docs: qualify Sarathi chunked-prefill budget case

docs/reference-matrix.md

@@ -151,6 +151,7 @@ not fabricated implementation SHAs or a version-matched runtime stack.
 |---|---|---|---|
 | Vidur — [2405.05465v2](https://arxiv.org/abs/2405.05465v2) | [A readset](../src/inferrsi/data/skill_library/methods/performance-ceiling-analysis/references/papers-20261015/READSET.md); §§3–7, A.1–A.2 | mined — operator/queue model calibration and joint-SLO shortlist → [calibrate-serving-simulation](../src/inferrsi/data/skill_library/methods/calibrate-serving-simulation/SKILL.md), ceiling analysis | Empirical prediction, not physical bound; near-capacity CPU overhead and 12.65% maximum error in the specified appendix case; no current runtime calibration |
 | Orca — OSDI 2022 proceedings | [Exact readset](../src/inferrsi/data/skill_library/repositories/vllm/diagnose-vllm-admission-pressure/references/papers-20261015/references/orca-readset.md); §§3–6, Algorithm 1 | mined — iteration eligibility, selective batching, FCFS/reservation → [assess-serving-preemption](../src/inferrsi/data/skill_library/methods/assess-serving-preemption/SKILL.md) | Historical engine/harness and normalized-median denominators; not present vLLM behavior or fairness proof |
+| Sarathi-Serve — [OSDI 2024 proceedings, pp. 117–134](https://www.usenix.org/conference/osdi24/presentation/agrawal) | Fresh official-edition grounding, 2026-10-30; full text and artifact appendix; §§4.1–4.3, §5 setup, Figures 9/14 and Table 4 checked for the [case record](../src/inferrsi/data/skill_library/methods/optimization-search-policy/cases/sarathi-chunk-budget.json) | mined — decode/partial-prefill/new-request token accounting and chunk-cost tradeoffs → [one optimization-search-policy case](../src/inferrsi/data/skill_library/methods/optimization-search-policy/cases/sarathi-chunk-budget.md) | Source-derived candidate; empirical not_run. Combined design has worse P50 TTFT than hybrid-only in both Table 4 traces; no current-engine, universal-budget or fairness claim. Raw source reuse rights UNKNOWN; no archive recovery or novelty over unavailable supplier synthesis claimed. |
 | FastServe — [2305.05920v3](https://arxiv.org/abs/2305.05920v3) | [Exact readset](../src/inferrsi/data/skill_library/repositories/vllm/diagnose-vllm-admission-pressure/references/papers-20261015/references/fastserve-readset.md) | mined — service priority, promotion and suspended-state host residency → serving-preemption candidate | Aggregate tail improvement is not tenant fairness; same-engine contrast distinct from historical-vLLM comparison |
 | Llumnix — [2406.03243v1](https://arxiv.org/abs/2406.03243v1) | [Exact readset](../src/inferrsi/data/skill_library/repositories/vllm/diagnose-vllm-admission-pressure/references/papers-20261015/references/llumnix-readset.md) | mined — replica migration, virtual headroom, pre-copy/downtime and host staging → [evaluate-elastic-serving-placement](../src/inferrsi/data/skill_library/methods/evaluate-elastic-serving-placement/SKILL.md) | Virtual headroom is not physical capacity; normal-class cost and evaluated-artifact fault-tolerance limit retained |
 | LoongServe — [2404.09526v2](https://arxiv.org/abs/2404.09526v2) | [Exact readset](../src/inferrsi/data/skill_library/repositories/vllm/diagnose-vllm-admission-pressure/references/papers-20261015/references/loongserve-readset.md) | mined — elastic sequence groups, phase-dependent state placement and ongoing attention communication → elastic-placement candidate | Distributed device KV is not host RAM; group transitions have different migration costs; no current engine/NPU claim |
@@ -176,6 +177,9 @@ the supplied local directory. Its six-paper extraction and standalone candidates
 are not imported or counted here. Existing DeepSpeed configuration evidence and
 Mooncake transport discovery are not substitutes for those paper readsets.
 
+The separate Sarathi-Serve case above uses a freshly pinned official edition;
+it does not recover or import that archive or its supplier-authored synthesis.
+
 ### Negative findings (scoped to inspected paths and revisions)
 
 - **Disk-level storage I/O mechanics** (mmap/pinned staging buffers,

docs/searchspace-matrix.md

@@ -46,7 +46,7 @@ The primary compute and HBM resource. Most current work lives here.
 | **Diagnose / Configure / Verify** | AKG static CUDA Affine loop tails: effective target/option gates, full point-loop bounds and protected input/output accesses | [kernel-writing-guide](../src/inferrsi/data/skill_library/methods/kernel-writing-guide/SKILL.md) | [Tail-guard reference](../src/inferrsi/data/skill_library/methods/kernel-writing-guide/references/akg-affine-tail-guards.md): separation precedence, non-divisible fixture and diagnostic boundary | AKG @ 5aa15f3fe4d856eb60cd3b27447e04f5f1fb4dff; source + FileCheck expectations only | Candidate; compiler/IR command, CUDA runtime and agent efficacy not run; no new measured coverage |
 | **Optimize** | Operator-level: staged accuracy-gated tuning campaign (code-level → scheduling → instruction-level), metric→action routing; AscendC tiling design per operator class (reduction/broadcast/elewise/matmul/conversion) | `ascend-operator-tuning`, `ascendc-tiling-design` | PyPTO frontend/swimlane/incore playbooks; fullload-vs-split, four-element tiling template, chip-generation (DAV_2201/3510) routing | awesome-ascend-skills pypto-op-perf-tune + ascendc-tiling-design @ 75ec89c | 🟢 |
 | **Optimize** | Casebook of past NPU optimization decisions: AI-CPU fallback, memory-bound, comm-exposed, graph-compilation, scheduling idle gaps | `ascend-optimization-casebook`, `inference-optimization-casebook` | Five bottleneck classes with symptom→diagnosis→action→evidence schema | MindIE-SD profiling-analyze + dit-perf-opt + perf-gate + dit-parallel-opt @ 8481d39b | 🟢 |
-| **Optimize** | Bounded experiment selection, distinct mechanism hypotheses, incumbent preservation and final-artifact integrity | `optimization-search-policy` | Methodology provenance stays in Reference Matrix §2; no hardware coverage from agent-paper scores | vllm-evolve @ 43cc82c; scoped A/F paper amendments | 🟢 (procedure only) |
+| **Optimize** | Bounded experiment selection, distinct mechanism hypotheses, incumbent preservation and final-artifact integrity | `optimization-search-policy` | [Sarathi chunk-budget case](../src/inferrsi/data/skill_library/methods/optimization-search-policy/cases/sarathi-chunk-budget.md): decode-first accounting, prefill overhead and TTFT/TBT tradeoffs; no universal budget. Methodology provenance stays in Reference Matrix §2; no hardware coverage from agent-paper scores | vllm-evolve @ 43cc82c; scoped A/F paper amendments; separate official OSDI 2024 Sarathi case/readset in Reference Matrix | 🟢 (source-derived candidate procedure; empirical not_run) |
 | **Optimize** | Bounded multistream overlap: dependency-DAG constrained (data/state/event/collective_order edges; shared input ≠ dependency), minimal-first, execution-mode sync contract (eager streams vs TorchAir) | `cann-attribution-overlap` | Metric separation (wall / union-of-busy / summed durations / summed duration+wait; repeated-instance product = estimate); device-memory domain qualification; formula conflicts carried unresolved | cann/oam-tools @ 8440935f (pack-mediated, wave 6) | 🟢 |
 | **Verify** | Evidence taxonomy (L0 correctness → L5 production), claim-evidence matching | `dense-performance-evidence`, `rsi-evidence`, `claim-contract` | InferRSI PASS/FAIL/INCONCLUSIVE protocol | InferRSI @ 4c8e1285 | 🟢 |
 | **Diagnose / Verify** | Fixed-stage PyTorch/Ascend model report: bind current stage and actually loaded candidate to quality-validator output and comparable baseline | [ascend-performance-analysis](../src/inferrsi/data/skill_library/hardware/ascend/ascend-performance-analysis/SKILL.md) | [Stage-evidence rule](../src/inferrsi/data/skill_library/hardware/ascend/ascend-performance-analysis/references/cannbot-stage-evidence.md): workflow instructions versus registered marker/keyword checks; completion is not numerical validation | canonical cann/cannbot @ c0e2ad23ee1a73e3c0f84c8df1cc459e2aae3110; selected CANN-2.0 scope | Candidate procedure only; host/model/backend not_run, agent efficacy not evaluated; no new measured coverage |

docs/skill-coverage.md

@@ -4,7 +4,7 @@ Generated from the registry, the skill sidecars and the two maintained
 matrices. **Do not edit by hand**; regenerate after any registry, skill,
 matrix or frozen-plan change.
 
-- Input digest: `sha256:b6f5cec72e5e52fcc4113bf50c68612417688b40c54cc79cbd24a611db574f3c`
+- Input digest: `sha256:bf6ac4fe724e1f91f7373ed7352852919749aa9cb1e53e8d27e68e3e1df4b8ff`
 - Regenerate: `python tools/build_skill_coverage.py`
 - Verify without writing: `python tools/build_skill_coverage.py --check`
 
@@ -129,7 +129,7 @@ optimization is fully covered or that any procedure has measured benefit.
 | [inference-benchmark-design](../src/inferrsi/data/skill_library/methods/inference-benchmark-design/SKILL.md) | common | Select task quality gates, reference parity checks and performance workloads for text, agentic coding, work agents ... | **unknown** | — | — | 2 · Methodology and corpus references (JiusiServe/vllm-evolve @ 43cc82c)<br>Task-specific benchmark selection (issue #60) ([Issue #60 report](https://github.com/InferMatrixLab/InferRSI/issues/6 ...) | candidate · not_run · 2 readset(s) | no-SearchSpace-layer, no-applicability |
 | [inference-optimization-casebook](../src/inferrsi/data/skill_library/methods/inference-optimization-casebook/SKILL.md) | common | How to structure and use a casebook of past optimization decisions as a decision-support tool. Extracts the casebook ... | **unknown** | — | 1 Accelerator · Optimize | 1 · Systems sources (NVIDIA/TensorRT-LLM @ f33e0e5fd) | candidate · not_run · 0 readset(s) | no-applicability, no-eval |
 | [kernel-writing-guide](../src/inferrsi/data/skill_library/methods/kernel-writing-guide/SKILL.md) | common | Decision guide for writing custom inference kernels: when to write one, which framework to choose (Triton, CuTe DSL ... | **unknown** | — | 1 Accelerator · Optimize / Diagnose / Configure / Verify | 1 · Systems sources (Dao-AILab/causal-conv1d @ `cd81f0413cad2fc1e6f17e785ac39f59aae690cd` ...)<br>1 · Systems sources (NVIDIA/TensorRT-LLM @ f33e0e5fd)<br>1 · Systems sources (huggingface/transformers @ `0dad7b822255a0ae261ec45ae937371e859ffd1a` ...)<br>1 · Systems sources (mindspore-ai/akg @ `5aa15f3fe4d856eb60cd3b27447e04f5f1fb4dff` ...)<br>1 · Systems sources (sgl-project/sglang @ `16176278d0180f1a1a578e53e6eef432577876c2` ...)<br>1 · Systems sources (vllm-project/vllm @ `537a79f0fc5ba564d2bef43d54915b2f80f9f77d` (root ...) | candidate · not_run · 7 readset(s) | no-applicability, no-eval |
-| [optimization-search-policy](../src/inferrsi/data/skill_library/methods/optimization-search-policy/SKILL.md) | common | Policy for how an optimization agent should select, order, and stop optimization experiments. Covers candidate ... | **unknown** | — | 1 Accelerator · Optimize | 2 · Methodology and corpus references (JiusiServe/vllm-evolve @ 43cc82c)<br>2 · Methodology and corpus references (RRSI — arXiv 2609.24972v3)<br>2 · Methodology and corpus references (arXiv 2608.01867 (CRISP))<br>2 · Methodology and corpus references (microsoft/SkillOpt @ 343db22) | candidate · not_run · 6 readset(s) | no-applicability, no-eval |
+| [optimization-search-policy](../src/inferrsi/data/skill_library/methods/optimization-search-policy/SKILL.md) | common | Policy for how an optimization agent should select, order, and stop optimization experiments. Covers candidate ... | **unknown** | — | 1 Accelerator · Optimize | 2 · Methodology and corpus references (JiusiServe/vllm-evolve @ 43cc82c)<br>2 · Methodology and corpus references (RRSI — arXiv 2609.24972v3)<br>2 · Methodology and corpus references (arXiv 2608.01867 (CRISP))<br>2 · Methodology and corpus references (microsoft/SkillOpt @ 343db22) | candidate · not_run · 7 readset(s) | no-applicability, no-eval |
 | [performance-ceiling-analysis](../src/inferrsi/data/skill_library/methods/performance-ceiling-analysis/SKILL.md) | common | Estimate optimization headroom before modifying an inference system. Produces a three-level ceiling model (analytical ... | inference serving systems (LLM, diffusion, audio) | — | 1 Accelerator · Diagnose / Verify<br>5 OS · Configure | 1 · Systems sources (NVIDIA/TensorRT-LLM @ f33e0e5fd)<br>1 · Systems sources (ebugger/Empirical-Roofline-Toolkit @ 0771fa9)<br>1 · Systems sources (hahnyuan/LLM-Viewer @ 1893e4b) | candidate · not_run · 16 readset(s) | no-eval |
 | [plan-activation-residency](../src/inferrsi/data/skill_library/methods/plan-activation-residency/SKILL.md) | common | Propose training activation and gradient memory reductions by distinguishing lifetime reordering, address packing and ... | **unknown** | — | 1 Accelerator · Diagnose / Optimize / Verify | 1 · Systems sources (Checkmate — [1910.02653v3](https://arxiv.org/abs/1910.02653v3)) | candidate · not_run · 1 readset(s) | no-applicability, no-eval |
 | [repository-evolution-analysis](../src/inferrsi/data/skill_library/methods/repository-evolution-analysis/SKILL.md) | common | Analyze a repository's commit history to build a capability-evolution model: what the system can do, how it got there ... | any software repository with a recoverable git history | — | — | 2 · Methodology and corpus references (LearningToCommit/LearningToCommit @ 7e5a1d7)<br>2 · Methodology and corpus references (yetone/magpie @ 77800032) | candidate · not_run · 0 readset(s) | no-SearchSpace-layer, no-eval |

src/inferrsi/data/repo_packs/inferrsi/knowledge/README.md

@@ -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: [bde70d82bb65cc2e5dcd0f9d33cda5d38091120b](https://github.com/InferMatrixLab/InferRSI/tree/bde70d82bb65cc2e5dcd0f9d33cda5d38091120b/src/inferrsi/data/skill_library).
+Source snapshot: [c5e13944c06a8a03ab7015a8933981a983e42186](https://github.com/InferMatrixLab/InferRSI/tree/c5e13944c06a8a03ab7015a8933981a983e42186/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 bde70d82bb65cc2e5dcd0f9d33cda5d38091120b --check
+python tools/build_skill_index.py --source-revision c5e13944c06a8a03ab7015a8933981a983e42186 --check
 ```
 
 `--check` detects missing, stale and unexpected generated records without writing.

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/bde70d82bb65cc2e5dcd0f9d33cda5d38091120b/src/inferrsi/data/skill_library/hardware/ascend/ascend-aclgraph-capture-replay/SKILL.md",
-      "commit": "bde70d82bb65cc2e5dcd0f9d33cda5d38091120b",
+      "url": "https://github.com/InferMatrixLab/InferRSI/blob/c5e13944c06a8a03ab7015a8933981a983e42186/src/inferrsi/data/skill_library/hardware/ascend/ascend-aclgraph-capture-replay/SKILL.md",
+      "commit": "c5e13944c06a8a03ab7015a8933981a983e42186",
       "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",

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/bde70d82bb65cc2e5dcd0f9d33cda5d38091120b/src/inferrsi/data/skill_library/hardware/ascend/ascend-device-health/SKILL.md",
-      "commit": "bde70d82bb65cc2e5dcd0f9d33cda5d38091120b",
+      "url": "https://github.com/InferMatrixLab/InferRSI/blob/c5e13944c06a8a03ab7015a8933981a983e42186/src/inferrsi/data/skill_library/hardware/ascend/ascend-device-health/SKILL.md",
+      "commit": "c5e13944c06a8a03ab7015a8933981a983e42186",
       "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",

src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/ascend-distributed-parallel.json

@@ -27,8 +27,8 @@
   "sources": [
     {
       "type": "git",
-      "url": "https://github.com/InferMatrixLab/InferRSI/blob/bde70d82bb65cc2e5dcd0f9d33cda5d38091120b/src/inferrsi/data/skill_library/hardware/ascend/ascend-distributed-parallel/SKILL.md",
-      "commit": "bde70d82bb65cc2e5dcd0f9d33cda5d38091120b",
+      "url": "https://github.com/InferMatrixLab/InferRSI/blob/c5e13944c06a8a03ab7015a8933981a983e42186/src/inferrsi/data/skill_library/hardware/ascend/ascend-distributed-parallel/SKILL.md",
+      "commit": "c5e13944c06a8a03ab7015a8933981a983e42186",
       "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",

src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/ascend-graph-mode.json

@@ -26,8 +26,8 @@
   "sources": [
     {
       "type": "git",
-      "url": "https://github.com/InferMatrixLab/InferRSI/blob/bde70d82bb65cc2e5dcd0f9d33cda5d38091120b/src/inferrsi/data/skill_library/hardware/ascend/ascend-graph-mode/SKILL.md",
-      "commit": "bde70d82bb65cc2e5dcd0f9d33cda5d38091120b",
+      "url": "https://github.com/InferMatrixLab/InferRSI/blob/c5e13944c06a8a03ab7015a8933981a983e42186/src/inferrsi/data/skill_library/hardware/ascend/ascend-graph-mode/SKILL.md",
+      "commit": "c5e13944c06a8a03ab7015a8933981a983e42186",
       "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",

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/bde70d82bb65cc2e5dcd0f9d33cda5d38091120b/src/inferrsi/data/skill_library/hardware/ascend/ascend-model-adaptation/SKILL.md",
-      "commit": "bde70d82bb65cc2e5dcd0f9d33cda5d38091120b",
+      "url": "https://github.com/InferMatrixLab/InferRSI/blob/c5e13944c06a8a03ab7015a8933981a983e42186/src/inferrsi/data/skill_library/hardware/ascend/ascend-model-adaptation/SKILL.md",
+      "commit": "c5e13944c06a8a03ab7015a8933981a983e42186",
       "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",

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/bde70d82bb65cc2e5dcd0f9d33cda5d38091120b/src/inferrsi/data/skill_library/hardware/ascend/ascend-operator-profiling/SKILL.md",
-      "commit": "bde70d82bb65cc2e5dcd0f9d33cda5d38091120b",
+      "url": "https://github.com/InferMatrixLab/InferRSI/blob/c5e13944c06a8a03ab7015a8933981a983e42186/src/inferrsi/data/skill_library/hardware/ascend/ascend-operator-profiling/SKILL.md",
+      "commit": "c5e13944c06a8a03ab7015a8933981a983e42186",
       "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",

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/bde70d82bb65cc2e5dcd0f9d33cda5d38091120b/src/inferrsi/data/skill_library/hardware/ascend/ascend-operator-tuning/SKILL.md",
-      "commit": "bde70d82bb65cc2e5dcd0f9d33cda5d38091120b",
+      "url": "https://github.com/InferMatrixLab/InferRSI/blob/c5e13944c06a8a03ab7015a8933981a983e42186/src/inferrsi/data/skill_library/hardware/ascend/ascend-operator-tuning/SKILL.md",
+      "commit": "c5e13944c06a8a03ab7015a8933981a983e42186",
       "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",