires a synchronized checkpoint` | Token history and live graph frontier are not valid together | Invalidate/rebuild through the owning session path; retain cancellation/restore context and check the continuation after synchronization | Do not force the validity bit or truncate only the token vector |
+| Error after tool syntax begins, or repeated long-context replay | Speculative block crossed a sampling-mode boundary; recent frontier restore unavailable | Preserve trace of kept/discarded rows and resampling mode; distinguish recent exact rewind from stale rewind/rebuild | Real recurrent/kernel restore needs the actual backend; CPU assertions cannot clear it |
+| V4.1 restore initially produces no drafts | Saved target caches do not include a complete draft-feature window | Record ordinary warm-window decoding and subsequent eligibility; compare steady-state separately from restore cost | Do not invent missing draft features or call the warmup a speculation failure |
+| V4.1 network DSpark appears accepted by code but a guide says unsupported | Conflicting revision-local documentation and model-specific implementation paths | Keep the conflict explicit and request revision-specific native both-rank evidence for the exact transport/checkpoint/mode | Without that evidence the capability remains unverified, not definitely supported or definitely absent |
+| CUDA network TP rejects SSD streaming or local CUDA TP options | Mutually exclusive placement modes | Select one mechanism: local pairing, resident two-rank network TP, or pipeline; recompute its budget and workload | Do not remove the validator to make an unsupported combination start |
+| `tp hello` version/role/model mismatch | Peers disagree on protocol, complementary roles, model/layout or gate schedule | Check same source/build intent and matched artifact identities on both peers; repair the declared configuration | Hello metadata agreement is not model-byte integrity; do not bypass checks |
+| TCP succeeds, RDMA stalls or silently falls back | Device/GID/link/runtime not qualified or transport negotiated differently | Preserve both logs, negotiated transport and device/link status; use explicitly selected TCP only when the changed experiment scope is accepted | Ping/local loopback is not RDMA evidence; no automatic network or privileged host changes |
+| Peer timeout or failed verify commit | Route or mirrored frontier is invalid | Retain the failure, invalidate every affected session, restore a complete route and replay a valid saved prefix before new work | A retry that later succeeds does not erase the failed trial |
+| Parser/protocol tests pass; deployment still fails | CPU framing checks do not execute kernels, real network, model or service workload | Follow the proof ladder in the verification contract; report each passed/failed/not-run layer separately | Missing physical/native evidence remains `BLOCKED_REQUIRED_BACKEND`; candidate admission needs informed acceptance |
+
+Do not publish raw prompts, tool traces, peer addresses or user paths while
+diagnosing. Keep logs task-owned and redact before sharing without changing
+failure status or the experimental conclusion. Do not terminate or signal an
+unidentified process; physical failure injection needs its own authorized scope.
diff --git a/src/inferrsi/data/skill_library/repositories/dwarfstar/sub-skills/qualification/references/verification-contract.md b/src/inferrsi/data/skill_library/repositories/dwarfstar/sub-skills/qualification/references/verification-contract.md
new file mode 100644
index 000000000..efb75a6d6
--- /dev/null
+++ b/src/inferrsi/data/skill_library/repositories/dwarfstar/sub-skills/qualification/references/verification-contract.md
@@ -0,0 +1,156 @@
+# Verification contract
+
+## Evidence and mode identity
+
+The following contracts were distilled from the pinned DwarfStar source named
+by the root provenance. Source paths and line ranges below identify inspected
+evidence; this reference contains the operational interpretation and does not
+require those files to remain available.
+
+### Sampling semantics
+
+| Mode | What it preserves or changes | Appropriate check |
+| --- | --- | --- |
+| Ordinary target-only | Selected target graph and its sampling policy; no accepted draft block | Frozen reference for the same build/model/backend/configuration |
+| Greedy speculation | Proposed tokens are accepted when they agree with the target verifier's argmax | Check accepted-prefix/state consistency and continuation quality; batched arithmetic need not be byte-identical to serial decoding |
+| Opportunistic, nonzero temperature | Ordinary boundary tokens are sampled, while matching greedy drafts can be committed directly | Treat as a distinct sampling policy; compare quality and speed, not as proof of the target distribution |
+| Exact sampled speculation | Greedy point-mass or sampled proposals use the applicable acceptance/correction rule | Verify filtering and probability correction, then backend state and model behavior; additional random draws defeat simple seeded-identity expectations |
+
+For sampled proposals, the inspected V4.1 implementation builds draft and target
+probabilities using the selected temperature/top-k/top-p/min-p filters, accepts
+using the target-to-draft ratio and draws a correction from residual mass on
+rejection. The point-mass/greedy-proposal path has its own correction. Do not
+substitute an argmax comparison for that probability calculation.
+
+Evidence: `docs/SPECULATIVE_DECODING.md:112-137`; `ds4.c:86072-86195`;
+`tests/test_sampling.c:279-354`; `QA_BEFORE_RELEASES.md:257-309`.
+The histogram test exercises selected synthetic logits; it does not prove the
+complete model's distribution, GPU kernels or end-to-end quality. Its embedded
+tolerance is that test's contract, not a new reusable campaign gate.
+
+### Model-specific route matrix
+
+| Target | Draft source and selected constraints | Qualification caution |
+| --- | --- | --- |
+| DeepSeek V4 Flash 0731 or Vision Experimental | Separate, checkpoint-matched DSpark support GGUF; not a standalone model; not interchangeable between checkpoints | The V4 Flash external DSpark route requires a resident target; do not extend this restriction or support to V4.1 by name similarity. PRO is not a DSpark target in the inspected guide. |
+| DeepSeek V4.1 Flash | Its own support weights and draft graph; the inspected code maintains a draft-feature window and falls back to ordinary decode when that window is incomplete after restore | Source describes resident and SSD single-host routes. Network-TP DSpark documentation conflicts; retain the unresolved physical-backend qualification gap below. |
+| GLM | Built-in MTP block in the main GGUF; selected code caps its cycle and has mode-dependent fallbacks | Exact nonzero-temperature GLM sampling under active network TP falls back to a one-token evaluation in the inspected dispatcher. Do not count the fallback as speculative acceleration. |
+| Qwen3.8 Flash Next | Built-in MTP; adaptive second chained draft is separate from fixed-depth diagnostic overrides | Exact sampled behavior is model/backend/serving-mode specific. The server guide describes a Metal batched exception to a broader no-speculation statement in the speculation guide. |
+
+Evidence: `docs/SPECULATIVE_DECODING.md:13-110,139-140`;
+`ds4.c:87073-87157`; `docs/SERVER.md:86-95` for the batched Qwen exception.
+Do not promote documentation's capacity or speed examples into portable defaults.
+
+### Unresolved V4.1 network-DSpark conflict
+
+`docs/SPECULATIVE_DECODING.md:66-78` describes V4.1 DSpark on two Metal or
+CUDA hosts. `docs/DISTRIBUTED.md:89-98,139-144` instead says V4.1 speculation
+is unsupported in the described network routes. Inspected implementation contains
+a V4.1 worker verification/prefix-commit path (`ds4.c:82967-83022`) and
+model-specific speculative dispatcher, not merely a parser flag.
+
+This supports the bounded statement **implementation paths exist**, not a
+resolved support/QA guarantee. No physical native run in this extraction clears
+the conflict. Before using that tuple, obtain revision-specific support evidence
+and run both ranks, exact chosen transport and matched checkpoint with rejection,
+partial commit, cancellation and restore coverage. Until then, report the tuple
+unverified; do not choose one document as proof that it works or cannot work.
+
+## Accepted-prefix and rewind invariants
+
+1. Before speculative evaluation, session token length, valid checkpoint and
+   graph position must agree. An invalid checkpoint is an error, not an empty
+   reusable prefix.
+2. The verifier may evaluate more rows than it commits. Only the accepted prefix
+   may advance visible target state. The V4.1 worker validates commit mode and
+   count; failures invalidate its graph/checkpoint.
+3. Recent exact rewind can restore a pre-verify frontier and commit a retained
+   prefix. It also invalidates draft-capture state and restores the drafter
+   window where required. Arbitrary or stale rewinds have a different
+   invalidate/rebuild contract; truncating token history is insufficient.
+4. A verified block crossing into or out of greedy tool syntax can require
+   resampling at the boundary under exact sampling. Preserve the changed
+   sampling mode, not the previously drafted suffix.
+5. A failed peer verification or transport gate must not leave the local
+   session apparently reusable.
+
+Evidence: `ds4.c:82967-82985,86072-86195,87084-87095,87250-87289`;
+`ds4_server.c:14236-14265`. These inspected bodies do not constitute execution
+evidence for their accelerator paths.
+
+## Topology is part of the algorithm
+
+| Mechanism | Distilled setup/ownership contract | Distinct failure surface |
+| --- | --- | --- |
+| Local CUDA layer placement | Complete layers placed on selected GPUs; budget includes graph/context headroom | Unsupported CPU spill or insufficient per-device capacity |
+| In-process CUDA TP | `--cuda-tensor-parallel` pairs local devices; layer-home devices precede partners; routed experts are split while selected dense work is replicated | Wrong physical pairing, interconnect bottleneck or unsupported grouped shape |
+| Network TP | `--tensor-parallel` with complementary coordinator/worker roles; exactly two ranks in the described route | Identity/gate-schedule mismatch, transport timeout, cancellation disagreement or failed prefix commit |
+| Pipeline | Complete non-overlapping layer ranges; final range owns the output head; activations pass between stages | Missing route, overlap/gap in coverage, worker disconnect or transport-precision change |
+
+The CUDA network validator rejects combining network TP with SSD streaming or
+in-process CUDA TP. Pipeline and network-TP modes are exclusive. Do not infer
+that a flag accepted by one backend enables it on another.
+
+Network hello checks include protocol/role, model/layout attributes and gate
+schedule. These checks are not cryptographic weight-payload authentication.
+Freeze actual model bytes/hashes separately. A successful ping or local socket
+exchange does not prove an active RDMA data plane. Keep transport selection
+explicit and report the one actually negotiated. The described network
+protocols provide neither authentication nor encryption; trusted-network scope
+is a precondition, not a feature supplied by the engine.
+
+Evidence: `docs/CUDA_MULTI_GPU.md:24-68`;
+`docs/DISTRIBUTED.md:5-32,80-98,100-171,192-205`;
+`ds4_tp.c:700-738,2327-2405`.
+
+## Qualification ladder and minimal discriminating checks
+
+| Question | Smallest useful evidence | What it cannot establish |
+| --- | --- | --- |
+| Are probability filters/correction and selected RNG invariants implemented coherently? | Model-free sampling tests with synthetic logits and rejected/accepted proposals | GPU recurrence, target model quality or exact seeded parity |
+| Do command frames reject bad flags, desynchronization, malformed counts and failed acknowledgments? | Model-free socketpair protocol tests; both success and failure cases | Physical TCP/RDMA, peer GPU execution or hardware throughput |
+| Does backend rewind restore actual KV/recurrent/indexer state? | Native accelerator fixture with serial control, partial acceptance and continuation after rewind | Complete matched-model quality or concurrent service stability |
+| Does the chosen target/drafter/quant tuple behave correctly? | Matched checkpoint vectors and ordinary/opportunistic/exact model continuations under frozen policy | Another model family, backend or transport |
+| Is there a useful service gain? | Authorized end-to-end workload with request latency, concurrency, correctness/quality, memory and cleanup | A universal gain or a portable default beyond that RunSpec |
+
+Source native candidate identities are `tests/test_sampling` and
+`tests/test_tp_commands`; the inspected Makefile compiles the former with CPU
+test hooks and the latter against model-free CPU core objects. Those identities
+describe verification provenance, not instructions requiring the original
+checkout. Both remain subject to actual execution/result reporting by the
+verification owner. Physical backend/model tests remain distinct required
+evidence even if these CPU candidates pass.
+
+## Frozen experiment worksheet
+
+Before an authorized run, fill and freeze:
+
+- **Identity:** baseline/candidate revisions and patches; compiler/toolkit/driver;
+  backend, devices, peer layout and negotiated transport; target/support payload
+  hashes and quantization/layout.
+- **Workload:** exact tokenized inputs, image/tool state where applicable,
+  prefill/context/output limits, request count/concurrency, sampling/filtering,
+  thinking policy, seed/repetition policy, cold/warm/restore controls.
+- **Change and gates:** one hypothesis, declared mechanism, numerical or
+  distributional equivalence standard, quality protocol, latency/throughput and
+  memory limits, validator/input hashes, authorized resources and stop condition.
+- **Observations:** preparation/prefill/draft/verify/commit/replay/transport
+  timings; proposed/accepted lengths, full/partial commits, correction and
+  fallback counts, verifier errors; first divergence and complete quality output;
+  per-request and aggregate metrics; peak memory; both-peer logs and cleanup.
+
+For already installed binaries and already authorized local model files, a
+single-host diagnostic can compare a target-only invocation such as
+`ds4 -m <MODEL_GGUF> --ctx <TOKENS> --temp 0 -p <PROMPT>` with a matching
+candidate adding `--dspark --mtp-model <MATCHED_SUPPORT_GGUF>`. Keep all other
+controls fixed and record whether the requested path actually activated.
+This template is not a download/build instruction, a safe automatic run or a
+complete service experiment. A server benchmark needs its own request-level
+workload, not only CLI tokens per second.
+
+Source-script decisions: model download/conversion helpers are excluded from
+this route because they perform large network/storage operations and need
+separate model/license decisions. Hardware-specific server launchers are not
+portable wrappers. Long-running acceptance and benchmark scripts are
+reference-only: their controls are distilled above; executing them is not
+necessary to use this knowledge procedure and is not authorized by reading it.
diff --git a/src/inferrsi/data/skill_library/repositories/dwarfstar/sub-skills/runtime-memory/SKILL.md b/src/inferrsi/data/skill_library/repositories/dwarfstar/sub-skills/runtime-memory/SKILL.md
new file mode 100644
index 000000000..5b7054aef
--- /dev/null
+++ b/src/inferrsi/data/skill_library/repositories/dwarfstar/sub-skills/runtime-memory/SKILL.md
@@ -0,0 +1,83 @@
+---
+name: runtime-memory
+description: "Diagnose DwarfStar model fit, resident versus SSD expert placement, shared prefill workspace ownership and per-session memory without confusing cache capacity with total process memory."
+license: NO_LICENSE
+metadata:
+  disco-role: operating
+disable-model-invocation: true
+---
+
+# DwarfStar runtime and memory
+
+Use this route when a model will not fit, a larger expert cache makes decoding
+worse, extra server slots unexpectedly consume memory, or a multi-GPU placement
+is being proposed. This is source-grounded candidate guidance, not a qualified
+backend recipe. Required Metal/CUDA/ROCm inference evidence remains blocked
+until obtained on the matching backend, model and workload.
+
+## Establish the working set
+
+Read the [memory ledger](references/memory-ledger.md) before choosing a knob.
+Record these inputs, keeping unknown values explicit:
+
+- Engine revision and actual backend/build; model family, exact checkpoint,
+  GGUF tensor layout and quantization; optional vision/draft weights.
+- Physical RAM, available device memory, topology and available local storage.
+- Context limit, prefill chunk, planned slots, active slots and request lengths.
+- Resident or SSD mode, requested and effective expert-cache budget, full-layer
+  reservation, and cold/preloaded state.
+- Measured startup, prefill, decode, append/replay and post-release phases.
+
+Do not infer compatibility from a `.gguf` suffix, equal bit width, successful
+metadata inspection or the existence of an SSD flag. DwarfStar is deliberately
+specialized; unsupported checkpoints and backend/model combinations are a stop,
+not a reason to relax checks or install a different engine silently.
+
+## Procedure
+
+1. **Separate representations.** Identify file-backed model bytes, resident
+   non-routed weights, the routed-expert cache, vision/draft components and
+   per-request state. File size alone neither proves fit nor predicts bandwidth.
+   A mapped file may still generate page-cache pressure and I/O.
+2. **Separate capacities.** A `GB`-suffixed expert-cache request is a binary-GiB
+   target; a bare integer requests expert slots. Neither is a total-memory cap.
+   Record the effective startup allocation rather than copying the request into
+   a result. Include non-routed weights, graph/scratch and each slot's state.
+3. **Resolve the owner and lifetime.** Shared prefill scratch may be counted once
+   where enabled; KV, recurrent state, history and slot views remain logically
+   isolated. Do not free a shared arena with a borrowing session, or assume
+   sharing implies simultaneous unsafe access is permitted. Check capacity,
+   embedding tier, borrower lifetime and graph invalidation boundaries.
+4. **Budget the selected topology.** Metal GLM accounts for existing and planned
+   sessions when evaluating the next graph. Ordinary CUDA placement assigns
+   complete layers; in-process Flash TP splits routed experts inside device
+   pairs while replicating other work. Summed VRAM is not one interchangeable
+   allocation, and an even GPU count is not a fit proof.
+5. **Choose one discriminating change.** For context/slot pressure, reduce that
+   source of state before increasing expert cache. For expert misses with spare
+   capacity, vary the cache while fixing context, slot count and preload state.
+   For prefill I/O, compare selected-read ordering separately from next-layer
+   prefetch. Never conflate this with changed routing or weight precision.
+6. **Verify on the actual path.** First confirm startup's effective placement and
+   memory report; then compare cold prefill, continued prefill and short decode
+   separately. Preserve output/quality controls and failed allocations. CPU
+   arithmetic/bookkeeping tests cannot clear an accelerator or storage-runtime
+   gate. Do not publish an improvement from a successful launch alone.
+
+For a concrete symptom and recovery, use
+[troubleshooting](references/troubleshooting.md). Session identity, disk KV and
+replay ownership belong to the sibling serving-state route; sampling and physical
+distributed qualification belong to the sibling qualification route.
+
+## Output and stop conditions
+
+Return a phase/owner table with requested versus effective bytes, shared versus
+per-slot state, overlap, release boundary, observed versus inferred values and
+one falsifier for the proposed cause. State the activated backend path and any
+fallback. Stop at missing compatible weights, unsupported placement, insufficient
+mandatory memory, missing hardware or the authorized experiment boundary.
+
+Do not disable memory guards, change system wired-memory limits, download model
+weights, launch remote peers or treat this knowledge as execution permission.
+Absent native evidence remains `BLOCKED_REQUIRED_BACKEND`; neither the candidate
+nor an imported catalog entry is a performance or quality acceptance.
diff --git a/src/inferrsi/data/skill_library/repositories/dwarfstar/sub-skills/runtime-memory/references/memory-ledger.md b/src/inferrsi/data/skill_library/repositories/dwarfstar/sub-skills/runtime-memory/references/memory-ledger.md
new file mode 100644
index 000000000..6ab72b1db
--- /dev/null
+++ b/src/inferrsi/data/skill_library/repositories/dwarfstar/sub-skills/runtime-memory/references/memory-ledger.md
@@ -0,0 +1,124 @@
+# DwarfStar memory ledger and placement invariants
+
+## Source and evidence boundary
+
+This reference distills `antirez/ds4` source revision
+`ee7e8405232c8971000cc7ba2b945b00bf6d20d3`. The bounded readset is listed below.
+Source observations and proposed checks are not measurements. No Metal/CUDA/ROCm
+inference, real model fit, SSD latency or distributed execution was verified by
+this extraction. The source license file is MIT; the skill's `NO_LICENSE` value
+is AREX's unavailable authenticated-CLI lookup fallback, not a legal conclusion.
+
+## Representation and ownership
+
+| Item | Owner/lifetime | Budget implication | Smallest discriminating check |
+| --- | --- | --- | --- |
+| Model mapping and non-routed weights | Engine/model; backend-dependent residency | Routed-cache size excludes other weights; mapping bytes are not identical to resident physical pages | Observe startup model accounting and physical/process memory in cold and warm phases |
+| Routed expert cache | Engine/backend streaming state | Requested bytes or slots are constrained by model, graph and backend headroom | Compare requested/effective cache and actual expert I/O under a fixed prompt |
+| Prefill workspace | Shared engine-owned arena where supported; sessions borrow pointers/views | Count shared storage once, but preserve capacity/tier compatibility and scheduling rules | Change prefill chunk with slot count fixed, then slot count with chunk fixed |
+| KV/recurrent/history state | Per session, possibly views into an engine pool | A shared physical pool does not merge logical session identities | Interleave independent sessions and compare against isolated continuations |
+| Speculative, vision and host buffers | Selected model path and session | Additional weights, logits, frontier snapshots and staging can overlap other phases | Compare the identical configuration with the optional path absent, without claiming semantic equivalence |
+| File I/O/page cache and OS headroom | Host/storage, outside a tensor counter | Unified-memory expert cache and pinned staging compete with OS memory | Retain physical pressure, compressed/swap memory and storage-read observations |
+
+`ds4_engine` stores expert-cache bytes, prefill headroom, full-layer bytes,
+decode-map bytes and extra map bytes separately. Its transient guard adds
+dynamic expert-cache bytes, prefill headroom and decode-map extras using
+saturating arithmetic. The startup report additionally estimates raw/compressed
+KV and active/resident model storage. These internal counters are not a complete
+OS peak-memory measurement.
+
+## Cache controls: meaning before tuning
+
+| Control | Interpretation | Constraint |
+| --- | --- | --- |
+| `--ssd-streaming` | Stream supported routed experts; retain a bounded cache | Model/backend/tensor-layout support is specific, not universal |
+| `--ssd-streaming-cache-experts 32GB` | Request a 32-GiB expert-cache target | The `GB` suffix maps to powers of 1024; effective allocation can be reduced |
+| `--ssd-streaming-cache-experts 4000` | Request 4000 dynamic expert slots | A bare number is not a GiB value; non-routed weights and state remain additional |
+| `--ssd-streaming-full-layers N` | GLM reservation for full routed prefix layers | Trades cache distribution, not a guarantee that every other allocation fits |
+| `--ssd-streaming-cold` / `--ssd-streaming-preload-experts N` | Control preload state for selected measurements | Fix these between comparisons; a warm cache is not a cold-start result |
+
+The argument parser accepts positive decimal integer slots or positive integer
+`GB`-suffixed budgets, case-insensitively; it rejects zero, fractional values,
+negative values, other suffixes and overflow. The bare-integer GiB parser used
+elsewhere is not the expert-cache parser: do not transfer its semantics.
+
+`ds4_ssd_auto_cache_plan` computes a target from the recommended memory and its
+selected percentage, clamps it to a nonzero model limit, subtracts non-routed
+bytes where possible, then derives expert count. It enforces at least one slot
+and caps by the available model experts and integer capacity. Consequently the
+effective cache is slot-quantized and can exceed a tiny nominal remainder; this
+helper is not final proof that mandatory runtime allocations fit. Its
+`DS4_SSD_AUTO_CACHE_PCT` override accepts the implementation's bounded percentage
+range; invalid values warn and fall back. Do not turn this override into a
+generic cross-hardware tuning default.
+
+More cache is not monotonic progress. On Metal GLM, keeping non-routed weights
+resident avoids paging frequently needed data after prefill. An oversized
+routed cache can displace those weights or reduce room for sessions and scratch.
+First identify whether the failing/slow phase is cache-miss-bound or
+mandatory-state-bound.
+
+## Sharing is an ownership contract
+
+`metal_graph_transfer_prefill_workspace` moves allocation ownership to the
+destination and clears the source's ownership flag; pointer copying does not
+duplicate tensor allocations. `metal_graph_free_prefill_workspace` frees only
+the owner. A borrowed workspace is checked against the required prefill
+capacity and embedding tier. Whole-graph teardown invalidates captured decode
+graphs before releasing their baked-in buffer addresses. Preserve these
+boundaries when proposing lifetime or reuse changes.
+
+The engine also owns Qwen shared workspace and recurrent-state pools, with
+borrower/slot accounting. Sessions still hold their graph, checkpoint, image
+identities, CPU cache/scratch and speculative state. Inspect the selected path
+before asserting that every backend or model shares the same workspace.
+
+Metal GLM's next-graph budget is existing reserved session graph bytes plus
+the next graph's bytes multiplied by still-planned slots (at least one). It
+saturates on overflow. Other backends return the next-graph input unchanged in
+this helper; that is not evidence they have no separate memory policy. The
+model-free test exercises these bookkeeping distinctions, not actual GPU
+allocation or the safety of an arbitrary slot count.
+
+## CUDA read policy and topology
+
+V4.1 CUDA's selected-read path protects existing cache hits before assigning
+misses, and can order misses by file offset while preserving router/remap
+order. Its prefill default coalesces selected reads; single-token decode retains
+its original ordering unless the experimental override is present. Whole-next-
+layer prefetch is a distinct opt-in policy with its own cache-capacity and
+single-GPU prerequisites. Do not attribute a difference to routing/quantization
+when only read scheduling changed.
+
+These diagnostic switches use **presence**, not string-valued booleans:
+
+- `DS4_CUDA_ENABLE_SSD_PREFETCH` enables the V4.1 prefetch call site.
+- `DS4_CUDA_DISABLE_SSD_PREFETCH` overrides prefetch enablement.
+- `DS4_CUDA_SSD_SORT_SELECTED_READS` extends ordering to the selected
+  single-token path; `DS4_CUDA_DISABLE_SSD_SORT_SELECTED_READS` wins over it.
+- Unset the switches to recover defaults. `NAME=0` still means present.
+
+Ordinary CUDA placement assigns complete layers using per-device free-VRAM
+budgets after reserving graph/context space. Flash in-process TP instead pairs
+GPUs: routed experts are split, dense attention/router/shared experts are
+replicated within a pair, and the output head is vocabulary-sharded. Device
+ordering encodes layer-home devices followed by partners. Do not copy one
+eight-card ordering to a different interconnect, add all VRAM as fungible RAM,
+or substitute network TP flags for this mode.
+
+## Bounded source readset
+
+- `README.md:5-33`: specialized models/layouts and stated hardware targets.
+- `docs/SSD_STREAMING.md:5-93`: cache controls, headroom and CUDA switch semantics.
+- `docs/CUDA_MULTI_GPU.md:24-90`: layer placement, pair ordering and launcher limits.
+- `docs/SERVER.md:60-91`: shared prefill versus per-session state and fallback distinctions.
+- `ds4_ssd.h:7-36`, `ds4_ssd.c:16-140`: parser, byte/slot plan and arithmetic.
+- `ds4.c:17900-17962`, `19545-19579`: workspace copy/transfer/free and compatibility.
+- `ds4.c:43174-43321`, `43324-43385`: engine fields, graph budget and selected startup accounting.
+- `ds4.c:61171-61282`: selected session ownership/state fields.
+- `ds4_cuda.cu:29098-29155`, `29408-29450`: selected-cache ordering and prefetch prerequisites.
+- `ds4.c:42625-42638`: V4.1 prefetch opt-in call site.
+- `tests/test_session_state.c:163-190`: model-free memory bookkeeping assertions.
+
+The selected source paths agree on the distinctions above; this readset does
+not establish a complete support matrix or inspect every allocator/kernel.
diff --git a/src/inferrsi/data/skill_library/repositories/dwarfstar/sub-skills/runtime-memory/references/troubleshooting.md b/src/inferrsi/data/skill_library/repositories/dwarfstar/sub-skills/runtime-memory/references/troubleshooting.md
new file mode 100644
index 000000000..863ab6236
--- /dev/null
+++ b/src/inferrsi/data/skill_library/repositories/dwarfstar/sub-skills/runtime-memory/references/troubleshooting.md
@@ -0,0 +1,40 @@
+# Runtime-memory troubleshooting
+
+Use symptoms to choose a small discriminating observation, not to infer a
+universal optimal configuration. All native experiments require separate
+authorization, compatible hardware and matching model files.
+
+| Symptom | Possible cause and concrete next check | Recovery or stop |
+| --- | --- | --- |
+| A `32GB` cache request does not match the startup cache | The request is an expert-cache target; slot quantization, prefill headroom, mandatory model/graph/context allocations and backend guards constrain the effective result | Record both values and startup's breakdown. Reduce context/slots or cache only after identifying the pressured component; do not disable guards |
+| `--ssd-streaming-cache-experts 32` seems unexpectedly small | A bare decimal means 32 slots, not GiB | Use an explicit integer `GB` suffix for a byte target, then inspect the effective allocation. Reject `32G`, `1.5GB`, zero and negative values rather than silently guessing intent |
+| Larger expert cache slows generation after prefill | Non-routed weights may no longer remain resident; OS pressure or staging competes with the cache | Keep prompt/context/slots fixed, compare memory pressure and reads in append/decode, and reduce the routed cache. A changed cold/warm state invalidates the comparison |
+| One session fits but four do not | Per-slot KV/recurrent state and planned mandatory graph allocation remain even where prefill scratch is shared | Budget planned plus active slots and the actual context limit. Reduce slots or context; never extrapolate from model-file size alone |
+| Additional slot looks almost free in tensor counters | It may borrow prefill scratch or pooled views; this does not prove no private state or host allocation exists | Inspect the full phase/owner ledger and physical process memory, then test independent-session state. Retain uncertainty if allocator/OS accounting is unavailable |
+| Shared workspace reports incompatible capacity or embedding tier | Borrowed arena is smaller than the requested prefill capacity or belongs to a different embedding tier | Rebuild/reallocate through the owning engine path with compatible capacity/tier; do not alias the pointers manually or free a borrower's arena |
+| CUDA `...=0` does not turn off a diagnostic switch | The relevant SSD switches test environment-variable presence | Unset the switch. Record all related enable/disable switches; explicit disable takes precedence |
+| CUDA prefill comparison changes but decode does not | Selected-read ordering is enabled by default for the batched prefill path, not ordinary single-token reads | Measure prefill and decode separately. Treat whole-next-layer prefetch as a separate policy with its own prerequisites |
+| CUDA placement refuses despite enough summed VRAM | Whole-layer placement and replicated/shared/graph/context allocations must fit individual devices | Check the selected per-device budgets, ordering and active mode. Choose a supported smaller model or reduce state; no unsupported CPU spill assumption |
+| A model loads metadata but cannot execute or stream | Unsupported tensor layout/model/backend pairing; DwarfStar is not a generic GGUF runtime | Stop before allocating/downloading more. Reconfirm the exact project-compatible checkpoint and backend; a matching bit width is insufficient |
+
+## Measurement hygiene
+
+Keep cold load, initial prefill, continuation prefill, first-token delay and
+steady decode separate. Record preloaded experts, cache occupancy, context,
+concurrency, sampling and optional draft/vision paths. A short generation can
+expose an unusable interactive configuration but does not qualify long contexts,
+quality or service behavior. Successful CPU bookkeeping or source inspection
+does not resolve `BLOCKED_REQUIRED_BACKEND` for actual accelerator memory,
+storage overlap, model quality or throughput.
+
+## Script decisions
+
+Model download scripts are excluded: they perform large network/storage changes
+and require checkpoint-license and capacity decisions. The upstream NVIDIA TP
+launcher is reference-only because it embeds deployment-specific paths and
+policy; use the placement invariants in the bundled memory ledger, not a copied
+launcher. Benchmark/conversion tooling is outside this route's bounded scope.
+No small independent runtime-memory script was copied: the safe offline request
+builder is owned at the root, while memory-fit arithmetic without the full
+backend/model allocator would risk presenting an estimate as an allocation
+guarantee.
diff --git a/src/inferrsi/data/skill_library/repositories/dwarfstar/sub-skills/serving-state/SKILL.md b/src/inferrsi/data/skill_library/repositories/dwarfstar/sub-skills/serving-state/SKILL.md
new file mode 100644
index 000000000..e9351b473
--- /dev/null
+++ b/src/inferrsi/data/skill_library/repositories/dwarfstar/sub-skills/serving-state/SKILL.md
@@ -0,0 +1,88 @@
+---
+name: serving-state
+description: "Diagnose DwarfStar HTTP serving, session-slot reuse, image identity, disk KV compatibility and exact tool replay without confusing a visible transcript with a valid model checkpoint."
+license: NO_LICENSE
+disable-model-invocation: true
+metadata:
+  disco-role: operating
+---
+
+# DwarfStar serving and state
+
+Use for a stale answer, an unexpected full prefill, a tool-result continuation,
+an image-history change, a server restart or a concurrency change. This is
+source-derived candidate guidance, not native accelerator or service validation.
+Read [state contracts](references/state-contracts.md) for API examples, reuse
+predicates and the evidence readset; use [troubleshooting](references/troubleshooting.md)
+when an observed symptom violates an expected boundary.
+
+For offline request construction, use the bundled
+[request template](../../scripts/request_template.py); it prints JSON only and
+does not contact the server. Its `--model` value is an API alias, not a file
+selector.
+
+## Establish the request identity
+
+Record the engine revision and build backend, actual GGUF/checkpoint and optional
+vision encoder, context, slot count, prefill chunk, sampling/thinking settings,
+API, complete message history, ordered images and tool-call IDs. Record the
+cache policy and whether this is a warm live slot or a disk restore after a
+restart. Redact the content before sharing traces; retain the diagnostic
+relationship between requests.
+
+An API model alias is not a checkpoint selector. A rendered string is not an
+exact token history. An image marker is not the image fingerprint. A tool-call
+ID is not a durable remote conversation. Keep all four identities separate.
+
+## Route and inspect
+
+1. **Start local, not public.** An already-authorized, installed server may be
+   configured with `-m <MODEL_PATH> --ctx <CONTEXT>` and bound to loopback. The
+   default API address is `http://127.0.0.1:8000`. Do not broaden the bind,
+   install software, download weights or launch inference merely because this
+   procedure is being read. CORS is not authentication; a remote deployment
+   needs a separately authorized access-control/TLS design.
+2. **Identify the reuse layer.** Distinguish a live slot's valid backend
+   checkpoint, the disk entry's rendered-byte-prefix lookup and the exact
+   tokens/model state inside its payload. Apply the compatibility and suffix
+   construction rules in the reference before calling a miss a regression.
+3. **Check images before text shortcuts.** The routing probe requires a valid
+   historical fingerprint/row-count prefix. Materialization must also preserve
+   the token prefix and valid image positions. Changed pixels, reordered or
+   removed historical images, or different embedding row counts are not safe
+   extensions even when visible text is unchanged.
+4. **Check tool replay before canonicalization.** Known call IDs can recover
+   the sampled DSML spelling; all calls in a message must resolve to one
+   retained block. Missing/mixed mappings use canonical rendering instead.
+   Replacing previously sampled tokens requires a real backend-frontier
+   restore or rebuild, not an edit to the token vector alone.
+5. **Reconcile scheduling and ownership.** Each slot retains its own KV,
+   recurrent state and history. A supported model may share prefill workspace;
+   it does not share request identity. `--batched-session <SLOTS>` preallocates
+   states, and full occupancy queues requests. Ordered fallback is concurrency
+   and fairness, not proof of native grouped decode or throughput gain.
+6. **Choose the smallest discriminating check.** Compare the same request with
+   a known live prefix, after a controlled fresh session, and after an
+   authorized restart with compatible disk cache. Change only one of text,
+   image identity, tool IDs or cache policy at a time. Inspect the chosen reuse
+   route and effective prompt; compare next-token/logit or model output only
+   when the required backend and model are available.
+7. **Recover without manufacturing validity.** On an incompatible payload,
+   invalid checkpoint or unsupported rewrite, retain the error and rebuild
+   from a verified older prefix or the full request. On interruption, distinguish
+   completed synchronized chunks from an incomplete forward. Never set a
+   validity flag merely because a token count looks plausible.
+
+## Output and stopping condition
+
+Return a short ledger with `request identity`, `slot/backend validity`,
+`reuse layer`, `matched prefix`, `image/tool gate`, `suffix or rebuild`,
+`observed result` and `remaining check`. State the counterexample that would
+falsify the diagnosis. Stop before an unavailable accelerator/model, unsafe
+cache deletion, private trace publication or unapproved service exposure.
+
+CPU bookkeeping tests and source inspection can check guidance and state
+predicates. They cannot establish Metal/CUDA/ROCm rollback, physical TP restore,
+model quality, HTTP load/SLOs or performance. Those remain
+`BLOCKED_REQUIRED_BACKEND` when required resources are absent; do not upgrade
+them to a pass through synthetic assertions or green packaging CI.
diff --git a/src/inferrsi/data/skill_library/repositories/dwarfstar/sub-skills/serving-state/references/state-contracts.md b/src/inferrsi/data/skill_library/repositories/dwarfstar/sub-skills/serving-state/references/state-contracts.md
new file mode 100644
index 000000000..6293b4319
--- /dev/null
+++ b/src/inferrsi/data/skill_library/repositories/dwarfstar/sub-skills/serving-state/references/state-contracts.md
@@ -0,0 +1,167 @@
+# Serving-state contracts
+
+This is a bounded static read of DwarfStar at
+`ee7e8405232c8971000cc7ba2b945b00bf6d20d3`. The numbered source readset below
+identifies evidence, not files the reader must reopen to use this procedure.
+No accelerator, real-model session or service-performance result is implied.
+
+## API and lifecycle surface
+
+The installed `ds4-server` binary selects its backend through its build; `-m`
+selects the loaded GGUF. A Flash/PRO compatibility alias in an API request does
+not load a different model. The default listener is loopback port 8000.
+
+| Request | Purpose |
+| --- | --- |
+| `GET /v1/models` | Identify the loaded model surface |
+| `POST /v1/chat/completions` | OpenAI-style chat |
+| `POST /v1/responses` | Responses-style requests and continuations |
+| `POST /v1/completions` | Text completions |
+| `POST /v1/messages` | Anthropic-style messages |
+
+For an already-running authorized service, this is a small text-only Chat body
+that can be constructed offline before any network request:
+
+```json
+{
+  "model": "deepseek-v4-flash",
+  "messages": [{"role": "user", "content": "Explain prefix reuse briefly."}],
+  "stream": false,
+  "temperature": 0,
+  "max_tokens": 32
+}
+```
+
+Chat, Responses and Anthropic interfaces support SSE and tools. Reasoning is
+separate from visible text in each interface's own shape. At this source
+snapshot DeepSeek thinking is normally enabled; `reasoning_effort: "max"`
+selects Think Max only if context permits, whereas `"xhigh"` means normal
+thinking. Keep the effective setting fixed when investigating a cache change.
+
+Image requests require a matching loaded vision encoder. OpenAI Chat/Responses
+accept inline PNG/JPEG data URIs; Anthropic accepts base64 image sources.
+Remote URLs and server-side paths are rejected. The documented limits are 16
+images and a 64 MiB HTTP body. Preserve image order. These are this snapshot's
+interface constraints, not universal acceptance or security certification.
+
+## Three distinct cache identities
+
+| Layer | Match and ownership | Consequence of mismatch |
+| --- | --- | --- |
+| Live slot | A valid checkpoint's exact tokens plus model-specific backend state; image-conditioned states also need historical image identities | Sync resets/rebuilds rather than extending an invalid history |
+| Disk lookup | Rendered-byte-prefix hash plus model, saved-context and optional same-quant checks | A miss is legitimate even when a human sees equivalent text |
+| Disk payload | Exact saved tokens and versioned/model-specific graph state; optional tool/visible-text trailers | Load validates the payload; a byte-key hit alone cannot establish restore success |
+
+The disk search rejects a different model ID, an entry whose saved context is
+larger than the current context, entries below its minimum token policy, and
+different quantization when `--kv-cache-reject-different-quant` is set. Payload
+load additionally checks version, layout, row counts and capacity as applicable.
+Do not generalize these implementation checks into compatibility with arbitrary
+checkpoints that happen to share a model ID or bit width.
+
+After a successful byte-prefix hit, the effective prompt starts with the exact
+saved token history and tokenizes only the rendered-text suffix. Retokenizing
+the entire visible prefix can produce a different BPE spelling and detach the
+new token vector from the restored backend state. A copied cache filename or
+matching visible text is not proof of token equality or valid graph state.
+The suffix uses the rendered-chat tokenizer, not plain-text BPE, because it may
+start with role or thinking markers. Preserve that distinction in adaptations.
+
+With `--kv-disk-dir <CACHE_DIR> --kv-disk-space-mb <BUDGET_MB>`, disk cache is an
+additional persistence layer; it does not replace each slot's live state.
+Cache contents include prompts and model state. Treat the directory as private,
+retain a redacted failure record, and stop the server before any separately
+authorized clearing. Do not delete a cache just to hide a failed load.
+
+## Vision gate: identity before reuse
+
+1. A routing probe first rejects invalid checkpoints and historical image
+   fingerprint/embedding-row-count mismatches. This fingerprint-prefix probe
+   deliberately ignores positions and can admit appended request images.
+2. That is only a probe. The materialized live continuation must still satisfy
+   text/token history and span constraints. Rebased positions require an
+   independently matched continuation, such as the live tool-call binding;
+   image identity alone does not authenticate the text history.
+3. The strict session predicate checks historical token positions, row counts
+   and fingerprints. New images start at or beyond the retained token frontier.
+   Reordering, dropping or changing a historical image is not prefix extension.
+4. Vision-conditioned live state must not be written into the text-keyed disk
+   cache. After a restart, do not promise restoration of image-conditioned KV
+   merely because the visible transcript and cache directory remain.
+
+A minimal falsifier is two histories with identical visible text but one changed
+historical image fingerprint: a live reuse route must not accept the old state.
+The source's model-free identity tests express that predicate; only a real
+backend/model check can establish correct multimodal output and restoration.
+
+## Exact tool replay and backend-frontier repair
+
+For DeepSeek, generated DSML tool blocks are retained behind tool-call IDs.
+The replay map is bounded by `--tool-memory-max-ids`; a retained disk trailer may
+restore mappings across restart. Mapping restore is model-scoped. All calls
+within one message must resolve to the same retained block to attach its exact
+sampled text. IDs missing from the map, or IDs resolving to mixed blocks, use
+canonical rendering rather than fabricated exact replay.
+
+`--disable-exact-dsml-tool-replay` is a diagnostic comparison, not a remedy for
+lost model state. The live tool binding is not durable remote conversation
+storage. A tool-result-only request with no replayable prefix may fail; keep
+complete history when the client/server contract requires reconstructing it.
+
+Canonicalization must not rewrite only the token vector behind the live end.
+DeepSeek's state includes raw/compressed KV, indexer rows and compressor
+frontiers; safe replacement needs the whole corresponding frontier. The
+rewrite API reports rebuild-needed without changing the session when that
+frontier is unavailable. The server can then use a valid older disk prefix or
+full replay. It avoids text-only canonicalization for image requests to preserve
+the correctly image-conditioned sampled frontier.
+
+General rewind is also model-specific. A Qwen verify snapshot may restore
+certain recent frontiers; otherwise its recurrent state resets for replay.
+Absent a valid saved frontier, truncating DeepSeek tokens invalidates the
+checkpoint. A stale `checkpoint_valid` bit or token-count-only repair is unsafe.
+
+## Batching, fairness and distributed restoration
+
+`--batched-session <SLOTS>` preallocates independent states and queues excess
+requests; active requests are not evicted. Supported paths share prefill
+workspace, not the slot's recurrent state, KV or request history. Context and
+slot count must be budgeted together. Long prefills yield to active decoders in
+bounded intervals; `--mixed-prefill-quantum` is the diagnostic control.
+
+Execution is model/backend specific: ordered fallback runs rows separately.
+In this snapshot single-GPU CUDA, Metal GLM 5.2, and Metal SSD-streamed V4.1
+are documented examples of ordered fallback. Native grouped paths exist only
+for supported layouts. Image-conditioned V4.1 requests use ordered fallback.
+Do not label every scheduled slot a native GPU batch.
+
+Most session-batched paths use ordinary target decoding. The documented
+exception is Qwen3.8 on Metal with `--mtp`; nonzero-temperature exact-sampling
+requests use ordinary batches there. Do not stack speculative and multi-slot
+claims without checking the actual selected path.
+
+TP payload restore retains the exact saved tokens but rebuilds both ranks
+through mirrored sync; a local file cannot directly restore a remote GPU.
+Pipeline restore follows its own layer-state route. Physical peers, protocol
+failure handling and model correctness still need actual target execution.
+
+## Bounded source readset
+
+- `docs/SERVER.md`, lines 5–149: endpoints, settings, slots, images, persistence
+  and tool history.
+- `ds4.h`, lines 430–485: sync/rewrite return contracts and vision predicates.
+- `ds4.c`, lines 61171–61280: per-session state ownership;
+  64250–64367: payload checks and TP token rebuild;
+  76345–76460: prefix and image identity predicates;
+  76537–76721: sync/vision validation and failure handling;
+  76733–76845: selected Qwen/CPU prefix and interruption behavior;
+  77826–77894: rewrite/frontier contract and checked return behavior;
+  87383–87449: rewind validity.
+- `ds4_kvstore.h`, lines 11–98, and `ds4_kvstore.c`, lines 673–704 and 1196–1325:
+  byte-key, compatibility, payload and exact-token-plus-suffix construction.
+- `ds4_server.c`, lines 10440–10472, 10653–10777, 11094–11133,
+  11260–11285, 11430–11455, 11593–11670 and 12992–13075:
+  live bindings, exact replay, disk map restore, vision save guards and
+  route/materialization/canonicalization boundaries.
+- `tests/test_session_state.c`, lines 1–160: synthetic image identity and
+  invalidation assertions, inspected only; not a native GPU execution result.
diff --git a/src/inferrsi/data/skill_library/repositories/dwarfstar/sub-skills/serving-state/references/troubleshooting.md b/src/inferrsi/data/skill_library/repositories/dwarfstar/sub-skills/serving-state/references/troubleshooting.md
new file mode 100644
index 000000000..efe5ba290
--- /dev/null
+++ b/src/inferrsi/data/skill_library/repositories/dwarfstar/sub-skills/serving-state/references/troubleshooting.md
@@ -0,0 +1,43 @@
+# Serving-state troubleshooting
+
+Use these as source-derived diagnostic hypotheses. A failed hypothesis is useful
+evidence; do not change cache policy, erase files or expose a service merely to
+obtain a cleaner trace. Record the effective request and build configuration.
+
+| Symptom | Distinguish first | Smallest useful check | Recovery / stop condition |
+| --- | --- | --- | --- |
+| The requested API model name appears ignored | A compatibility alias versus the GGUF loaded at startup | Compare startup model identity and `/v1/models` with the configured `-m` | Select the intended compatible file in an authorized restart; a request alias does not hot-load it |
+| A repeated chat re-prefills after restart | Live slot versus disk prefix/payload and cache policy | Check whether a matching disk entry exists, is model/context compatible, and actually loaded | Full-history replay is valid fallback; preserve the failed-load reason rather than calling a byte-key hit a restore |
+| Equal visible text has different token counts | Sampled token spelling versus canonical BPE tokenization | Compare exact saved tokens and the suffix boundary, not only rendered text | Restore exact history plus suffix; replacing sampled tokens behind the live end needs full frontier restore/rebuild |
+| Changed image appears to reuse stale KV | Fingerprint/row count and order versus a text-only prefix | Change one historical image while keeping visible text fixed; the old image-conditioned route must miss | Retain ordered image identity and rebuild; never force reuse from matching text or placeholder IDs |
+| Appended images never reuse a live prefix | Fingerprint-only routing probe versus strict materialized spans | Verify historical fingerprints/row counts, the independent text/tool continuation, and the rebased/new-image positions | Diagnose the failed gate; do not skip token matching merely because images match |
+| Image conversation disappears from disk cache | Deliberate vision save guard versus corrupt persistence | Check whether the slot contains or is syncing vision state | Text-keyed disk persistence is not a multimodal state contract; after restart re-send complete image history and accept prefill |
+| Tool result causes a large rebuild | Missing/evicted IDs, mixed DSML blocks, or canonical token rewrite | Keep original IDs and complete history, inspect replay versus canonical route; compare a controlled diagnostic with exact replay disabled only if authorized | Rebuild a valid frontier, or resend reconstructible history; a tool ID is not durable remote state |
+| Tool-result-only request fails after restart | In-memory live binding versus retained compatible disk/tool trailers | Identify whether any full replayable prefix still exists | Return a clear missing-context limitation; do not invent a stored conversation |
+| `unsupported session payload version` or layout/capacity error | Wrong runtime/layout, saved context too large, incomplete file or incompatible model | Preserve the error; compare source/build/model/context identity and declared payload lengths | Rebuild from full history or a verified compatible older cache; never patch header/version fields to force a load |
+| `prompt length ... exceeds context ...` | Prompt-only length versus required generation room | Count the materialized prompt, including template/images/tools | Reduce history/output plan or increase context only if supported and budgeted; sync requires at least one generation token of room |
+| Interrupted request later returns stale logits | Token vector versus completed chunk's backend/logit checkpoint | Check actual interrupted status and last completed synchronized frontier | Re-sync a valid prefix; do not mark an incomplete forward valid based only on token count |
+| More slots fit but do not improve throughput | Independent states and scheduling versus native grouped decode | Record the actual model/backend path and fallback messages with fixed workload | Ordered fallback is expected in supported cases; measure before claiming gain and stop at missing GPU evidence |
+| Disk restore is unexpectedly slow on TP | Token replay across both ranks versus direct per-rank GPU restore | Compare both-rank sync activity with the retained token count | Expect prefill; physical transport and distributed correctness require an authorized real-peer check |
+
+## Safe local interface preparation
+
+Keep loopback binding unless remote access is explicitly requested. `--cors`
+only changes browser cross-origin headers; it does not supply authentication,
+change the bind address or make an Internet deployment safe. Keep cache and
+trace paths private. Public examples should use `<MODEL_PATH>`, `<CACHE_DIR>`
+and redacted request identities rather than private paths or conversation text.
+
+Build a tiny deterministic JSON request offline first, then send it only to a
+confirmed authorized local service. Do not add image URLs or server-side file
+paths when the endpoint requires inline image content. Free-form traces may
+contain prompts, tool arguments, IDs and generated text; review/redact before
+publication, retaining enough structure to explain reuse or invalidation.
+
+## Report the boundary
+
+For each diagnosis provide: observed symptom, selected identity layer, source
+invariant, alternate explanation, discriminating check, result and next safe
+step. Source reads and CPU identity tests do not prove model output, native
+rollback, device-memory safety, service SLOs or cross-rank restoration. Required
+unavailable backend/model evidence remains `BLOCKED_REQUIRED_BACKEND`.
diff --git a/src/inferrsi/data/skill_library/skill-index.json b/src/inferrsi/data/skill_library/skill-index.json
index 53c3ccefd..af8e10d50 100644
--- a/src/inferrsi/data/skill_library/skill-index.json
+++ b/src/inferrsi/data/skill_library/skill-index.json
@@ -240,6 +240,10 @@
     "reason-hopper-kernels": {
       "path": "hardware/cuda/reason-hopper-kernels",
       "backend": "cuda"
+    },
+    "dwarfstar": {
+      "path": "repositories/dwarfstar",
+      "backend": "common"
     }
   }
 }
diff --git a/tests/check_installed.py b/tests/check_installed.py
index 40c599c76..a7c9f80dc 100644
--- a/tests/check_installed.py
+++ b/tests/check_installed.py
@@ -17,7 +17,7 @@
     "This check requires an installed package"
 )
 records = knowledge_records()
-assert len(records) == 108, "Knowledge is missing from the installation"
+assert len(records) == 109, "Knowledge is missing from the installation"
 assert (
     files("inferrsi")
     .joinpath("data", "repo_packs", "vllm_omni", "knowledge", "README.md")
@@ -43,7 +43,7 @@
     ) == files("inferrsi").joinpath(*mlx_document.split("/")).read_bytes()
 assert files("inferrsi").joinpath("data/platform_packs/mlx/pack.json").is_file()
 skill_index = search_knowledge(repo="inferrsi", limit=100)
-assert len(skill_index) == 60, "Missing installed skill abstracts"
+assert len(skill_index) == 61, "Missing installed skill abstracts"
 assert all(record["index_type"] == "packaged-skill" for record in skill_index)
 indexed = KnowledgeReader().read("inferrsi.ascend.skills.ascend_operator_profiling")
 assert "data/skill_library/hardware/ascend/ascend-operator-profiling/SKILL.md" in indexed.content
@@ -143,7 +143,7 @@
         data = resource.read_bytes().replace(b"\r\n", b"\n")
         assert hashlib.sha256(data).hexdigest() == expected_hash, relative
 registry = json.loads(library.joinpath("skill-index.json").read_text(encoding="utf-8"))["skills"]
-assert len(registry) == 60
+assert len(registry) == 61
 for entry in registry.values():
     assert library.joinpath(*entry["path"].split("/"), "SKILL.md").is_file()
 router = library.joinpath("workflows", "inference-engine-research")
diff --git a/tests/fixtures/repo_to_skill/dwarfstar/reports/final/final-skill-report.md b/tests/fixtures/repo_to_skill/dwarfstar/reports/final/final-skill-report.md
new file mode 100644
index 000000000..95a271f3d
--- /dev/null
+++ b/tests/fixtures/repo_to_skill/dwarfstar/reports/final/final-skill-report.md
@@ -0,0 +1,45 @@
+# Final DwarfStar skill report
+
+## Source snapshot
+antirez/ds4 at ee7e8405232c8971000cc7ba2b945b00bf6d20d3; clean tracked source
+before local build. Native C/Make identity, not a fabricated Python package.
+AREX method pin ac3fe1afa80fb9a09775ecfb2b6cc3ba850a2db6; compatible-agent
+manual orchestration with explicitly accepted native/partial adaptation.
+
+## Generated graph and coverage
+Root dwarfstar routes runtime-memory, serving-state and qualification. Six
+focused references plus immutable source inventory, provenance, routing fragment,
+MIT notice and one offline request helper are bundled. Selected source mechanism
+coverage is substantive; full conversion/calibration, all kernels and release QA
+are explicit gaps, not a complete upstream manual.
+
+## Usability and static verification
+10 assertion-bearing cases cover all three routes, ordinary offline
+request support and two integrated difficult scenarios. Source/body/link/JSON/YAML
+checks and helper valid/invalid fixtures pass. AREX recursive license validation
+reports valid=true across four SKILL.md files. Fresh-agent independent content
+review is separately pending and is not author-prefilled.
+
+## Native ground truth
+CPU build and two CLI help checks passed. Session state, synthetic sampling and
+114/114 placement bookkeeping passed. TP command test failed its exact socket
+receive-timeout assertion (exit134); independent host probe requested50ms and
+observed52ms. The test remains NATIVE_FAIL; no upstream source was patched.
+Timing ratios printed by the sampling microtest are not engine/service claims.
+
+## Required backend gaps
+Metal/CUDA/ROCm kernels, actual model/SSD execution, GPU recurrent rollback,
+physical distributed behavior, service quality/performance and skill efficacy
+remain unverified. Two required capability groups remain
+BLOCKED_REQUIRED_BACKEND. CPU/synthetic checks do not clear them.
+
+## License resolution
+The actual AREX GitHub CLI resolver was unauthenticated, so the consistent
+runtime value is NO_LICENSE. This is a warning fallback, not a legal conclusion.
+Pinned source MIT notice and third-party attribution are retained independently.
+
+## Import readiness
+Not fully verified; not eligible for automatic import. Partial drafting was
+accepted, but final informed candidate admission after these exact gaps is
+pending. No DisCo managed import/export was executed. InferRSI packaging and
+CI evidence, independent review and merge authorization are separate records.
diff --git a/tests/fixtures/repo_to_skill/dwarfstar/reports/integration/integration-notes.md b/tests/fixtures/repo_to_skill/dwarfstar/reports/integration/integration-notes.md
new file mode 100644
index 000000000..6efd6262c
--- /dev/null
+++ b/tests/fixtures/repo_to_skill/dwarfstar/reports/integration/integration-notes.md
@@ -0,0 +1,63 @@
+# Integrated DwarfStar graph
+
+The root author inspected the actual files from all three stable-ID lanes before
+native cases were run. Lane outputs are complete; none was silently omitted.
+
+| Lane / material author | Integrated capability | Native anchor | Review decision |
+| --- | --- | --- | --- |
+| runtime-memory / Codex runtime_memory_author | Representation, SSD byte/slot parser, effective versus total budget, shared scratch/private state and CUDA topology | session-state and placement bookkeeping | Accepted: concrete controls, ownership boundaries, falsifiers and recovery; no GPU allocation claim |
+| serving-state / Codex serving_state_author | Live token/image identity, disk byte key versus exact payload, tool replay, batching/fallback and restore | session-state identity cases | Accepted: token/fingerprint/materialization gates separated; image disk-cache prohibition retained |
+| qualification / Codex qualification_author | Opportunistic/exact sampling, accepted frontier, topology/protocol and proof ladder | sampling and TP command framing | Accepted: actual V4.1 network-DSpark and Qwen batched-MTP documentation conflicts retained |
+
+All material authors use gpt-6.1-sol at xhigh; the root author is Codex
+approved_arex_author with the same configured identity. Reviewer provenance is
+recorded separately, not prefilled as approval here.
+
+## Coverage/depth and troubleshooting map
+
+- Memory inputs/representation/owner/overlap/release → runtime-memory and its two references.
+- API request/configuration, images/tools/cache identity and recovery → serving-state and its two references.
+- Model/mode compatibility, numerical/rollback correctness, physical topology and evidence classification → qualification and its two references.
+- Offline request construction → root request_template.py adapted from the source's small request example. It never sends requests.
+- Root provenance/source hashes/notices → freshness and source identity; the catalog is only a projection.
+
+Root and sibling routes use consistent lifetime and validity terminology. Local
+runtime links stay inside the DwarfStar graph. Source filenames/line spans are
+provenance, not instructions requiring a retained source checkout.
+
+## Source script import map and long-tail gaps
+
+The source-script decisions in plan.json are preserved: model-download and full
+conversion/calibration tooling are excluded; host-specific launchers and
+model/hardware benchmark harnesses are reference-only. Reusable request JSON
+construction is bundled. There is no safe independent total-memory calculator:
+a partial allocator formula would risk presenting estimates as guaranteed fit.
+
+Full GGUF conversion/imatrix, all model/kernel bodies, native agent tool side
+effects, privileged host/network setup and complete platform release QA are
+outside the bounded scope. Required accelerated/model/physical distributed
+execution remains blocked, not downgraded to optional. Actual SSD I/O, quality,
+service performance and skill efficacy have no local acceptance result.
+
+## Difficult-case reconciliation
+
+Every lane has at least two difficult synthetic cases in addition to native
+source anchors. The root adds ordinary offline-request support and two
+integrated difficult cases: cache/session/quality budgeting and image/tool
+restart state. Synthetic judgments test guidance, not unavailable native paths.
+The exact case index and assertion grades are recorded during verification.
+
+## Native environment adaptation
+
+The user explicitly accepted native C adaptation and partial drafting. A Linux
+CPU build and model-free CLI help are used instead of fictitious Python package
+metadata. GCC 14.2 and make are available; no Metal/CUDA/ROCm toolchain/device is
+available. The first build attempt used the workspace instead of the source
+working directory and failed with `No rule to make target cpu`; the corrected
+`make -C ds4 cpu -j2` completed. Both ds4 and ds4-server `--help` returned 0.
+No model download or inference was performed. The private command logs remain
+outside the published runtime graph.
+
+Final admission is manual and pending after actual verification; the partial
+drafting permission is not treated as final import/merge approval. No managed
+DisCo collection/import/export is part of this destination adaptation.
diff --git a/tests/fixtures/repo_to_skill/dwarfstar/reports/integration/plan.json b/tests/fixtures/repo_to_skill/dwarfstar/reports/integration/plan.json
new file mode 100644
index 000000000..54d10dcc7
--- /dev/null
+++ b/tests/fixtures/repo_to_skill/dwarfstar/reports/integration/plan.json
@@ -0,0 +1,33 @@
+{
+  "workflow": "AREX create-repo-skill / prepare-repo-skill-env / verify-repo-skill",
+  "method_revision": "ac3fe1afa80fb9a09775ecfb2b6cc3ba850a2db6",
+  "execution": "portable compatible-agent manual orchestration; not DisCo CLI or model API execution",
+  "source": "antirez/ds4",
+  "source_commit": "ee7e8405232c8971000cc7ba2b945b00bf6d20d3",
+  "source_state": "clean before local build; source inputs unchanged",
+  "scope_decision": "user accepted native-tool adaptation and partial drafting on 2026-10-27",
+  "included": ["README.md", "docs", "ds4.h", "ds4.c selected runtime and session bodies", "ds4_ssd.c", "ds4_kvstore.c", "ds4_server.c selected state/replay", "ds4_tp.c selected protocol contracts", "Makefile", "tests selected model-free candidates"],
+  "excluded": ["vendored quantization code bodies", "model weights and downloads", "release automation", "uninspected kernel bodies", "full GGUF conversion/imatrix tooling", "benchmark charts as local evidence"],
+  "subskills": [
+    {"id":"runtime-memory","owner":"runtime_memory_author","responsibility":"model representation, resident/SSD expert placement, shared workspace and per-session memory","evidence":["ds4_ssd.c","ds4_ssd.h","ds4.c:17900-17965","ds4.c:43174-43310","docs/SSD_STREAMING.md","docs/CUDA_MULTI_GPU.md"],"references":["memory-ledger.md","troubleshooting.md"],"native_candidates":["tests/test_session_state","tests/test_engine_mgpu_placement"],"synthetic_cases":["expert-cache-is-not-total-memory","presence-switches-and-model-specific-streaming"]},
+    {"id":"serving-state","owner":"serving_state_author","responsibility":"session prefix/vision identity, persistent KV and exact tool replay, multi-session serving","evidence":["ds4.h:430-485","ds4.c:61171-61300","ds4.c:76537-76690","ds4.c:87383-87460","ds4_kvstore.h","ds4_server.c selected replay/state paths","docs/SERVER.md","tests/test_session_state.c"],"references":["state-contracts.md","troubleshooting.md"],"native_candidates":["tests/test_session_state"],"synthetic_cases":["same-text-changed-image","disk-key-versus-token-state"]},
+    {"id":"qualification","owner":"qualification_author","responsibility":"sampling/rollback, model-specific speculative paths, physical topology and qualification ladder","evidence":["docs/SPECULATIVE_DECODING.md","docs/DISTRIBUTED.md","docs/TESTING.md","tests/test_sampling.c","tests/test_tp_commands.c","Makefile"],"references":["verification-contract.md","troubleshooting.md"],"native_candidates":["tests/test_sampling","tests/test_tp_commands"],"synthetic_cases":["opportunistic-is-not-exact-sampling","parser-test-is-not-physical-tp"]}
+  ],
+  "backend_plan": {
+    "required": [
+      {"capability":"accelerated inference and recurrent rollback","backend_requirement":"accelerator-specific","backend_criticality":"required","cpu_substitute":"partial","dependency_variant":"matching native Metal/CUDA/ROCm build and checkpoint GGUF","hardware_prerequisites":"compatible supported accelerator and model capacity","verification_expectation":"native backend state/kernel/model checks","status":"BLOCKED_REQUIRED_BACKEND"},
+      {"capability":"physical distributed inference","backend_requirement":"accelerator-specific","backend_criticality":"required","cpu_substitute":"none","dependency_variant":"same native revision on every peer and supported transport","hardware_prerequisites":"supported peer GPUs and actual TCP/RDMA topology","verification_expectation":"both-rank runtime and failure tests","status":"BLOCKED_REQUIRED_BACKEND"}
+    ],
+    "alternative": [{"capability":"model-free session/sampling bookkeeping","backend_requirement":"cpu","backend_criticality":"alternative","cpu_substitute":"full","dependency_variant":"GCC native CPU objects","hardware_prerequisites":"Linux C compiler","verification_expectation":"selected tests after whole-skill integration"}]
+  },
+  "source_script_inventory": [
+    {"source":"download_model.sh","decision":"exclude","reason":"large network downloads and checkpoint license/capacity decisions outside authority"},
+    {"source":"run-nvidia-tp-server.sh","decision":"reference-only","reason":"host-specific deployment and model/cache paths; no portable execution wrapper"},
+    {"source":"speed-bench/*","decision":"reference-only","reason":"model-backed benchmark harness, large data and hardware; distill measurement controls only"},
+    {"source":"docs/SERVER.md curl example","decision":"adapt","target":"scripts/request_template.py","reason":"safe offline request construction without network or source checkout"},
+    {"source":"gguf-tools/*","decision":"exclude","reason":"conversion/calibration is not selected extraction scope"}
+  ],
+  "integration_cases":["ssd-cache-session-budget-and-quality","image-tool-prefix-after-restart"],
+  "rubric":["canonical identity and routing","concrete source-backed invariants and falsifiers","self-contained linked resources","actionable failure handling","no implicit execution authority","blocked native evidence remains blocked"],
+  "final_candidate_admission":"pending informed decision after verification; no merge or managed import yet"
+}
diff --git a/tests/fixtures/repo_to_skill/dwarfstar/reports/license-resolution.json b/tests/fixtures/repo_to_skill/dwarfstar/reports/license-resolution.json
new file mode 100644
index 000000000..8f045aed6
--- /dev/null
+++ b/tests/fixtures/repo_to_skill/dwarfstar/reports/license-resolution.json
@@ -0,0 +1,12 @@
+{
+  "repository": "antirez/ds4",
+  "source_commit": "ee7e8405232c8971000cc7ba2b945b00bf6d20d3",
+  "source": "GitHub CLI (gh api)",
+  "value": "NO_LICENSE",
+  "status": "unavailable",
+  "reason": "GitHub CLI is not authenticated",
+  "old_license_value": "unavailable",
+  "runtime_files_updated": 4,
+  "independent_source_license": "MIT",
+  "independent_source_evidence": "Pinned source LICENSE copied with its ds4.c/ggml/DeepSeek notices; helper status unchanged."
+}
diff --git a/tests/fixtures/repo_to_skill/dwarfstar/reports/self-refine/evals.json b/tests/fixtures/repo_to_skill/dwarfstar/reports/self-refine/evals.json
new file mode 100644
index 000000000..2d36d3928
--- /dev/null
+++ b/tests/fixtures/repo_to_skill/dwarfstar/reports/self-refine/evals.json
@@ -0,0 +1,178 @@
+{
+  "skill_name": "dwarfstar",
+  "evals": [
+    {
+      "id": 1,
+      "prompt": "A DwarfStar tool conversation had an image, then I restarted the server. The visible chat and tool IDs look the same, but one old image changed. Can I force the disk KV cache to restore it and skip prefill? I need a safe diagnosis without running a GPU.",
+      "expected_output": "The answer rejects reuse of a changed historical image even when text matches.; The answer explains that vision-conditioned state is not saved into text-keyed disk KV.; The answer treats tool IDs as bounded replay mappings rather than durable remote state.; The answer preserves BLOCKED_REQUIRED_BACKEND for actual model/rollback correctness.",
+      "files": [
+        "sub-skills/serving-state/references/state-contracts.md",
+        "sub-skills/qualification/references/verification-contract.md"
+      ],
+      "assertions": [
+        "The answer rejects reuse of a changed historical image even when text matches.",
+        "The answer explains that vision-conditioned state is not saved into text-keyed disk KV.",
+        "The answer treats tool IDs as bounded replay mappings rather than durable remote state.",
+        "The answer preserves BLOCKED_REQUIRED_BACKEND for actual model/rollback correctness."
+      ]
+    },
+    {
+      "id": 2,
+      "prompt": "My DwarfStar model starts with a 32GB SSD expert cache, but four 32K server slots fail. Increasing the expert-cache request should fix it, right? I also want to enable DSpark and report faster serving based on a local CPU state test.",
+      "expected_output": "The answer distinguishes expert cache, mandatory non-routed weights, shared scratch and per-slot KV/recurrent state.; The answer rejects a monotonic more-cache solution without phase evidence.; The answer checks model-specific DSpark/SSD compatibility and sampling controls.; The answer does not turn CPU bookkeeping into accelerated/service performance evidence.",
+      "files": [
+        "sub-skills/runtime-memory/references/memory-ledger.md",
+        "sub-skills/qualification/references/verification-contract.md"
+      ],
+      "assertions": [
+        "The answer distinguishes expert cache, mandatory non-routed weights, shared scratch and per-slot KV/recurrent state.",
+        "The answer rejects a monotonic more-cache solution without phase evidence.",
+        "The answer checks model-specific DSpark/SSD compatibility and sampling controls.",
+        "The answer does not turn CPU bookkeeping into accelerated/service performance evidence."
+      ]
+    },
+    {
+      "id": 3,
+      "prompt": "I have an already-installed DwarfStar server but do not want any network requests yet. Prepare a deterministic short Chat Completions JSON body for an API alias named local-model. Does changing the alias select a new GGUF?",
+      "expected_output": "The answer uses the bundled offline request helper and does not send a request.; The answer distinguishes API alias from startup GGUF selection.; The answer keeps network/model execution under separate authorization.",
+      "files": [
+        "scripts/request_template.py",
+        "sub-skills/serving-state/references/state-contracts.md"
+      ],
+      "assertions": [
+        "The answer uses the bundled offline request helper and does not send a request.",
+        "The answer distinguishes API alias from startup GGUF selection.",
+        "The answer keeps network/model execution under separate authorization."
+      ]
+    },
+    {
+      "id": 4,
+      "prompt": "I already have an installed DwarfStar binary and a checkpoint-matched Flash 0731 target and DSpark support file. I want to compare single-host greedy decoding with and without DSpark. Give me a small experiment worksheet and command templates, keep my files untouched, and say exactly what this could and could not qualify.",
+      "expected_output": "The response provides a target-only command template and a candidate adding --dspark --mtp-model with placeholder model/support paths.; The response freezes same target/support hashes, backend/build, prompt/context/output length, greedy sampling, repetition and cache policy before running.; The response records accepted-prefix correctness, first divergence/quality, draft/verify/replay costs and actual path activation instead of ranking acceptance alone.; The response explains that batched verifier arithmetic may differ from serial greedy decoding, so the frozen correctness standard must be explicit.; The response distinguishes single-host diagnostic evidence from service latency, physical TP/RDMA, other models and backend qualification.; The response does not execute a command, download a model or claim a completed result.",
+      "files": [
+        "sub-skills/qualification/SKILL.md",
+        "sub-skills/qualification/references/verification-contract.md",
+        "sub-skills/qualification/references/troubleshooting.md"
+      ],
+      "assertions": [
+        "The response provides a target-only command template and a candidate adding --dspark --mtp-model with placeholder model/support paths.",
+        "The response freezes same target/support hashes, backend/build, prompt/context/output length, greedy sampling, repetition and cache policy before running.",
+        "The response records accepted-prefix correctness, first divergence/quality, draft/verify/replay costs and actual path activation instead of ranking acceptance alone.",
+        "The response explains that batched verifier arithmetic may differ from serial greedy decoding, so the frozen correctness standard must be explicit.",
+        "The response distinguishes single-host diagnostic evidence from service latency, physical TP/RDMA, other models and backend qualification.",
+        "The response does not execute a command, download a model or claim a completed result."
+      ]
+    },
+    {
+      "id": 5,
+      "prompt": "I enabled DSpark for a matched DwarfStar Flash target/support pair at temperature 1. The accepted drafts look good and the CPU sampling checks passed. Can I call the default mode distribution-preserving? If I add --mtp-exact-sampling, must the same seed reproduce every token of ordinary decoding? Give a correctness-first comparison plan; do not run or download anything.",
+      "expected_output": "The response routes to qualification and reads the bundled verification contract.; The response states that default nonzero-temperature opportunistic greedy-draft acceptance is not a target-distribution preservation guarantee.; The response identifies --mtp-exact-sampling as the selected exact path where implemented and describes acceptance/correction rather than only argmax agreement.; The response does not promise identical seeded token sequences across ordinary and exact speculative algorithms because additional random draws and batched arithmetic differ.; The response keeps CPU synthetic-logit sampling checks separate from actual accelerator state, matched-model quality and service verification.; The response freezes mode, checkpoint hashes, sampling filters, workload and quality/performance gates before comparison and refuses to execute or download.",
+      "files": [
+        "sub-skills/qualification/SKILL.md",
+        "sub-skills/qualification/references/verification-contract.md",
+        "sub-skills/qualification/references/troubleshooting.md"
+      ],
+      "assertions": [
+        "The response routes to qualification and reads the bundled verification contract.",
+        "The response states that default nonzero-temperature opportunistic greedy-draft acceptance is not a target-distribution preservation guarantee.",
+        "The response identifies --mtp-exact-sampling as the selected exact path where implemented and describes acceptance/correction rather than only argmax agreement.",
+        "The response does not promise identical seeded token sequences across ordinary and exact speculative algorithms because additional random draws and batched arithmetic differ.",
+        "The response keeps CPU synthetic-logit sampling checks separate from actual accelerator state, matched-model quality and service verification.",
+        "The response freezes mode, checkpoint hashes, sampling filters, workload and quality/performance gates before comparison and refuses to execute or download."
+      ]
+    },
+    {
+      "id": 6,
+      "prompt": "Our DwarfStar command tests pass using local socketpairs. I want to deploy two CUDA hosts with V4.1 DSpark, --ssd-streaming and --cuda-tensor-parallel, and label RDMA verified because the hosts ping each other. One guide says V4.1 DSpark works across two hosts and another says it does not. Review this plan without starting a service or changing the hosts.",
+      "expected_output": "The response routes to qualification and distinguishes in-process CUDA TP, network two-rank TP and pipeline placement.; The response rejects combining CUDA network TP with SSD streaming or in-process --cuda-tensor-parallel in the inspected validator.; The response states that local socketpair success and ping do not establish a physical RDMA data plane, GPU inference, or model correctness.; The response preserves the V4.1 network-DSpark source documentation conflict and notes implementation paths without asserting resolved support or absence.; The response requires exact matching peer/model/build identities, actual negotiated transport, both-rank runtime and cancellation/partial-commit/restore/failure evidence.; The response keeps missing accelerator and physical transport evidence BLOCKED_REQUIRED_BACKEND and does not recommend bypassing validators, changing host settings or starting services.",
+      "files": [
+        "sub-skills/qualification/SKILL.md",
+        "sub-skills/qualification/references/verification-contract.md",
+        "sub-skills/qualification/references/troubleshooting.md"
+      ],
+      "assertions": [
+        "The response routes to qualification and distinguishes in-process CUDA TP, network two-rank TP and pipeline placement.",
+        "The response rejects combining CUDA network TP with SSD streaming or in-process --cuda-tensor-parallel in the inspected validator.",
+        "The response states that local socketpair success and ping do not establish a physical RDMA data plane, GPU inference, or model correctness.",
+        "The response preserves the V4.1 network-DSpark source documentation conflict and notes implementation paths without asserting resolved support or absence.",
+        "The response requires exact matching peer/model/build identities, actual negotiated transport, both-rank runtime and cancellation/partial-commit/restore/failure evidence.",
+        "The response keeps missing accelerator and physical transport evidence BLOCKED_REQUIRED_BACKEND and does not recommend bypassing validators, changing host settings or starting services."
+      ]
+    },
+    {
+      "id": 7,
+      "prompt": "I have a project-compatible GLM GGUF and want four DwarfStar sessions at a longer context on a Mac. I requested --ssd-streaming-cache-experts 32GB, but startup shows a smaller effective expert cache and the fourth session is refused. One session works. Should I disable the memory guard or assume that shared prefill means four sessions cost almost the same as one? Give me a phase/owner budget and the smallest safe comparison before changing anything. I have not authorized downloads or a benchmark run.",
+      "expected_output": "The response routes the request to runtime-memory and uses the bundled memory ledger.; The response distinguishes the 32-GiB expert-cache target from total process memory and records requested versus effective values.; The response separately budgets non-routed weights, shared prefill scratch, per-session KV/recurrent state, and existing plus planned GLM sessions.; The response does not assume shared prefill merges session identities or makes additional sessions free.; The response preserves the memory guard and suggests a bounded controlled comparison without running it.; The response states that CPU bookkeeping is not accelerator allocation, quality, or service qualification.; The response neither requires the original repository docs nor downloads weights or invents measured memory values.",
+      "files": [
+        "sub-skills/runtime-memory/SKILL.md",
+        "sub-skills/runtime-memory/references/memory-ledger.md",
+        "sub-skills/runtime-memory/references/troubleshooting.md"
+      ],
+      "assertions": [
+        "The response routes the request to runtime-memory and uses the bundled memory ledger.",
+        "The response distinguishes the 32-GiB expert-cache target from total process memory and records requested versus effective values.",
+        "The response separately budgets non-routed weights, shared prefill scratch, per-session KV/recurrent state, and existing plus planned GLM sessions.",
+        "The response does not assume shared prefill merges session identities or makes additional sessions free.",
+        "The response preserves the memory guard and suggests a bounded controlled comparison without running it.",
+        "The response states that CPU bookkeeping is not accelerator allocation, quality, or service qualification.",
+        "The response neither requires the original repository docs nor downloads weights or invents measured memory values."
+      ]
+    },
+    {
+      "id": 8,
+      "prompt": "For DwarfStar V4.1 CUDA I set DS4_CUDA_ENABLE_SSD_PREFETCH=0 and DS4_CUDA_SSD_SORT_SELECTED_READS=0 to recover the default read policy. I also set DS4_CUDA_DISABLE_SSD_PREFETCH=0. My friend says 4000 and 4000GB mean the same expert cache, and that any GGUF works if its bit width matches. Before I run a comparison, explain which settings I must unset, how to record prefill versus decode separately, and what compatibility or memory checks should make me stop. Do not run commands or download anything.",
+      "expected_output": "The response routes to runtime-memory and explains that the named CUDA SSD switches use presence semantics, so values of 0 remain active.; The response unsets both enable and disable prefetch overrides and the experimental selected-read ordering override to restore the documented defaults.; The response states that explicit prefetch disable overrides enablement and sort-disable overrides experimental ordering when present.; The response distinguishes 4000 expert slots from a 4000-GiB requested target and does not call either total process memory.; The response separates prefill selected-read ordering, whole-next-layer prefetch, and ordinary single-token decode in its comparison plan.; The response requires the project-compatible checkpoint tensor layout and actual supported model/backend pairing rather than a GGUF suffix or matching bit width.; The response performs no execution, download, or unverified performance claim and retains the required native backend gap.",
+      "files": [
+        "sub-skills/runtime-memory/SKILL.md",
+        "sub-skills/runtime-memory/references/memory-ledger.md",
+        "sub-skills/runtime-memory/references/troubleshooting.md"
+      ],
+      "assertions": [
+        "The response routes to runtime-memory and explains that the named CUDA SSD switches use presence semantics, so values of 0 remain active.",
+        "The response unsets both enable and disable prefetch overrides and the experimental selected-read ordering override to restore the documented defaults.",
+        "The response states that explicit prefetch disable overrides enablement and sort-disable overrides experimental ordering when present.",
+        "The response distinguishes 4000 expert slots from a 4000-GiB requested target and does not call either total process memory.",
+        "The response separates prefill selected-read ordering, whole-next-layer prefetch, and ordinary single-token decode in its comparison plan.",
+        "The response requires the project-compatible checkpoint tensor layout and actual supported model/backend pairing rather than a GGUF suffix or matching bit width.",
+        "The response performs no execution, download, or unverified performance claim and retains the required native backend gap."
+      ]
+    },
+    {
+      "id": 9,
+      "prompt": "Our DwarfStar server restarts with the same disk KV directory. A cache filename matches the rendered conversation prefix, but retokenizing that prefix produces a different BPE token sequence from the saved payload. The new request also returns tool results. Should we replace the saved token vector with the canonical one and skip prefill? I need a recovery plan that covers model/context compatibility, missing tool mappings and a two-rank TP deployment; do not start services or delete caches.",
+      "expected_output": "The response separates rendered-byte-prefix lookup from validated exact-token/model-state payload restoration.; It preserves saved exact tokens and tokenizes only the suffix after the byte prefix, rather than retokenizing the whole prefix into the old backend state.; It checks model identity, saved context not exceeding current context, optional same-quant policy, and payload version/layout constraints without claiming arbitrary checkpoint compatibility.; It rejects token-vector-only rewrites behind the live end and calls for verified older-frontier restoration or full replay.; It explains that all calls in a message must resolve to one retained DSML block, missing/mixed mappings use canonical fallback, and tool IDs alone are not durable conversation storage.; It states that TP payload reload rebuilds both ranks from exact saved tokens and still needs real distributed-backend verification.; It neither starts a service nor deletes or patches cache data, and distinguishes the proposed recovery from an executed result.",
+      "files": [
+        "sub-skills/serving-state/SKILL.md",
+        "sub-skills/serving-state/references/state-contracts.md",
+        "sub-skills/serving-state/references/troubleshooting.md"
+      ],
+      "assertions": [
+        "The response separates rendered-byte-prefix lookup from validated exact-token/model-state payload restoration.",
+        "It preserves saved exact tokens and tokenizes only the suffix after the byte prefix, rather than retokenizing the whole prefix into the old backend state.",
+        "It checks model identity, saved context not exceeding current context, optional same-quant policy, and payload version/layout constraints without claiming arbitrary checkpoint compatibility.",
+        "It rejects token-vector-only rewrites behind the live end and calls for verified older-frontier restoration or full replay.",
+        "It explains that all calls in a message must resolve to one retained DSML block, missing/mixed mappings use canonical fallback, and tool IDs alone are not durable conversation storage.",
+        "It states that TP payload reload rebuilds both ranks from exact saved tokens and still needs real distributed-backend verification.",
+        "It neither starts a service nor deletes or patches cache data, and distinguishes the proposed recovery from an executed result."
+      ]
+    },
+    {
+      "id": 10,
+      "prompt": "I'm debugging a local DwarfStar server. A client repeats exactly the same visible conversation and tool-call IDs but replaces the first image with a different PNG. Another request appends a second image without changing the first. Can we route both requests to the same retained KV state because the text is a prefix? Explain the checks and the smallest safe diagnostic, without downloading a model or running inference.",
+      "expected_output": "The response routes this cache question to serving-state and its bundled references.; It rejects old image-conditioned reuse when a historical fingerprint or embedding row count differs, even if visible text and tool IDs match.; It distinguishes the fingerprint-only routing probe from subsequent token-history, span-position and independent continuation checks.; It treats appending an image as conditional on preserving historical identity and placing the new image at or beyond the retained frontier, not as unconditional reuse.; It proposes a one-variable image-identity diagnostic without a model download or inference execution.; It keeps required native accelerator/model verification blocked and does not describe a synthetic or CPU predicate check as equivalent evidence.",
+      "files": [
+        "sub-skills/serving-state/SKILL.md",
+        "sub-skills/serving-state/references/state-contracts.md",
+        "sub-skills/serving-state/references/troubleshooting.md"
+      ],
+      "assertions": [
+        "The response routes this cache question to serving-state and its bundled references.",
+        "It rejects old image-conditioned reuse when a historical fingerprint or embedding row count differs, even if visible text and tool IDs match.",
+        "It distinguishes the fingerprint-only routing probe from subsequent token-history, span-position and independent continuation checks.",
+        "It treats appending an image as conditional on preserving historical identity and placing the new image at or beyond the retained frontier, not as unconditional reuse.",
+        "It proposes a one-variable image-identity diagnostic without a model download or inference execution.",
+        "It keeps required native accelerator/model verification blocked and does not describe a synthetic or CPU predicate check as equivalent evidence."
+      ]
+    }
+  ]
+}
diff --git a/tests/fixtures/repo_to_skill/dwarfstar/reports/self-refine/iteration-1.md b/tests/fixtures/repo_to_skill/dwarfstar/reports/self-refine/iteration-1.md
new file mode 100644
index 000000000..3ede4a427
--- /dev/null
+++ b/tests/fixtures/repo_to_skill/dwarfstar/reports/self-refine/iteration-1.md
@@ -0,0 +1,5 @@
+# Content self-refine
+
+Root author reviewed all three complete lane outputs, every runtime reference and helper, and every assertion. Coverage, canonical IDs, role/license/invocation fields, references, source script decisions and concrete failure recovery are present. Static resource checks pass; a fresh independent agent usability attempt remains separately pending.
+
+Corrections before final verification: added native-C install/inspection boundary and provenance schema; retained source-file body limits and two conflicting model/mode support descriptions; removed an accidental unrelated numerical assertion edit before commit. No assertion judgment substitutes for unavailable backend execution.
diff --git a/tests/fixtures/repo_to_skill/dwarfstar/reports/verification/native-verification-report.json b/tests/fixtures/repo_to_skill/dwarfstar/reports/verification/native-verification-report.json
new file mode 100644
index 000000000..4bd53e759
--- /dev/null
+++ b/tests/fixtures/repo_to_skill/dwarfstar/reports/verification/native-verification-report.json
@@ -0,0 +1,75 @@
+{
+  "schema_version": 1,
+  "source_commit": "ee7e8405232c8971000cc7ba2b945b00bf6d20d3",
+  "environment": "Linux x86_64, GCC 14.2, native CPU objects; no model weights or accelerator",
+  "stage_order": "Native tests ran only after root/subskill/reference/helper integration and review.",
+  "cases": [
+    {
+      "id": "cpu-build",
+      "command": "make -C ds4 cpu -j2",
+      "status": "PASS",
+      "scope": "native CPU binaries built; does not run inference"
+    },
+    {
+      "id": "cli-help",
+      "command": "ds4 --help; ds4-server --help",
+      "status": "PASS",
+      "scope": "both exit 0 without weights"
+    },
+    {
+      "id": "session-state",
+      "command": "tests/test_session_state",
+      "status": "PASS",
+      "scope": "model-free CPU state/vision/bookkeeping assertions"
+    },
+    {
+      "id": "sampling",
+      "command": "tests/test_sampling",
+      "status": "PASS",
+      "scope": "synthetic logits/filtering/correction; emitted timing numbers are not imported as engine performance evidence"
+    },
+    {
+      "id": "tp-command-framing",
+      "command": "tests/test_tp_commands",
+      "status": "NATIVE_FAIL",
+      "exit_code": 134,
+      "reason": "Source line54 exact SO_RCVTIMEO millisecond assertion failed after bulk exchange; not patched or relabeled.",
+      "diagnostic": "Independent socketpair probe requested 50ms and observed 52000us on this Linux host, consistent with timeout granularity. This observation does not prove all test paths pass or establish an upstream fix."
+    },
+    {
+      "id": "placement-bookkeeping",
+      "command": "tests/test_engine_mgpu_placement",
+      "status": "PASS",
+      "scope": "114/114 CPU checks; no actual GPU placement execution"
+    },
+    {
+      "id": "accelerated-model-state",
+      "status": "BLOCKED_REQUIRED_BACKEND",
+      "backend_requirement": "accelerator-specific",
+      "backend_criticality": "required",
+      "cpu_substitute": "partial",
+      "dependency_variant": "matching native Metal/CUDA/ROCm and checkpoint GGUF",
+      "hardware_prerequisites": "supported accelerator and model capacity",
+      "verification_expectation": "actual kernel/model/recurrent rollback/SSD runtime cases",
+      "assigned_environment": "unavailable"
+    },
+    {
+      "id": "physical-distributed",
+      "status": "BLOCKED_REQUIRED_BACKEND",
+      "backend_requirement": "accelerator-specific",
+      "backend_criticality": "required",
+      "cpu_substitute": "none",
+      "dependency_variant": "matching peers and native transport",
+      "hardware_prerequisites": "actual supported peer GPUs/TCP or RDMA topology",
+      "verification_expectation": "both-rank model/state/failure tests",
+      "assigned_environment": "unavailable"
+    }
+  ],
+  "prior_failures": [
+    {
+      "command": "make cpu -j2 from workspace rather than source",
+      "result": "failed: no rule; corrected source -C invocation passed"
+    }
+  ],
+  "final_admission": "pending new informed manual decision after verification; no auto-import"
+}
diff --git a/tests/fixtures/repo_to_skill/dwarfstar/test-cases/index.md b/tests/fixtures/repo_to_skill/dwarfstar/test-cases/index.md
new file mode 100644
index 000000000..882ffbe8b
--- /dev/null
+++ b/tests/fixtures/repo_to_skill/dwarfstar/test-cases/index.md
@@ -0,0 +1,25 @@
+# DwarfStar assertion-backed usability cases
+
+Cases verify guidance; they do not execute or qualify GPU/model behavior.
+
+| Case | Target | Assertions | Evidence basis |
+| --- | --- | --- | --- |
+| integration/image-tool-prefix-after-restart | integration | 4 | docs/SERVER.md, docs/SSD_STREAMING.md, docs/SPECULATIVE_DECODING.md, tests/test_session_state.c |
+| integration/ssd-cache-session-budget-and-quality | integration | 4 | docs/SERVER.md, docs/SSD_STREAMING.md, docs/SPECULATIVE_DECODING.md, tests/test_session_state.c |
+| root/offline-request | dwarfstar | 3 | docs/SERVER.md, docs/SSD_STREAMING.md, docs/SPECULATIVE_DECODING.md, tests/test_session_state.c |
+| sub-skills/qualification/checkpoint-matched-control | qualification | 6 | docs/SPECULATIVE_DECODING.md:13-38,112-137, QA_BEFORE_RELEASES.md:118-124,257-309 |
+| sub-skills/qualification/opportunistic-is-not-exact-sampling | qualification | 6 | docs/SPECULATIVE_DECODING.md:112-137, tests/test_sampling.c:279-354, QA_BEFORE_RELEASES.md:257-309, ds4.c:86072-86195 |
+| sub-skills/qualification/parser-test-is-not-physical-tp | qualification | 6 | tests/test_tp_commands.c:20-153, ds4_tp.c:700-738, docs/DISTRIBUTED.md:13-32,89-106,139-144, docs/SPECULATIVE_DECODING.md:66-78, ds4.c:82967-83022 |
+| sub-skills/runtime-memory/expert-cache-is-not-total-memory | runtime-memory | 7 | docs/SSD_STREAMING.md:44-73, docs/SERVER.md:60-91, ds4.c:43279-43289, tests/test_session_state.c:163-190 |
+| sub-skills/runtime-memory/presence-switches-and-model-specific-streaming | runtime-memory | 7 | README.md:5-25, docs/SSD_STREAMING.md:44-93, ds4_ssd.c:46-69, ds4.c:42625-42638, ds4_cuda.cu:29123-29129, ds4_cuda.cu:29408-29425 |
+| sub-skills/serving-state/disk-key-versus-token-state | serving-state | 7 | ds4_kvstore.h:36-56, ds4_kvstore.c:1196-1308, ds4.c:64250-64292, ds4.c:77826-77860, ds4_server.c:10679-10731, ds4_server.c:11094-11133 |
+| sub-skills/serving-state/same-text-changed-image | serving-state | 6 | ds4.h:448-472, ds4.c:76362-76425, ds4_server.c:11593-11614, tests/test_session_state.c:6-117 |
+
+Each sub-skill has at least two new difficult synthetic cases. Qualification also
+has a source-native-anchored sampling case. Two integrated cases combine source
+state/memory/sampling contracts; these are adapted source-backed scenarios, not
+a substitute for the unavailable real-model/physical-backend tests.
+
+Assertion-bearing cases: 10. Assertions: 56.
+All included routes have source-native anchors and synthetic assertions; no model
+or accelerator capability is validated solely by these assertion grades.
diff --git a/tests/fixtures/repo_to_skill/dwarfstar/test-cases/integration/image-tool-prefix-after-restart/README.md b/tests/fixtures/repo_to_skill/dwarfstar/test-cases/integration/image-tool-prefix-after-restart/README.md
new file mode 100644
index 000000000..d54667661
--- /dev/null
+++ b/tests/fixtures/repo_to_skill/dwarfstar/test-cases/integration/image-tool-prefix-after-restart/README.md
@@ -0,0 +1,19 @@
+# Image Tool And Disk State
+
+## User Persona
+An engine operator who needs a bounded diagnosis rather than an automatic experiment.
+
+## Scenario Coverage
+- Target: integration
+- Difficulty: advanced
+- Evidence: source docs/SERVER.md, docs/SSD_STREAMING.md, docs/SPECULATIVE_DECODING.md and tests/test_session_state.c.
+- Basis: adapted source-backed cross-capability scenario; it is not a native execution result.
+
+## Expected Successful Behavior
+- The answer rejects reuse of a changed historical image even when text matches.
+- The answer explains that vision-conditioned state is not saved into text-keyed disk KV.
+- The answer treats tool IDs as bounded replay mappings rather than durable remote state.
+- The answer preserves BLOCKED_REQUIRED_BACKEND for actual model/rollback correctness.
+
+## Failure Signals
+Implicit model/network execution, forced invalid state, missing route, or a CPU check presented as GPU/service proof.
diff --git a/tests/fixtures/repo_to_skill/dwarfstar/test-cases/integration/image-tool-prefix-after-restart/assertions.json b/tests/fixtures/repo_to_skill/dwarfstar/test-cases/integration/image-tool-prefix-after-restart/assertions.json
new file mode 100644
index 000000000..95724b31c
--- /dev/null
+++ b/tests/fixtures/repo_to_skill/dwarfstar/test-cases/integration/image-tool-prefix-after-restart/assertions.json
@@ -0,0 +1,22 @@
+{
+  "schema": "disco.usability-case.v1",
+  "target_skill_area": "integration",
+  "target_capability": "image tool and disk state",
+  "difficulty": "advanced",
+  "evidence_basis": [
+    "docs/SERVER.md",
+    "docs/SSD_STREAMING.md",
+    "docs/SPECULATIVE_DECODING.md",
+    "tests/test_session_state.c"
+  ],
+  "expected_skill_files": [
+    "sub-skills/serving-state/references/state-contracts.md",
+    "sub-skills/qualification/references/verification-contract.md"
+  ],
+  "assertions": [
+    "The answer rejects reuse of a changed historical image even when text matches.",
+    "The answer explains that vision-conditioned state is not saved into text-keyed disk KV.",
+    "The answer treats tool IDs as bounded replay mappings rather than durable remote state.",
+    "The answer preserves BLOCKED_REQUIRED_BACKEND for actual model/rollback correctness."
+  ]
+}
diff --git a/tests/fixtures/repo_to_skill/dwarfstar/test-cases/integration/image-tool-prefix-after-restart/user_request.txt b/tests/fixtures/repo_to_skill/dwarfstar/test-cases/integration/image-tool-prefix-after-restart/user_request.txt
new file mode 100644
index 000000000..3c27e8cd8
--- /dev/null
+++ b/tests/fixtures/repo_to_skill/dwarfstar/test-cases/integration/image-tool-prefix-after-restart/user_request.txt
@@ -0,0 +1 @@
+A DwarfStar tool conversation had an image, then I restarted the server. The visible chat and tool IDs look the same, but one old image changed. Can I force the disk KV cache to restore it and skip prefill? I need a safe diagnosis without running a GPU.
diff --git a/tests/fixtures/repo_to_skill/dwarfstar/test-cases/integration/ssd-cache-session-budget-and-quality/README.md b/tests/fixtures/repo_to_skill/dwarfstar/test-cases/integration/ssd-cache-session-budget-and-quality/README.md
new file mode 100644
index 000000000..e8ab9f7a6
--- /dev/null
+++ b/tests/fixtures/repo_to_skill/dwarfstar/test-cases/integration/ssd-cache-session-budget-and-quality/README.md
@@ -0,0 +1,19 @@
+# Memory And Qualification
+
+## User Persona
+An engine operator who needs a bounded diagnosis rather than an automatic experiment.
+
+## Scenario Coverage
+- Target: integration
+- Difficulty: advanced
+- Evidence: source docs/SERVER.md, docs/SSD_STREAMING.md, docs/SPECULATIVE_DECODING.md and tests/test_session_state.c.
+- Basis: adapted source-backed cross-capability scenario; it is not a native execution result.
+
+## Expected Successful Behavior
+- The answer distinguishes expert cache, mandatory non-routed weights, shared scratch and per-slot KV/recurrent state.
+- The answer rejects a monotonic more-cache solution without phase evidence.
+- The answer checks model-specific DSpark/SSD compatibility and sampling controls.
+- The answer does not turn CPU bookkeeping into accelerated/service performance evidence.
+
+## Failure Signals
+Implicit model/network execution, forced invalid state, missing route, or a CPU check presented as GPU/service proof.
diff --git a/tests/fixtures/repo_to_skill/dwarfstar/test-cases/integration/ssd-cache-session-budget-and-quality/assertions.json b/tests/fixtures/repo_to_skill/dwarfstar/test-cases/integration/ssd-cache-session-budget-and-quality/assertions.json
new file mode 100644
index 000000000..47ad3b535
--- /dev/null
+++ b/tests/fixtures/repo_to_skill/dwarfstar/test-cases/integration/ssd-cache-session-budget-and-quality/assertions.json
@@ -0,0 +1,22 @@
+{
+  "schema": "disco.usability-case.v1",
+  "target_skill_area": "integration",
+  "target_capability": "memory and qualification",
+  "difficulty": "advanced",
+  "evidence_basis": [
+    "docs/SERVER.md",
+    "docs/SSD_STREAMING.md",
+    "docs/SPECULATIVE_DECODING.md",
+    "tests/test_session_state.c"
+  ],
+  "expected_skill_files": [
+    "sub-skills/runtime-memory/references/memory-ledger.md",
+    "sub-skills/qualification/references/verification-contract.md"
+  ],
+  "assertions": [
+    "The answer distinguishes expert cache, mandatory non-routed weights, shared scratch and per-slot KV/recurrent state.",
+    "The answer rejects a monotonic more-cache solution without phase evidence.",
+    "The answer checks model-specific DSpark/SSD compatibility and sampling controls.",
+    "The answer does not turn CPU bookkeeping into accelerated/service performance evidence."
+  ]
+}
diff --git a/tests/fixtures/repo_to_skill/dwarfstar/test-cases/integration/ssd-cache-session-budget-and-quality/user_request.txt b/tests/fixtures/repo_to_skill/dwarfstar/test-cases/integration/ssd-cache-session-budget-and-quality/user_request.txt
new file mode 100644
index 000000000..d57d7f48b
--- /dev/null
+++ b/tests/fixtures/repo_to_skill/dwarfstar/test-cases/integration/ssd-cache-session-budget-and-quality/user_request.txt
@@ -0,0 +1 @@
+My DwarfStar model starts with a 32GB SSD expert cache, but four 32K server slots fail. Increasing the expert-cache request should fix it, right? I also want to enable DSpark and report faster serving based on a local CPU state test.
diff --git a/tests/fixtures/repo_to_skill/dwarfstar/test-cases/root/offline-request/README.md b/tests/fixtures/repo_to_skill/dwarfstar/test-cases/root/offline-request/README.md
new file mode 100644
index 000000000..1234c1322
--- /dev/null
+++ b/tests/fixtures/repo_to_skill/dwarfstar/test-cases/root/offline-request/README.md
@@ -0,0 +1,18 @@
+# Offline Request Preparation
+
+## User Persona
+An engine operator who needs a bounded diagnosis rather than an automatic experiment.
+
+## Scenario Coverage
+- Target: dwarfstar
+- Difficulty: advanced
+- Evidence: source docs/SERVER.md, docs/SSD_STREAMING.md, docs/SPECULATIVE_DECODING.md and tests/test_session_state.c.
+- Basis: adapted source-backed cross-capability scenario; it is not a native execution result.
+
+## Expected Successful Behavior
+- The answer uses the bundled offline request helper and does not send a request.
+- The answer distinguishes API alias from startup GGUF selection.
+- The answer keeps network/model execution under separate authorization.
+
+## Failure Signals
+Implicit model/network execution, forced invalid state, missing route, or a CPU check presented as GPU/service proof.
diff --git a/tests/fixtures/repo_to_skill/dwarfstar/test-cases/root/offline-request/assertions.json b/tests/fixtures/repo_to_skill/dwarfstar/test-cases/root/offline-request/assertions.json
new file mode 100644
index 000000000..e101ea389
--- /dev/null
+++ b/tests/fixtures/repo_to_skill/dwarfstar/test-cases/root/offline-request/assertions.json
@@ -0,0 +1,21 @@
+{
+  "schema": "disco.usability-case.v1",
+  "target_skill_area": "dwarfstar",
+  "target_capability": "offline request preparation",
+  "difficulty": "advanced",
+  "evidence_basis": [
+    "docs/SERVER.md",
+    "docs/SSD_STREAMING.md",
+    "docs/SPECULATIVE_DECODING.md",
+    "tests/test_session_state.c"
+  ],
+  "expected_skill_files": [
+    "scripts/request_template.py",
+    "sub-skills/serving-state/references/state-contracts.md"
+  ],
+  "assertions": [
+    "The answer uses the bundled offline request helper and does not send a request.",
+    "The answer distinguishes API alias from startup GGUF selection.",
+    "The answer keeps network/model execution under separate authorization."
+  ]
+}
diff --git a/tests/fixtures/repo_to_skill/dwarfstar/test-cases/root/offline-request/user_request.txt b/tests/fixtures/repo_to_skill/dwarfstar/test-cases/root/offline-request/user_request.txt
new file mode 100644
index 000000000..f9835e524
--- /dev/null
+++ b/tests/fixtures/repo_to_skill/dwarfstar/test-cases/root/offline-request/user_request.txt
@@ -0,0 +1 @@
+I have an already-installed DwarfStar server but do not want any network requests yet. Prepare a deterministic short Chat Completions JSON body for an API alias named local-model. Does changing the alias select a new GGUF?
diff --git a/tests/fixtures/repo_to_skill/dwarfstar/test-cases/sub-skills/qualification/checkpoint-matched-control/README.md b/tests/fixtures/repo_to_skill/dwarfstar/test-cases/sub-skills/qualification/checkpoint-matched-control/README.md
new file mode 100644
index 000000000..5578becb9
--- /dev/null
+++ b/tests/fixtures/repo_to_skill/dwarfstar/test-cases/sub-skills/qualification/checkpoint-matched-control/README.md
@@ -0,0 +1,30 @@
+# Plan an ordinary target-only control before a DSpark comparison
+
+## User Persona
+New DwarfStar user with authorized local model files but requesting a plan only.
+
+## Scenario Coverage
+- Skill area: `qualification`
+- Capability: correctness-first single-host comparison planning
+- Difficulty: intermediate
+- Prompt file: `user_request.txt`
+- Expected references: bundled `verification-contract.md` and `troubleshooting.md`.
+- Trigger expectation: sampling or physical-topology qualification requires this route.
+
+## Expected Successful Behavior
+- The response provides a target-only command template and a candidate adding --dspark --mtp-model with placeholder model/support paths.
+- The response freezes same target/support hashes, backend/build, prompt/context/output length, greedy sampling, repetition and cache policy before running.
+- The response records accepted-prefix correctness, first divergence/quality, draft/verify/replay costs and actual path activation instead of ranking acceptance alone.
+- The response explains that batched verifier arithmetic may differ from serial greedy decoding, so the frozen correctness standard must be explicit.
+- The response distinguishes single-host diagnostic evidence from service latency, physical TP/RDMA, other models and backend qualification.
+- The response does not execute a command, download a model or claim a completed result.
+
+## Evidence and Construction
+Ordinary task anchors command generation and the experiment worksheet to the documented matched-checkpoint comparison, complementing the two difficult synthetic cases.
+This case grades guidance, not actual engine runtime or behavioral efficacy.
+
+## Failure Signals
+- Any assertion fails.
+- A response upgrades source/CPU/synthetic evidence into GPU, model, RDMA or service qualification.
+- It requires source-checkout documentation or leaks private environment details.
+- It executes outside the requested planning-only scope.
diff --git a/tests/fixtures/repo_to_skill/dwarfstar/test-cases/sub-skills/qualification/checkpoint-matched-control/assertions.json b/tests/fixtures/repo_to_skill/dwarfstar/test-cases/sub-skills/qualification/checkpoint-matched-control/assertions.json
new file mode 100644
index 000000000..3a584d1a6
--- /dev/null
+++ b/tests/fixtures/repo_to_skill/dwarfstar/test-cases/sub-skills/qualification/checkpoint-matched-control/assertions.json
@@ -0,0 +1,23 @@
+{
+  "schema": "disco.usability-case.v1",
+  "target_skill_area": "qualification",
+  "target_capability": "correctness-first single-host comparison planning",
+  "difficulty": "intermediate",
+  "evidence_basis": [
+    "docs/SPECULATIVE_DECODING.md:13-38,112-137",
+    "QA_BEFORE_RELEASES.md:118-124,257-309"
+  ],
+  "expected_skill_files": [
+    "sub-skills/qualification/SKILL.md",
+    "sub-skills/qualification/references/verification-contract.md",
+    "sub-skills/qualification/references/troubleshooting.md"
+  ],
+  "assertions": [
+    "The response provides a target-only command template and a candidate adding --dspark --mtp-model with placeholder model/support paths.",
+    "The response freezes same target/support hashes, backend/build, prompt/context/output length, greedy sampling, repetition and cache policy before running.",
+    "The response records accepted-prefix correctness, first divergence/quality, draft/verify/replay costs and actual path activation instead of ranking acceptance alone.",
+    "The response explains that batched verifier arithmetic may differ from serial greedy decoding, so the frozen correctness standard must be explicit.",
+    "The response distinguishes single-host diagnostic evidence from service latency, physical TP/RDMA, other models and backend qualification.",
+    "The response does not execute a command, download a model or claim a completed result."
+  ]
+}
diff --git a/tests/fixtures/repo_to_skill/dwarfstar/test-cases/sub-skills/qualification/checkpoint-matched-control/user_request.txt b/tests/fixtures/repo_to_skill/dwarfstar/test-cases/sub-skills/qualification/checkpoint-matched-control/user_request.txt
new file mode 100644
index 000000000..fb5d4bd85
--- /dev/null
+++ b/tests/fixtures/repo_to_skill/dwarfstar/test-cases/sub-skills/qualification/checkpoint-matched-control/user_request.txt
@@ -0,0 +1 @@
+I already have an installed DwarfStar binary and a checkpoint-matched Flash 0731 target and DSpark support file. I want to compare single-host greedy decoding with and without DSpark. Give me a small experiment worksheet and command templates, keep my files untouched, and say exactly what this could and could not qualify.
diff --git a/tests/fixtures/repo_to_skill/dwarfstar/test-cases/sub-skills/qualification/opportunistic-is-not-exact-sampling/README.md b/tests/fixtures/repo_to_skill/dwarfstar/test-cases/sub-skills/qualification/opportunistic-is-not-exact-sampling/README.md
new file mode 100644
index 000000000..a194d30aa
--- /dev/null
+++ b/tests/fixtures/repo_to_skill/dwarfstar/test-cases/sub-skills/qualification/opportunistic-is-not-exact-sampling/README.md
@@ -0,0 +1,30 @@
+# Separate distribution preservation from seeded output identity
+
+## User Persona
+Experienced inference user comparing algorithms, not yet familiar with DwarfStar sampling semantics.
+
+## Scenario Coverage
+- Skill area: `qualification`
+- Capability: opportunistic versus exact speculative sampling
+- Difficulty: advanced
+- Prompt file: `user_request.txt`
+- Expected references: bundled `verification-contract.md` and `troubleshooting.md`.
+- Trigger expectation: sampling or physical-topology qualification requires this route.
+
+## Expected Successful Behavior
+- The response routes to qualification and reads the bundled verification contract.
+- The response states that default nonzero-temperature opportunistic greedy-draft acceptance is not a target-distribution preservation guarantee.
+- The response identifies --mtp-exact-sampling as the selected exact path where implemented and describes acceptance/correction rather than only argmax agreement.
+- The response does not promise identical seeded token sequences across ordinary and exact speculative algorithms because additional random draws and batched arithmetic differ.
+- The response keeps CPU synthetic-logit sampling checks separate from actual accelerator state, matched-model quality and service verification.
+- The response freezes mode, checkpoint hashes, sampling filters, workload and quality/performance gates before comparison and refuses to execute or download.
+
+## Evidence and Construction
+Synthetic case extends the native histogram test by combining a misleading default-mode conclusion, exact-mode RNG expectations and an unavailable backend.
+This case grades guidance, not actual engine runtime or behavioral efficacy.
+
+## Failure Signals
+- Any assertion fails.
+- A response upgrades source/CPU/synthetic evidence into GPU, model, RDMA or service qualification.
+- It requires source-checkout documentation or leaks private environment details.
+- It executes outside the requested planning-only scope.
diff --git a/tests/fixtures/repo_to_skill/dwarfstar/test-cases/sub-skills/qualification/opportunistic-is-not-exact-sampling/assertions.json b/tests/fixtures/repo_to_skill/dwarfstar/test-cases/sub-skills/qualification/opportunistic-is-not-exact-sampling/assertions.json
new file mode 100644
index 000000000..358923c04
--- /dev/null
+++ b/tests/fixtures/repo_to_skill/dwarfstar/test-cases/sub-skills/qualification/opportunistic-is-not-exact-sampling/assertions.json
@@ -0,0 +1,25 @@
+{
+  "schema": "disco.usability-case.v1",
+  "target_skill_area": "qualification",
+  "target_capability": "opportunistic versus exact speculative sampling",
+  "difficulty": "advanced",
+  "evidence_basis": [
+    "docs/SPECULATIVE_DECODING.md:112-137",
+    "tests/test_sampling.c:279-354",
+    "QA_BEFORE_RELEASES.md:257-309",
+    "ds4.c:86072-86195"
+  ],
+  "expected_skill_files": [
+    "sub-skills/qualification/SKILL.md",
+    "sub-skills/qualification/references/verification-contract.md",
+    "sub-skills/qualification/references/troubleshooting.md"
+  ],
+  "assertions": [
+    "The response routes to qualification and reads the bundled verification contract.",
+    "The response states that default nonzero-temperature opportunistic greedy-draft acceptance is not a target-distribution preservation guarantee.",
+    "The response identifies --mtp-exact-sampling as the selected exact path where implemented and describes acceptance/correction rather than only argmax agreement.",
+    "The response does not promise identical seeded token sequences across ordinary and exact speculative algorithms because additional random draws and batched arithmetic differ.",
+    "The response keeps CPU synthetic-logit sampling checks separate from actual accelerator state, matched-model quality and service verification.",
+    "The response freezes mode, checkpoint hashes, sampling filters, workload and quality/performance gates before comparison and refuses to execute or download."
+  ]
+}
diff --git a/tests/fixtures/repo_to_skill/dwarfstar/test-cases/sub-skills/qualification/opportunistic-is-not-exact-sampling/user_request.txt b/tests/fixtures/repo_to_skill/dwarfstar/test-cases/sub-skills/qualification/opportunistic-is-not-exact-sampling/user_request.txt
new file mode 100644
index 000000000..5ab69da9d
--- /dev/null
+++ b/tests/fixtures/repo_to_skill/dwarfstar/test-cases/sub-skills/qualification/opportunistic-is-not-exact-sampling/user_request.txt
@@ -0,0 +1 @@
+I enabled DSpark for a matched DwarfStar Flash target/support pair at temperature 1. The accepted drafts look good and the CPU sampling checks passed. Can I call the default mode distribution-preserving? If I add --mtp-exact-sampling, must the same seed reproduce every token of ordinary decoding? Give a correctness-first comparison plan; do not run or download anything.
diff --git a/tests/fixtures/repo_to_skill/dwarfstar/test-cases/sub-skills/qualification/parser-test-is-not-physical-tp/README.md b/tests/fixtures/repo_to_skill/dwarfstar/test-cases/sub-skills/qualification/parser-test-is-not-physical-tp/README.md
new file mode 100644
index 000000000..5b44a8e58
--- /dev/null
+++ b/tests/fixtures/repo_to_skill/dwarfstar/test-cases/sub-skills/qualification/parser-test-is-not-physical-tp/README.md
@@ -0,0 +1,30 @@
+# Do not promote local protocol checks into network-TP qualification
+
+## User Persona
+Maintainer planning a heterogeneous deployment from partial CPU results.
+
+## Scenario Coverage
+- Skill area: `qualification`
+- Capability: topology/mode compatibility and physical-backend evidence
+- Difficulty: troubleshooting
+- Prompt file: `user_request.txt`
+- Expected references: bundled `verification-contract.md` and `troubleshooting.md`.
+- Trigger expectation: sampling or physical-topology qualification requires this route.
+
+## Expected Successful Behavior
+- The response routes to qualification and distinguishes in-process CUDA TP, network two-rank TP and pipeline placement.
+- The response rejects combining CUDA network TP with SSD streaming or in-process --cuda-tensor-parallel in the inspected validator.
+- The response states that local socketpair success and ping do not establish a physical RDMA data plane, GPU inference, or model correctness.
+- The response preserves the V4.1 network-DSpark source documentation conflict and notes implementation paths without asserting resolved support or absence.
+- The response requires exact matching peer/model/build identities, actual negotiated transport, both-rank runtime and cancellation/partial-commit/restore/failure evidence.
+- The response keeps missing accelerator and physical transport evidence BLOCKED_REQUIRED_BACKEND and does not recommend bypassing validators, changing host settings or starting services.
+
+## Evidence and Construction
+Synthetic combination of real validator constraints and conflicting documentation goes beyond local framing tests and catches unsafe evidence promotion.
+This case grades guidance, not actual engine runtime or behavioral efficacy.
+
+## Failure Signals
+- Any assertion fails.
+- A response upgrades source/CPU/synthetic evidence into GPU, model, RDMA or service qualification.
+- It requires source-checkout documentation or leaks private environment details.
+- It executes outside the requested planning-only scope.
diff --git a/tests/fixtures/repo_to_skill/dwarfstar/test-cases/sub-skills/qualification/parser-test-is-not-physical-tp/assertions.json b/tests/fixtures/repo_to_skill/dwarfstar/test-cases/sub-skills/qualification/parser-test-is-not-physical-tp/assertions.json
new file mode 100644
index 000000000..79580e9e9
--- /dev/null
+++ b/tests/fixtures/repo_to_skill/dwarfstar/test-cases/sub-skills/qualification/parser-test-is-not-physical-tp/assertions.json
@@ -0,0 +1,26 @@
+{
+  "schema": "disco.usability-case.v1",
+  "target_skill_area": "qualification",
+  "target_capability": "topology/mode compatibility and physical-backend evidence",
+  "difficulty": "troubleshooting",
+  "evidence_basis": [
+    "tests/test_tp_commands.c:20-153",
+    "ds4_tp.c:700-738",
+    "docs/DISTRIBUTED.md:13-32,89-106,139-144",
+    "docs/SPECULATIVE_DECODING.md:66-78",
+    "ds4.c:82967-83022"
+  ],
+  "expected_skill_files": [
+    "sub-skills/qualification/SKILL.md",
+    "sub-skills/qualification/references/verification-contract.md",
+    "sub-skills/qualification/references/troubleshooting.md"
+  ],
+  "assertions": [
+    "The response routes to qualification and distinguishes in-process CUDA TP, network two-rank TP and pipeline placement.",
+    "The response rejects combining CUDA network TP with SSD streaming or in-process --cuda-tensor-parallel in the inspected validator.",
+    "The response states that local socketpair success and ping do not establish a physical RDMA data plane, GPU inference, or model correctness.",
+    "The response preserves the V4.1 network-DSpark source documentation conflict and notes implementation paths without asserting resolved support or absence.",
+    "The response requires exact matching peer/model/build identities, actual negotiated transport, both-rank runtime and cancellation/partial-commit/restore/failure evidence.",
+    "The response keeps missing accelerator and physical transport evidence BLOCKED_REQUIRED_BACKEND and does not recommend bypassing validators, changing host settings or starting services."
+  ]
+}
diff --git a/tests/fixtures/repo_to_skill/dwarfstar/test-cases/sub-skills/qualification/parser-test-is-not-physical-tp/user_request.txt b/tests/fixtures/repo_to_skill/dwarfstar/test-cases/sub-skills/qualification/parser-test-is-not-physical-tp/user_request.txt
new file mode 100644
index 000000000..a01a6b748
--- /dev/null
+++ b/tests/fixtures/repo_to_skill/dwarfstar/test-cases/sub-skills/qualification/parser-test-is-not-physical-tp/user_request.txt
@@ -0,0 +1 @@
+Our DwarfStar command tests pass using local socketpairs. I want to deploy two CUDA hosts with V4.1 DSpark, --ssd-streaming and --cuda-tensor-parallel, and label RDMA verified because the hosts ping each other. One guide says V4.1 DSpark works across two hosts and another says it does not. Review this plan without starting a service or changing the hosts.
diff --git a/tests/fixtures/repo_to_skill/dwarfstar/test-cases/sub-skills/runtime-memory/expert-cache-is-not-total-memory/README.md b/tests/fixtures/repo_to_skill/dwarfstar/test-cases/sub-skills/runtime-memory/expert-cache-is-not-total-memory/README.md
new file mode 100644
index 000000000..30dd33976
--- /dev/null
+++ b/tests/fixtures/repo_to_skill/dwarfstar/test-cases/sub-skills/runtime-memory/expert-cache-is-not-total-memory/README.md
@@ -0,0 +1,30 @@
+# Expert-cache budget is not total memory
+
+## User Persona
+An operator familiar with the command line but not the engine's ownership and
+planned-session bookkeeping.
+
+## Scenario Coverage
+- Skill area: runtime-memory.
+- Capability: SSD model fit, shared prefill and per-session capacity.
+- Difficulty: troubleshooting.
+- Prompt file: `user_request.txt`.
+- Expected skill resources: the runtime-memory skill and its memory-ledger and
+  troubleshooting references.
+- Trigger expectation: direct mention of SSD cache, multiple slots and a
+  failed allocation should select runtime-memory.
+- Synthetic extension: combines the documented cache controls with the
+  model-free `test_session_memory` assertions; it is not a report of a real Mac
+  execution and extends those assertions into an operational diagnosis.
+
+## Expected Successful Behavior
+Distinguish requested and effective expert cache, count non-routed weights,
+shared prefill storage, private KV/recurrent state and planned slots, preserve
+the guard, and propose one controlled context/slot/cache comparison without
+executing it. Keep observed values separate from missing physical-memory data.
+
+## Failure Signals
+Calling 32GB a process cap, asserting all slots share KV, disabling the guard,
+inventing a safe slot count, launching tests/downloads, reopening original source
+docs instead of using the bundled references, or clearing accelerator evidence
+with CPU bookkeeping.
diff --git a/tests/fixtures/repo_to_skill/dwarfstar/test-cases/sub-skills/runtime-memory/expert-cache-is-not-total-memory/assertions.json b/tests/fixtures/repo_to_skill/dwarfstar/test-cases/sub-skills/runtime-memory/expert-cache-is-not-total-memory/assertions.json
new file mode 100644
index 000000000..eec8579ae
--- /dev/null
+++ b/tests/fixtures/repo_to_skill/dwarfstar/test-cases/sub-skills/runtime-memory/expert-cache-is-not-total-memory/assertions.json
@@ -0,0 +1,26 @@
+{
+  "schema": "disco.usability-case.v1",
+  "target_skill_area": "runtime-memory",
+  "target_capability": "SSD cache and multi-session memory ownership",
+  "difficulty": "troubleshooting",
+  "evidence_basis": [
+    "docs/SSD_STREAMING.md:44-73",
+    "docs/SERVER.md:60-91",
+    "ds4.c:43279-43289",
+    "tests/test_session_state.c:163-190"
+  ],
+  "expected_skill_files": [
+    "sub-skills/runtime-memory/SKILL.md",
+    "sub-skills/runtime-memory/references/memory-ledger.md",
+    "sub-skills/runtime-memory/references/troubleshooting.md"
+  ],
+  "assertions": [
+    "The response routes the request to runtime-memory and uses the bundled memory ledger.",
+    "The response distinguishes the 32-GiB expert-cache target from total process memory and records requested versus effective values.",
+    "The response separately budgets non-routed weights, shared prefill scratch, per-session KV/recurrent state, and existing plus planned GLM sessions.",
+    "The response does not assume shared prefill merges session identities or makes additional sessions free.",
+    "The response preserves the memory guard and suggests a bounded controlled comparison without running it.",
+    "The response states that CPU bookkeeping is not accelerator allocation, quality, or service qualification.",
+    "The response neither requires the original repository docs nor downloads weights or invents measured memory values."
+  ]
+}
diff --git a/tests/fixtures/repo_to_skill/dwarfstar/test-cases/sub-skills/runtime-memory/expert-cache-is-not-total-memory/user_request.txt b/tests/fixtures/repo_to_skill/dwarfstar/test-cases/sub-skills/runtime-memory/expert-cache-is-not-total-memory/user_request.txt
new file mode 100644
index 000000000..bad0b6eeb
--- /dev/null
+++ b/tests/fixtures/repo_to_skill/dwarfstar/test-cases/sub-skills/runtime-memory/expert-cache-is-not-total-memory/user_request.txt
@@ -0,0 +1 @@
+I have a project-compatible GLM GGUF and want four DwarfStar sessions at a longer context on a Mac. I requested --ssd-streaming-cache-experts 32GB, but startup shows a smaller effective expert cache and the fourth session is refused. One session works. Should I disable the memory guard or assume that shared prefill means four sessions cost almost the same as one? Give me a phase/owner budget and the smallest safe comparison before changing anything. I have not authorized downloads or a benchmark run.
diff --git a/tests/fixtures/repo_to_skill/dwarfstar/test-cases/sub-skills/runtime-memory/presence-switches-and-model-specific-streaming/README.md b/tests/fixtures/repo_to_skill/dwarfstar/test-cases/sub-skills/runtime-memory/presence-switches-and-model-specific-streaming/README.md
new file mode 100644
index 000000000..d4c78f3a0
--- /dev/null
+++ b/tests/fixtures/repo_to_skill/dwarfstar/test-cases/sub-skills/runtime-memory/presence-switches-and-model-specific-streaming/README.md
@@ -0,0 +1,28 @@
+# Presence switches and model-specific streaming
+
+## User Persona
+An experienced CUDA operator transferring familiar boolean environment settings
+and generic GGUF assumptions into a specialized runtime.
+
+## Scenario Coverage
+- Skill area: runtime-memory.
+- Capability: expert-cache units, read-policy controls and compatibility triage.
+- Difficulty: advanced.
+- Prompt file: `user_request.txt`.
+- Expected skill resources: runtime-memory, memory-ledger and troubleshooting.
+- Trigger expectation: CUDA SSD read-policy comparison belongs to this route.
+- Synthetic extension: combines parser and CUDA call-site facts into a
+  contradictory configuration that a simple native parser test does not cover.
+  This case tests advice only, not CUDA execution.
+
+## Expected Successful Behavior
+Explain presence semantics and disable precedence, unset all relevant policy
+overrides to restore defaults, distinguish bare slots from integer-GB budgets,
+separate selected-read prefill ordering from next-layer prefetch and ordinary
+decode, and reject an unsupported GGUF/model/backend tuple. Propose a frozen
+comparison record, not an unrequested run.
+
+## Failure Signals
+Treating `=0` as absent, suggesting only the enable switch needs unsetting while
+leaving the disable override present, accepting arbitrary GGUF formats, equating
+slots with GiB, promising speedups, or calling source inspection GPU validation.
diff --git a/tests/fixtures/repo_to_skill/dwarfstar/test-cases/sub-skills/runtime-memory/presence-switches-and-model-specific-streaming/assertions.json b/tests/fixtures/repo_to_skill/dwarfstar/test-cases/sub-skills/runtime-memory/presence-switches-and-model-specific-streaming/assertions.json
new file mode 100644
index 000000000..ea3fff1ca
--- /dev/null
+++ b/tests/fixtures/repo_to_skill/dwarfstar/test-cases/sub-skills/runtime-memory/presence-switches-and-model-specific-streaming/assertions.json
@@ -0,0 +1,28 @@
+{
+  "schema": "disco.usability-case.v1",
+  "target_skill_area": "runtime-memory",
+  "target_capability": "CUDA SSD policy controls and checkpoint compatibility",
+  "difficulty": "advanced",
+  "evidence_basis": [
+    "README.md:5-25",
+    "docs/SSD_STREAMING.md:44-93",
+    "ds4_ssd.c:46-69",
+    "ds4.c:42625-42638",
+    "ds4_cuda.cu:29123-29129",
+    "ds4_cuda.cu:29408-29425"
+  ],
+  "expected_skill_files": [
+    "sub-skills/runtime-memory/SKILL.md",
+    "sub-skills/runtime-memory/references/memory-ledger.md",
+    "sub-skills/runtime-memory/references/troubleshooting.md"
+  ],
+  "assertions": [
+    "The response routes to runtime-memory and explains that the named CUDA SSD switches use presence semantics, so values of 0 remain active.",
+    "The response unsets both enable and disable prefetch overrides and the experimental selected-read ordering override to restore the documented defaults.",
+    "The response states that explicit prefetch disable overrides enablement and sort-disable overrides experimental ordering when present.",
+    "The response distinguishes 4000 expert slots from a 4000-GiB requested target and does not call either total process memory.",
+    "The response separates prefill selected-read ordering, whole-next-layer prefetch, and ordinary single-token decode in its comparison plan.",
+    "The response requires the project-compatible checkpoint tensor layout and actual supported model/backend pairing rather than a GGUF suffix or matching bit width.",
+    "The response performs no execution, download, or unverified performance claim and retains the required native backend gap."
+  ]
+}
diff --git a/tests/fixtures/repo_to_skill/dwarfstar/test-cases/sub-skills/runtime-memory/presence-switches-and-model-specific-streaming/user_request.txt b/tests/fixtures/repo_to_skill/dwarfstar/test-cases/sub-skills/runtime-memory/presence-switches-and-model-specific-streaming/user_request.txt
new file mode 100644
index 000000000..dd4605320
--- /dev/null
+++ b/tests/fixtures/repo_to_skill/dwarfstar/test-cases/sub-skills/runtime-memory/presence-switches-and-model-specific-streaming/user_request.txt
@@ -0,0 +1 @@
+For DwarfStar V4.1 CUDA I set DS4_CUDA_ENABLE_SSD_PREFETCH=0 and DS4_CUDA_SSD_SORT_SELECTED_READS=0 to recover the default read policy. I also set DS4_CUDA_DISABLE_SSD_PREFETCH=0. My friend says 4000 and 4000GB mean the same expert cache, and that any GGUF works if its bit width matches. Before I run a comparison, explain which settings I must unset, how to record prefill versus decode separately, and what compatibility or memory checks should make me stop. Do not run commands or download anything.
diff --git a/tests/fixtures/repo_to_skill/dwarfstar/test-cases/sub-skills/serving-state/disk-key-versus-token-state/README.md b/tests/fixtures/repo_to_skill/dwarfstar/test-cases/sub-skills/serving-state/disk-key-versus-token-state/README.md
new file mode 100644
index 000000000..45604bb8a
--- /dev/null
+++ b/tests/fixtures/repo_to_skill/dwarfstar/test-cases/sub-skills/serving-state/disk-key-versus-token-state/README.md
@@ -0,0 +1,19 @@
+# Disk lookup is not backend state
+
+## User Persona
+A maintainer diagnosing a post-restart tool-continuation rebuild.
+
+## Scenario Coverage
+- Skill area: `serving-state`.
+- Capability: rendered-byte keys, exact-token payload, tool replay and distributed restore.
+- Difficulty: advanced.
+- Prompt file: `user_request.txt`.
+- Expected resources: `sub-skills/serving-state/SKILL.md` and bundled state-contracts/troubleshooting references.
+- Trigger expectation: the prompt combines persistent DwarfStar KV with a concrete BPE mismatch and TP deployment.
+- Synthetic extension: combines independently sourced byte-prefix loading, canonical-tail safety and TP rebuild behavior; source native candidates do not by themselves verify this integrated deployment.
+
+## Expected Successful Behavior
+Treat the filename match as a lookup candidate, not payload validation. Preserve the exact saved token history and tokenize only the new text suffix on a successful compatible load. Check model ID, saved context versus current capacity, optional same-quant policy and payload version/layout. Reject token-vector-only replacement behind the live frontier. Explain missing/mixed tool-map fallback and that a live ID binding is not durable remote conversation storage. TP reload rebuilds both ranks from exact tokens; it does not directly restore a second GPU. Preserve failures and avoid service mutation/cache deletion.
+
+## Failure Signals
+Equal bytes are assumed to prove token/backend equality; canonical tokens are spliced into old KV; larger saved context is forced into a smaller runtime; tool IDs are assumed durable; a local file is claimed to restore all ranks instantly; a proposed recovery is mislabeled as an executed native result.
diff --git a/tests/fixtures/repo_to_skill/dwarfstar/test-cases/sub-skills/serving-state/disk-key-versus-token-state/assertions.json b/tests/fixtures/repo_to_skill/dwarfstar/test-cases/sub-skills/serving-state/disk-key-versus-token-state/assertions.json
new file mode 100644
index 000000000..6853496ac
--- /dev/null
+++ b/tests/fixtures/repo_to_skill/dwarfstar/test-cases/sub-skills/serving-state/disk-key-versus-token-state/assertions.json
@@ -0,0 +1,28 @@
+{
+  "schema": "disco.usability-case.v1",
+  "target_skill_area": "serving-state",
+  "target_capability": "persistent prefix and exact backend-frontier recovery",
+  "difficulty": "advanced",
+  "evidence_basis": [
+    "ds4_kvstore.h:36-56",
+    "ds4_kvstore.c:1196-1308",
+    "ds4.c:64250-64292",
+    "ds4.c:77826-77860",
+    "ds4_server.c:10679-10731",
+    "ds4_server.c:11094-11133"
+  ],
+  "expected_skill_files": [
+    "sub-skills/serving-state/SKILL.md",
+    "sub-skills/serving-state/references/state-contracts.md",
+    "sub-skills/serving-state/references/troubleshooting.md"
+  ],
+  "assertions": [
+    "The response separates rendered-byte-prefix lookup from validated exact-token/model-state payload restoration.",
+    "It preserves saved exact tokens and tokenizes only the suffix after the byte prefix, rather than retokenizing the whole prefix into the old backend state.",
+    "It checks model identity, saved context not exceeding current context, optional same-quant policy, and payload version/layout constraints without claiming arbitrary checkpoint compatibility.",
+    "It rejects token-vector-only rewrites behind the live end and calls for verified older-frontier restoration or full replay.",
+    "It explains that all calls in a message must resolve to one retained DSML block, missing/mixed mappings use canonical fallback, and tool IDs alone are not durable conversation storage.",
+    "It states that TP payload reload rebuilds both ranks from exact saved tokens and still needs real distributed-backend verification.",
+    "It neither starts a service nor deletes or patches cache data, and distinguishes the proposed recovery from an executed result."
+  ]
+}
diff --git a/tests/fixtures/repo_to_skill/dwarfstar/test-cases/sub-skills/serving-state/disk-key-versus-token-state/user_request.txt b/tests/fixtures/repo_to_skill/dwarfstar/test-cases/sub-skills/serving-state/disk-key-versus-token-state/user_request.txt
new file mode 100644
index 000000000..490eeaa44
--- /dev/null
+++ b/tests/fixtures/repo_to_skill/dwarfstar/test-cases/sub-skills/serving-state/disk-key-versus-token-state/user_request.txt
@@ -0,0 +1 @@
+Our DwarfStar server restarts with the same disk KV directory. A cache filename matches the rendered conversation prefix, but retokenizing that prefix produces a different BPE token sequence from the saved payload. The new request also returns tool results. Should we replace the saved token vector with the canonical one and skip prefill? I need a recovery plan that covers model/context compatibility, missing tool mappings and a two-rank TP deployment; do not start services or delete caches.
diff --git a/tests/fixtures/repo_to_skill/dwarfstar/test-cases/sub-skills/serving-state/same-text-changed-image/README.md b/tests/fixtures/repo_to_skill/dwarfstar/test-cases/sub-skills/serving-state/same-text-changed-image/README.md
new file mode 100644
index 000000000..3746af0c5
--- /dev/null
+++ b/tests/fixtures/repo_to_skill/dwarfstar/test-cases/sub-skills/serving-state/same-text-changed-image/README.md
@@ -0,0 +1,19 @@
+# Identical text, changed image
+
+## User Persona
+An experienced application integrator investigating a stale multimodal cache.
+
+## Scenario Coverage
+- Skill area: `serving-state`.
+- Capability: historical image identity gates, append semantics and route/materialization agreement.
+- Difficulty: troubleshooting.
+- Prompt file: `user_request.txt`.
+- Expected resources: `sub-skills/serving-state/SKILL.md` and its state-contracts/troubleshooting references.
+- Trigger expectation: the user names DwarfStar KV reuse and a concrete image/tool-history interaction.
+- Synthetic extension: combines the source's image-predicate assertions with the server's independently matched tool continuation; not a claim of native execution.
+
+## Expected Successful Behavior
+Reject stale reuse after changing a historical image fingerprint or embedding row count even when visible text and call IDs match. Distinguish fingerprint-only routing from strict token/span validation. Permit consideration of an appended image only after historical identity, independent text/tool history and new-image position constraints hold. Do not promise that an image-prefix probe alone proves reuse. Propose a one-variable identity comparison and preserve the missing model/backend boundary.
+
+## Failure Signals
+Text-only cache keys, equal tool IDs or matching image counts are treated as sufficient; historical image ordering is ignored; a pure predicate test is described as GPU/model verification; the response downloads weights or runs inference despite the constraint.
diff --git a/tests/fixtures/repo_to_skill/dwarfstar/test-cases/sub-skills/serving-state/same-text-changed-image/assertions.json b/tests/fixtures/repo_to_skill/dwarfstar/test-cases/sub-skills/serving-state/same-text-changed-image/assertions.json
new file mode 100644
index 000000000..3399547f2
--- /dev/null
+++ b/tests/fixtures/repo_to_skill/dwarfstar/test-cases/sub-skills/serving-state/same-text-changed-image/assertions.json
@@ -0,0 +1,25 @@
+{
+  "schema": "disco.usability-case.v1",
+  "target_skill_area": "serving-state",
+  "target_capability": "multimodal live-prefix identity",
+  "difficulty": "troubleshooting",
+  "evidence_basis": [
+    "ds4.h:448-472",
+    "ds4.c:76362-76425",
+    "ds4_server.c:11593-11614",
+    "tests/test_session_state.c:6-117"
+  ],
+  "expected_skill_files": [
+    "sub-skills/serving-state/SKILL.md",
+    "sub-skills/serving-state/references/state-contracts.md",
+    "sub-skills/serving-state/references/troubleshooting.md"
+  ],
+  "assertions": [
+    "The response routes this cache question to serving-state and its bundled references.",
+    "It rejects old image-conditioned reuse when a historical fingerprint or embedding row count differs, even if visible text and tool IDs match.",
+    "It distinguishes the fingerprint-only routing probe from subsequent token-history, span-position and independent continuation checks.",
+    "It treats appending an image as conditional on preserving historical identity and placing the new image at or beyond the retained frontier, not as unconditional reuse.",
+    "It proposes a one-variable image-identity diagnostic without a model download or inference execution.",
+    "It keeps required native accelerator/model verification blocked and does not describe a synthetic or CPU predicate check as equivalent evidence."
+  ]
+}
diff --git a/tests/fixtures/repo_to_skill/dwarfstar/test-cases/sub-skills/serving-state/same-text-changed-image/user_request.txt b/tests/fixtures/repo_to_skill/dwarfstar/test-cases/sub-skills/serving-state/same-text-changed-image/user_request.txt
new file mode 100644
index 000000000..9f8fb2d58
--- /dev/null
+++ b/tests/fixtures/repo_to_skill/dwarfstar/test-cases/sub-skills/serving-state/same-text-changed-image/user_request.txt
@@ -0,0 +1 @@
+I'm debugging a local DwarfStar server. A client repeats exactly the same visible conversation and tool-call IDs but replaces the first image with a different PNG. Another request appends a second image without changing the first. Can we route both requests to the same retained KV state because the text is a prefix? Explain the checks and the smallest safe diagnostic, without downloading a model or running inference.
diff --git a/tests/test_repo_to_skill.py b/tests/test_repo_to_skill.py
new file mode 100644
index 000000000..3c9ccb224
--- /dev/null
+++ b/tests/test_repo_to_skill.py
@@ -0,0 +1,90 @@
+"""Portable AREX graph structure, not unavailable engine/backend verification."""
+
+import json
+import re
+import subprocess
+import sys
+from pathlib import Path
+from urllib.parse import unquote, urlsplit
+
+import yaml
+
+PROJECT = Path(__file__).resolve().parents[1]
+LIBRARY = PROJECT / "src/inferrsi/data/skill_library/repositories"
+FIXTURES = PROJECT / "tests/fixtures/repo_to_skill"
+GRAPHS = ("dwarfstar",)
+
+
+def frontmatter(path):
+    text = path.read_text(encoding="utf-8")
+    header = text.split("---", 2)[1]
+    assert re.search(r'^description: ".+"$', header, re.MULTILINE)
+    value = yaml.safe_load(header)
+    assert value["name"] == path.parent.name
+    assert re.fullmatch(r"[a-z0-9][a-z0-9-]{0,63}", value["name"])
+    assert value["license"] == "NO_LICENSE"
+    assert value["disable-model-invocation"] is True
+    assert value["metadata"]["disco-role"] == "operating"
+    return value
+
+
+def test_arex_runtime_graph_links_frontmatter_and_source_boundaries():
+    for name in GRAPHS:
+        root = LIBRARY / name
+        skills = list(root.rglob("SKILL.md"))
+        assert len(skills) == 4
+        for path in skills:
+            frontmatter(path)
+        for path in root.rglob("*.md"):
+            text = path.read_text(encoding="utf-8")
+            assert not any(value in text for value in ("/workspace/", "/home/agent/"))
+            for target in re.findall(r"\]\(([^)]+)\)", text):
+                link = urlsplit(target)
+                if link.scheme or link.netloc or not link.path:
+                    continue
+                selected = (path.parent / unquote(link.path)).resolve()
+                assert selected.is_relative_to(root.resolve()), (path, target)
+                assert selected.exists(), (path, target)
+        assert not list(root.rglob("__pycache__"))
+        assert not list(root.rglob("*.pyc"))
+        metadata = json.loads((root / "metadata.json").read_text(encoding="utf-8"))
+        assert metadata["status"] == "candidate"
+        assert metadata["construction"]["native_import"] == "not-run"
+        assert "BLOCKED_REQUIRED_BACKEND" in " ".join(metadata["limitations"])
+
+
+def test_arex_assertion_cases_resolve_their_expected_runtime_resources():
+    for name in GRAPHS:
+        root, cases = LIBRARY / name, FIXTURES / name / "test-cases"
+        assertions = list(cases.rglob("assertions.json"))
+        assert len(assertions) >= 9
+        assert len(list((cases / "integration").glob("*/assertions.json"))) == 2
+        for skill in (root / "sub-skills").iterdir():
+            assert len(list((cases / "sub-skills" / skill.name).glob("*/assertions.json"))) >= 2
+        for path in assertions:
+            case = json.loads(path.read_text(encoding="utf-8"))
+            assert case["schema"] == "disco.usability-case.v1"
+            assert case["assertions"] and case["evidence_basis"]
+            assert (path.parent / "user_request.txt").read_text().strip()
+            assert (path.parent / "README.md").is_file()
+            for relative in case["expected_skill_files"]:
+                selected = (root / relative).resolve()
+                assert selected.is_relative_to(root.resolve()) and selected.is_file()
+
+
+def test_dwarfstar_request_helper_is_offline_and_validates_its_small_contract():
+    script = LIBRARY / "dwarfstar/scripts/request_template.py"
+    result = subprocess.run(
+        [sys.executable, str(script), "--model", "local-model", "--prompt", "Hello"],
+        check=True, capture_output=True, text=True,
+    )
+    value = json.loads(result.stdout)
+    assert value == {
+        "model": "local-model", "messages": [{"role": "user", "content": "Hello"}],
+        "max_tokens": 64, "temperature": 0.0, "stream": False,
+    }
+    invalid = subprocess.run(
+        [sys.executable, str(script), "--model", "x", "--prompt", "y", "--max-tokens", "0"],
+        capture_output=True, text=True,
+    )
+    assert invalid.returncode == 2 and "positive" in invalid.stderr
diff --git a/tests/test_skill_index.py b/tests/test_skill_index.py
index 01985be34..a82129062 100644
--- a/tests/test_skill_index.py
+++ b/tests/test_skill_index.py
@@ -23,7 +23,7 @@ def test_complete_index_retains_metadata_states_and_unknown_license():
     records = search_knowledge(repo="inferrsi", limit=100)
     names = {record["name"] for record in skill_records(ROOT)}
     registry = json.loads((ROOT / "skill-index.json").read_text(encoding="utf-8"))["skills"]
-    assert len(records) == len(names) == 60
+    assert len(records) == len(names) == 61
     assert {record["skill_metadata"]["name"] for record in records} == names
     states = []
     for record in records:
@@ -41,7 +41,7 @@ def test_complete_index_retains_metadata_states_and_unknown_license():
         }
         assert set(record["source_files"]) == set(record["sources"][0]["paths"])
         states.append(metadata.get("status"))
-    assert states.count("candidate") == 55
+    assert states.count("candidate") == 56
     assert states.count(None) == 5
     profiling = next(r for r in records if r["title"] == "ascend-operator-profiling")
     assert profiling["skill_metadata"]["license"] == "UNKNOWN"
@@ -161,7 +161,7 @@ def test_package_index_and_checkout_skill_have_distinct_read_boundaries():
     assert reader.resolve("claim-contract") == "checkout:claim-contract"
     assert reader.read("claim-contract").metadata["kind"] == "skill"
     collected = reader.collect(repo="inferrsi", kind="knowledge", limit=100, material=True)
-    assert len(collected["items"]) == 60
+    assert len(collected["items"]) == 61
     assert not collected["collisions"] and not collected["omissions"]
     assert all(item["authority"] == "package" for item in collected["items"])
     assert all("record.json" in item["resources"] for item in collected["items"])
@@ -191,9 +191,9 @@ def test_index_cli_search_show_context_and_execution_boundary(capsys):
 
 
 def test_hardware_filters_include_only_matching_and_shared_skills():
-    for backend, count in (("ascend", 56), ("cuda", 47), ("cpu", 43)):
+    for backend, count in (("ascend", 57), ("cuda", 48), ("cpu", 44)):
         records = search_knowledge(repo="inferrsi", backend=backend, limit=100)
         assert len(records) == count
         assert all(record["backend"] in (backend, "common") for record in records)
     assert not (PROJECT / "skills").exists()
-    assert len(list(ROOT.rglob("SKILL.md"))) == 81
+    assert len(list(ROOT.rglob("SKILL.md"))) == 85
diff --git a/tests/test_skill_library.py b/tests/test_skill_library.py
index 70649accf..f4fab68ec 100644
--- a/tests/test_skill_library.py
+++ b/tests/test_skill_library.py
@@ -154,7 +154,7 @@ def test_library_root_exposes_only_owner_categories_and_required_entrypoints():
     assert {path.name for path in LIBRARY.iterdir() if path.is_dir()} == {
         "hardware", "frameworks", "repositories", "methods", "workflows",
     }
-    assert len(list(LIBRARY.rglob("SKILL.md"))) == 81
+    assert len(list(LIBRARY.rglob("SKILL.md"))) == 85
     assert not any(path.name == "zh-CN" for path in LIBRARY.rglob("*") if path.is_dir())
     references = LIBRARY / "workflows/inference-engine-research/references"
     assert (references / "import.md").is_file()