knowledge: audit Infera serving-search result provenance

docs/reference-matrix.md

@@ -151,6 +151,7 @@ not fabricated implementation SHAs or a version-matched runtime stack.
 | Original source / version | Inspected scope and local evidence | Mechanism / consumer | Evidence limits |
 |---|---|---|---|
 | 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 |
+| [AMD Infera @ ff1136e4](https://github.com/AMD-AGI/Infera/tree/ff1136e4a51998eba2933ed63cfb9c10120f86ea) | Direct bounded source inspection, 2026-11-03: serving CLI/evaluator/history, seed/objective parser, launcher/DES countercheck and named unexecuted test definitions; [16-path readset and rights](../src/inferrsi/data/skill_library/methods/calibrate-serving-simulation/references/infera-serving-search/source-record.json) | mined — `calibrate-serving-simulation`: one [calibration-owned provenance reference](../src/inferrsi/data/skill_library/methods/calibrate-serving-simulation/references/infera-serving-search/README.md); search-policy links its seed comparator and evaluated-trial ledger | Source-derived candidate; empirical not_run. Scalar tuner scoring and DES percentiles are separate; missing SLO fields and stale no-P/D-queues text are qualified. Selected original paths MIT; no runtime, ranking accuracy, quality, service or agent-benefit result. |
 | Mooncake — FAST '25 proceedings, pp. 155-170 | Direct public-paper read, retained qualification 2026-10-31; [source condition and readset](../src/inferrsi/data/skill_library/methods/performance-ceiling-analysis/references/mooncake-prefix-reuse-feasibility.md), §2.2, Table 1, Eqs. 1-2 | mined — conditional prefix-KV transfer versus saved-prefill-work comparison → `performance-ceiling-analysis` | Analytical gate only; current inputs and E2E benefit unverified. SSD mechanics excluded. Original transfer-gates extraction history remains positive; missing synthesis novelty UNKNOWN. Raw-paper redistribution permission UNKNOWN. |
 | FlexGen — [ICML 2023 / PMLR 202:31094–31116](https://proceedings.mlr.press/v202/sheng23a.html), official ten-author edition | Direct scoped public-paper qualification, 2026-10-31; [condition, exact digest and readset](../src/inferrsi/data/skill_library/methods/performance-ceiling-analysis/references/flexgen-offload-feasibility.md), §§3–4.3 and selected Appendix A.3 formulas; Figures 1–3, Algorithm 1, Eq. 1, Tables 2–3 | mined — joint weights/KV/activation placement, resident plus working-storage peaks and offline latency/throughput conditions → `performance-ceiling-analysis` | Source-derived candidate; empirical not_run. Perfect-overlap fitted model, relaxed-placement OOM and unresolved `bls < 20` versus `32 × 8` wording retained; no current runtime, quality or gain qualification. Attributed CC BY 4.0 synthesis; missing-archive novelty UNKNOWN. |
 | 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 |

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` | [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 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. [Infera provenance](../src/inferrsi/data/skill_library/methods/calibrate-serving-simulation/references/infera-serving-search/README.md) adds a seed-only comparator and evaluated-trial ledger example; its trace-ranking check is proposed. Methodology provenance stays in Reference Matrix §2; no hardware coverage from agent-paper scores | vllm-evolve @ 43cc82c; scoped A/F paper amendments; separate Sarathi case and bounded Infera Git 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 |
@@ -58,7 +58,7 @@ The primary compute and HBM resource. Most current work lives here.
 | **Observe / Diagnose / Verify** | KV receive/save completion, delayed-free ownership, in-flight-write fences and failure reporting | `audit-vllm-kv-lifetimes` | Event ledger is proposed, not a telemetry exporter; abstract obligations require concrete connector/caller evidence | vLLM @ 7f730b4cea9a, Dot bundle | 🟢 (contract intelligence; concrete backend unverified) |
 | **Diagnose / Configure / Optimize / Verify** | Runtime ACLGraph dispatch, stable input addresses, capture-resource limits and branch-specific replay/update ordering | `ascend-aclgraph-capture-replay` | Caller owns persistent buffers; DEBUG checks positional tensors only; reduced capture sets trade coverage for resources; [version/responsibility routing](../src/inferrsi/data/skill_library/README.md#vllm-ascend-version-and-responsibility-routing) | vllm-ascend @ 8a8e1e05ecb6, Dot bundle; wrapper/guide/tests inspected by supplier | 🟢 (conditional proposals; tests not run) |
 | **Diagnose / Configure / Optimize / Verify** | Device-HBM stream-use lifetime, fragmentation/counter interpretation and gated allocator controls | `torchnpu-device-memory-lifecycle` | Lifetime recording does not insert data dependencies; expandable segments change counters/control legality; GC synchronization/comment conflict retained | TorchNPU @ 67511c53cd4c, Dot bundle; device allocator ledger | 🟢 (source-supported procedure; NPU execution not run) |
-| **Diagnose / Verify** | Serving model: operator/context dependencies, calibration/load domain and joint-SLO prediction uncertainty | [calibrate-serving-simulation](../src/inferrsi/data/skill_library/methods/calibrate-serving-simulation/SKILL.md), `performance-ceiling-analysis` | Vidur near-capacity CPU-overhead counterexample; simulator output is not a bound | Vidur v2; Reference Matrix systems-paper row | 🟢 (prediction guidance; calibration not run) |
+| **Diagnose / Verify** | Serving model: operator/context dependencies, calibration/load domain, fidelity provenance and joint-SLO prediction uncertainty | [calibrate-serving-simulation](../src/inferrsi/data/skill_library/methods/calibrate-serving-simulation/SKILL.md), `performance-ceiling-analysis` | Vidur near-capacity CPU-overhead counterexample; simulator output is not a bound. [Infera source reference](../src/inferrsi/data/skill_library/methods/calibrate-serving-simulation/references/infera-serving-search/README.md) separates scalar tuner scoring from DES percentiles and retains missing-SLO/stale-document limits; no service guarantee | Vidur v2 and bounded Infera Git readset; Reference Matrix systems sources | 🟢 (prediction guidance; calibration not run) |
 | **Diagnose / Verify** | Reusable-prefix compute work versus host/network KV loading cost | [performance-ceiling-analysis](../src/inferrsi/data/skill_library/methods/performance-ceiling-analysis/SKILL.md), [Mooncake condition](../src/inferrsi/data/skill_library/methods/performance-ceiling-analysis/references/mooncake-prefix-reuse-feasibility.md) | Check bytes, FLOPs and effective rate at one GPU-group scope; favorable Equation 2 still requires exposed-cost/SLO measurements | Mooncake FAST '25 §2.2, Table 1, Eqs. 1-2; separate public-paper source, not recovered foundation archive | 🟢 (analytical source condition; current feasibility and E2E benefit unverified; no SSD credit) |
 | **Diagnose / Verify** | Offline offload feasibility across GPU, CPU RAM and disk | [performance-ceiling-analysis](../src/inferrsi/data/skill_library/methods/performance-ceiling-analysis/SKILL.md), [FlexGen condition](../src/inferrsi/data/skill_library/methods/performance-ceiling-analysis/references/flexgen-offload-feasibility.md) | Joint weights/KV/activation placement; RAM resident plus staging/working peaks, SSD state capacity and disk↔CPU cost, effective batch versus GPU batch, exposed CPU/transfer cost and block-latency objective | FlexGen official ICML 2023, §§3–4.3 and selected A.3; separate public source, foundation archive still unavailable | Source-only candidate; empirical not_run, fitted/perfect-overlap model and relaxed OOM limits; no blanket RAM/SSD, Network/OS or current-engine qualification |
 | **Diagnose / Configure / Optimize / Verify** | Serving service order versus HBM admission and suspended-KV recovery; replica migration versus elastic sequence groups | [assess-serving-preemption](../src/inferrsi/data/skill_library/methods/assess-serving-preemption/SKILL.md), [evaluate-elastic-serving-placement](../src/inferrsi/data/skill_library/methods/evaluate-elastic-serving-placement/SKILL.md) | [Hypothetical pressure/migration cases](../src/inferrsi/data/skill_library/repositories/vllm/diagnose-vllm-admission-pressure/references/papers-20261015/knowledge/scheduling-state-boundaries.md); cohort/fairness and destination-reservation costs | Orca proceedings, FastServe v3, Llumnix v1, LoongServe v2; implementation pins remain separate | 🟢 (conditional paper-derived plans) |

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: `sha256:7640b171256091efd1514a33d5e434eef3feaed238fc92fe3eeba02fee94e9ff`
+- Input digest: `sha256:930a3c698dd502804d2bff47cd6c49311e51492d68497d2986110f51bf18ceae`
 - Regenerate: `python tools/build_skill_coverage.py`
 - Verify without writing: `python tools/build_skill_coverage.py --check`
 
@@ -126,7 +126,7 @@ optimization is fully covered or that any procedure has measured benefit.
 | [assess-serving-preemption](../src/inferrsi/data/skill_library/methods/assess-serving-preemption/SKILL.md) | common | Assess whether LLM serving queue delays justify token-boundary priority preemption, accounting for suspended KV, host ... | Diagnosing whether LLM serving queue delay justifies token-boundary priority preemption ... | failure_modes, input_contract, produced_artifact, refusal_point<br>4 advice | 1 Accelerator · Diagnose / Configure / Optimize / Verify<br>2 RAM/Host · Diagnose / Configure / Optimize / Verify | 1 · Systems sources (Orca — OSDI 2022 proceedings) | candidate · not_run · 1 readset(s) | no-eval |
 | [assess-split-backward-schedule](../src/inferrsi/data/skill_library/methods/assess-split-backward-schedule/SKILL.md) | common | Evaluate memory-capped split-backward and V-shaped pipeline scheduling proposals, preserving synchronous update ... | Assessing memory-capped split-backward and V-shaped pipeline schedule proposals when a ... | failure_modes, input_contract, produced_artifact, refusal_point<br>3 advice | 1 Accelerator · Diagnose / Configure / Optimize / Verify | 1 · Systems sources (Zero Bubble — [2401.10241v1](https://arxiv.org/abs/2401.10241v1)) | candidate · not_run · 1 readset(s) | no-eval |
 | [audit-fixed-workload-claim](../src/inferrsi/data/skill_library/methods/audit-fixed-workload-claim/SKILL.md) | common | Audit a performance claim from supplied fixed-workload reports and an explicit measurement contract. Check aggregation ... | Auditing a performance claim against supplied fixed-workload reports and an explicit ... | failure_modes, input_contract, produced_artifact, refusal_point<br>3 advice | N/A: SKILL.md §§ Preserve the observation hierarchy, Check what was actually compared, and Return an auditable conclusion audit supplied measurements and verdicts; device identity is comparison metadata, not a hardware-resource mechanism. | — | candidate · not_run · 0 readset(s) | no-Reference-Matrix-row, no-eval |
-| [calibrate-serving-simulation](../src/inferrsi/data/skill_library/methods/calibrate-serving-simulation/SKILL.md) | common | Build an evidence-gated serving-configuration shortlist from a profiled simulator while separating empirical ... | Building an evidence-gated serving-configuration shortlist when a profiled simulator or ... | failure_modes, input_contract, produced_artifact, refusal_point<br>3 advice | 1 Accelerator · Diagnose / Verify | 1 · Systems sources (Vidur — [2405.05465v2](https://arxiv.org/abs/2405.05465v2)) | candidate · not_run · 1 readset(s) | no-eval |
+| [calibrate-serving-simulation](../src/inferrsi/data/skill_library/methods/calibrate-serving-simulation/SKILL.md) | common | Build an evidence-gated serving-configuration shortlist from a profiled simulator while separating empirical ... | Building an evidence-gated serving-configuration shortlist when a profiled simulator or ... | failure_modes, input_contract, produced_artifact, refusal_point<br>3 advice | 1 Accelerator · Diagnose / Verify | 1 · Systems sources (Vidur — [2405.05465v2](https://arxiv.org/abs/2405.05465v2))<br>1 · Systems sources ([AMD Infera @ ff1136e4](https://github.com/AMD-AGI/Infera/tree/ff1136e ...) | candidate · not_run · 3 readset(s) | no-eval |
 | [checkpoint-recovery-consistency](../src/inferrsi/data/skill_library/methods/checkpoint-recovery-consistency/SKILL.md) | common | Assess distributed checkpoint recovery using DLRover memory agreement, persistence completion and replica transport ... | Assessing distributed checkpoint recovery through DLRover memory agreement, persistence ... | failure_modes, input_contract, produced_artifact, refusal_point<br>4 advice | 2 RAM/Host · Configure / Verify<br>3 SSD · Configure / Verify<br>4 Network · Diagnose / Configure / Verify<br>5 OS · Diagnose / Configure | 1 · Systems sources (C3 / DLRover FlashCheckpoint @ `e35b1ad90e2928ca6bacf5d22117f71356 ...)<br>1 · Systems sources (C4 / NVIDIA Resiliency Extension @ ...) | candidate · not_run · 1 readset(s) | no-eval |
 | [choose-parallel-training-plan](../src/inferrsi/data/skill_library/methods/choose-parallel-training-plan/SKILL.md) | common | Compare stage, device-mesh and operator-sharding candidates for static distributed training while accounting for ... | Comparing stage, device-mesh and operator-sharding candidates for static distributed ... | failure_modes, input_contract, produced_artifact, refusal_point<br>3 advice | 1 Accelerator · Diagnose / Configure / Optimize / Verify<br>4 Network · Diagnose / Configure / Optimize / Verify | 1 · Systems sources (Alpa — [2201.12023v3](https://arxiv.org/abs/2201.12023v3)) | candidate · not_run · 1 readset(s) | no-eval |
 | [dense-performance-evidence](../src/inferrsi/data/skill_library/methods/dense-performance-evidence/SKILL.md) | common | Taxonomy of performance evidence an optimization agent needs, ordered by collection cost and diagnostic power. Maps ... | Deciding which performance evidence to collect next and which claim each evidence level ... | failure_modes, input_contract, produced_artifact, refusal_point<br>4 advice | 1 Accelerator · Observe / Profile / Verify<br>2 RAM/Host · Verify<br>4 Network · Verify | 1 · Systems sources (NVIDIA/TensorRT-LLM @ f33e0e5fd)<br>2 · Methodology and corpus references (JiusiServe/vllm-evolve @ 43cc82c)<br>2 · Methodology and corpus references (microsoft/SkillOpt @ 343db22) | candidate · not_run · 0 readset(s) | — |