Repository navigation
docs: describe conditions and advice for 22 hardware skills - #283
Conversation
Preserve source provenance and evidence states while making the declared scope explicit. Agent-assisted-by: Codex / gpt-6-astra [effort=xhigh] (author) Signed-off-by: Dong1017 <55723162+Dong1017@users.noreply.github.com>
Keep the canonical source snapshot reachable as a preceding commit. Agent-assisted-by: Codex / gpt-6-astra [effort=xhigh] (author) Signed-off-by: Dong1017 <55723162+Dong1017@users.noreply.github.com>
Codex Review SummaryThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
Skill evaluationPLANNED: not executed; there are no task outcomes yet.
Author verification steps (nonpaid; not yet run)
python tools/build_skill_index.py --source-revision fb4171e87b4c78d4d62c927f297374ed917949f5 --checkThe declared source pin matches this head's canonical library bytes. The command has not run: retain its exit code and full drift/success output. It checks the generated projection, not claim truth or skill efficacy. Evidence and logs · Retention: 30 days. Plan bindings, budget, frozen exam and advisory detailsAdvisory: obtain missing evidenceRetained evidence availability: planned / unexecuted. No approval or merge authorization is issued.
Model/settings:
Automatic execution / cumulative PR budget: {"reason":"Trusted configuration or pricing missing, invalid or expired","remaining_usd":"unknown","reserved_or_unresolved_usd":"unknown","spent_or_legacy_usd":"unknown","state":"blocked"}Projection check selection (not an executed result): {"declared_in":"src/inferrsi/data/repo_packs/inferrsi/knowledge/README.md","source_revision":"fb4171e87b4c78d4d62c927f297374ed917949f5","status":"ready"}Provider/settings: Uncovered / unsupported paths (103)
|
Dong1017
added
the
area:knowledge
ChatGPT Codex Connector
Dong1017
left a comment
Dong1017
left a comment
There was a problem hiding this comment.
LGTM for the bounded hardware-sidecar projection at 7746aea5; no blocking source-fidelity or packaging issue found.
- All 22 canonical sidecars change only applicability/recommendations. Existing applicability values are preserved, including the Hopper and Ascend video boundaries; source identities, maturity and empirical-validation fields remain unchanged.
- The new conditions and advice retain the owning procedures' hardware, ordering/lifetime, measurement and refusal distinctions. They describe source guidance and do not establish native correctness or a performance result.
- Current-head Ready and declaration checks pass. Existing author/integrator @Dong1017 should combine canonical sources before regenerating catalog/coverage and checking a new integration head. Skill Eval remains unexecuted; this is neither a merge decision nor acceptance of the overall tracker.
Evidence and scope
Head: 7746aea56298fa2f9df21f92b1a830b8eb8ea4ba; API base / assessed main: ecca7e9136d962747f42860f58e332f45b0aa358 / ecca7e9136d962747f42860f58e332f45b0aa358.
Hosted tested merge: 368ca43a97b8ccf84d43fdec86aeaaef492ef06c, with verified parents current main + this head. Ready run has successful Python 3.10/3.11 full-check and installed-package jobs plus KV task qualification. Latest declaration run succeeds; declaration compliance alone verifies neither author execution settings nor content.
Reviewed scope: All 22 changed hardware sidecars against their owning procedure contracts, the two preserved applicability declarations, canonical allowed-field diff, reachable source pin and affected catalog/coverage projections. Source snapshot a037b99a198c944a4ddbbe1d952bda1f42ab2ffa is public and the library inputs do not change between it and this head; later commits contain projections. This task inspected source and hosted evidence; it did not rerun pytest, generators, models, devices or paid evaluation.
Current Skill Eval plan: zero selected comparisons, 127 uncovered/unsupported changes, execute skipped. Explicit disposition: source/metadata software delivery only, with candidate/not_run and all empirical acceptance gaps retained. Green software checks are not skill-effect evidence.
Sibling combination is untested: #263 a559b12a1dcf1d549e801dd645a949b0f730a75f; #282 8b8ff658d10a143cabac0999900ba27d915bea19; #283 7746aea56298fa2f9df21f92b1a830b8eb8ea4ba; #285 57bc805db9bb56995e2ae988830e1c52a14f85d3; paper additions #277 e3db289c42d6d37374755a05f28b1107da4f73ae, #278 7ef53189343c4bf04fc9e31b4d80f9644482cdfc, #279 cf0b7022761fa83361f986e3b37c2031ad33fd8d. The known findings on #263/#285 are not cleared by this review.
No implementation involvement in this review task. COMMENTED because the posting account is also the PR author account; no formal self-approval is issued.
Retain all upstream source, case, metadata and matrix additions alongside the accepted hardware applicability and recommendations. Keep the source merge reachable for the subsequent generated projection commit. Agent-assisted-by: Codex / gpt-6-astra [effort=xhigh] (source integration) Signed-off-by: Dong1017 <55723162+Dong1017@users.noreply.github.com>
Pin the canonical catalog to the reachable source merge and refresh the coverage view from the same committed input bytes. Agent-assisted-by: Codex / gpt-6-astra [effort=xhigh] (generated projections and validation) Signed-off-by: Dong1017 <55723162+Dong1017@users.noreply.github.com>
Preserve all 22 hardware applicability and recommendation declarations. Retain upstream canonical sources and matrices before regenerating projections. Agent-assisted-by: Codex / gpt-6-astra [effort=xhigh] (author) Signed-off-by: Dong1017 <55723162+Dong1017@users.noreply.github.com>
Retain the DistServe integration and all 22 hardware guidance declarations. Preserve upstream canonical sources and matrices before final regeneration. Agent-assisted-by: Codex / gpt-6-astra [effort=xhigh] (author) Signed-off-by: Dong1017 <55723162+Dong1017@users.noreply.github.com>
Keep the canonical source snapshot reachable as a preceding commit. Agent-assisted-by: Codex / gpt-6-astra [effort=xhigh] (author) Signed-off-by: Dong1017 <55723162+Dong1017@users.noreply.github.com>
Preserve all 22 hardware declarations and the original applicability values. Retain every main source, matrix and historical evidence blob before regeneration. Agent-assisted-by: Codex / gpt-6-astra [effort=xhigh] (author) Signed-off-by: Dong1017 <55723162+Dong1017@users.noreply.github.com>
Preserve both histories and all canonical sources, matrices, and evidence. Resolve only generated projection conflicts before regenerating from this merge. Agent-assisted-by: Codex / gpt-6-astra [effort=xhigh] (author) Signed-off-by: Dong1017 <55723162+Dong1017@users.noreply.github.com>
Refresh the shared skill index and coverage report from the committed source. Agent-assisted-by: Codex / gpt-6-astra [effort=xhigh] (author) Signed-off-by: Dong1017 <55723162+Dong1017@users.noreply.github.com>
Dong1017
left a comment
Dong1017
left a comment
There was a problem hiding this comment.
Scoped LGTM for the hardware-sidecar integration at 14d5a51c.
- All 22 hardware sidecars are unchanged from
270fca4d; the previously reviewed field-to-procedure semantics remain applicable. The merged upstream sources are retained, and the final library tree exactly matches the reachable combined material pin. - Ready CI tests this head with assessed main
919ed42c. Both Python full-check and installed-package jobs, qualification/reference checks and current per-head policy succeed. - @Dong1017 retains the merge/integration decision. This clears the earlier current-main integration hold for this version, not a future sibling combination. Skill Eval remains planned with zero comparisons; source/metadata delivery does not establish hardware correctness, performance or overall tracker acceptance.
Evidence and limits
Reviewed head: 14d5a51c1de1171490f10fc5ed7b00c1dd914b56; API base: f58d09808eda2ab830c25dfbfbb2f24b737d03ca; assessed main: 919ed42c0d01dc0aa3b613085b9041d735d2d4db. The complete old-head delta changes no hardware canonical path; the original 22-owner semantic review is reused on that verified basis.
Material df149c73c7a8f2e90fdb7ad56a55707b95b3b4d8 and head both expose library tree 1237733439acfe11e90636e806034403ceab0348; only 79 catalog/coverage projection paths change afterwards. All 101 PR paths were inventoried, including the second files page. Ready CI checks out tested merge 4987efbc5a35d07f0eba1b2ba3daccef8c8a9727, whose verified parents are assessed main and reviewed head. Inspected Python 3.11 log: 3,381 tests and 284 subtests passed, 13 skips, one warning.
Per-head policy passes. pull_request_target Skill Eval publishes a plan and skips execution; 101 paths remain uncovered/unsupported. Verification was source and hosted-evidence inspection only, without local repository tests, generators, models/devices or paid evaluation. No implementation involvement; COMMENTED because the posting account also owns the PR.
Preserve canonical framework sources, hardware metadata, matrices, and frozen evidence. Keep generated projection resolution separate from source provenance. Agent-assisted-by: Codex / gpt-6-astra [effort=xhigh] (author) Signed-off-by: Dong1017 <55723162+Dong1017@users.noreply.github.com>
Regenerate the catalog and coverage report from the committed combined source. Agent-assisted-by: Codex / gpt-6-astra [effort=xhigh] (author) Signed-off-by: Dong1017 <55723162+Dong1017@users.noreply.github.com>
Governing Issue and delivery slice
Refs #91, Round 2 Batch 3.
The requester authorized the parallel bounded continuation. Fill source-supported applicability for 20 hardware owners and add procedure-specific advice across the frozen 22 hardware IDs. Preserve the two existing applicability declarations, including Hopper architecture and mechanism exclusions.
Only applicability/recommendations change in the selected sidecars; source bodies, provenance, maturity and empirical-validation fields are unchanged.
This branch starts from main
ffb77f9a168ba1cb22f7ba261f289d24dbb6698fand integrates main2baebf6ebb815ebf07d21c41e3b3754858c0798e, preserving the Sarathi, DistServe and Mooncake source/matrix additions and upstream evidence. Batch 2 is merged through #263 and retained in this integration; the other batches are separate deliveries. Shared projections must be regenerated from combined sources when integrating siblings, preserving all source and matrix additions. A closed applicability key allowlist remains deferred. No new source extraction, model/device execution, paid evaluation or efficacy claim is included.Checks and provenance
012bdcc8524a56c65482e14561469111736ac907; integrated main:2baebf6ebb815ebf07d21c41e3b3754858c0798e.fb4171e87b4c78d4d62c927f297374ed917949f5. A separate following commit changes only the generated catalog and coverage; the final skill-library inputs are identical to this snapshot.2baebf6ebb815ebf07d21c41e3b3754858c0798e; Python 3.11 reports 3,389 passed and 13 skipped.14d5a51c1de1171490f10fc5ed7b00c1dd914b56is reused for hardware semantics: an independent incremental inspection verified that all 22 sidecar blobs and owning bodies are unchanged, main sources are retained, and the combined source pin matches the final library tree. Local generator checks, Ruff and eight PyTorch/TorchNPU integration tests pass.These are source and software checks. Independent review applies only to its recorded revision; this description is not an approval. Skill Eval execution, efficacy, native-device qualification and whole-tracker acceptance are not claimed. Refs #91; the tracker remains open.
Frozen IDs and field-to-source review pointers
Pointers refer to unchanged source sections in the canonical snapshot above. Existing applicability values are preserved.
applicability.scope: SKILL.md:L8-L12 Trigger and boundaryapplicability.prerequisites: SKILL.md:L22-L26 Required evidence (P)applicability.input_contract: SKILL.md:L22-L24 Required evidence (P)applicability.produced_artifact: SKILL.md:L42-L46 Verification and failure handling (P)applicability.refusal_point: SKILL.md:L26 and L42-L46 Verification and failure handling (P)recommendations[0]: SKILL.md:L28-L31 Decision procedure and available actions (P)recommendations[1]: SKILL.md:L32-L33 Decision procedure and available actions (P)recommendations[2]: SKILL.md:L34-L44 Decision procedure; Verification and failure handling (P)applicability.scope: SKILL.md:L22-L28 When to use this skillapplicability.hardware: SKILL.md:L30-L62 Steps 1 and 2applicability.input_contract: SKILL.md:L30-L91 Steps 1-4applicability.produced_artifact: SKILL.md:L30-L111 Steps 1-5; Card-group change disciplineapplicability.failure_modes: SKILL.md:L62-L111 Steps 2-5; Card-group change disciplinerecommendations[0]: SKILL.md:L30-L76 Steps 1-4recommendations[1]: SKILL.md:L74-L91 Step 4recommendations[2]: SKILL.md:L95-L124 Step 5; Card-group change discipline; Environment variables checklistapplicability.scope: SKILL.md:L8-L22 and L91-L107applicability.input_contract: SKILL.md:L24-L52 Native HCCL AllReduce qualification; L91-L123applicability.produced_artifact: SKILL.md:L54-L70 and L253-L286 Attributing differences between parallel strategiesapplicability.failure_modes: SKILL.md:L171-L230 and L278-L286recommendations[0]: SKILL.md:L24-L47 and L101-L107 and L171-L181recommendations[1]: SKILL.md:L34-L52 Native HCCL AllReduce qualificationrecommendations[2]: SKILL.md:L72-L89 and L215-L230recommendations[3]: SKILL.md:L48-L52 and L109-L161 and L253-L276applicability.scope: SKILL.md:L18-L37 Establish the path before changing an optionapplicability.prerequisites: SKILL.md:L20-L23 and L59-L64 and L140-L147applicability.input_contract: SKILL.md:L20-L35 and L86-L104applicability.produced_artifact: SKILL.md:L130-L161 Discriminate failures; Validation recipe and completion boundaryapplicability.failure_modes: SKILL.md:L86-L128 Trace capture eligibility and fallback; Preserve input and output ownershiprecommendations[0]: SKILL.md:L39-L84 Use the installed configuration surfacerecommendations[1]: SKILL.md:L86-L104 and L130-L137recommendations[2]: SKILL.md:L106-L128 and L140-L154applicability.scope: SKILL.md:L14-L16 Choose the versioned ownerapplicability.prerequisites: SKILL.md:L10-L12 and L33-L34 and L99-L106applicability.input_contract: SKILL.md:L20-L65 Debug order; FP8 checkpoints; KV replication with tensor parallelism; L84-L96applicability.produced_artifact: SKILL.md:L33-L34 and L109-L124 Feature-status reportingapplicability.failure_modes: SKILL.md:L24-L65 and L99-L106recommendations[0]: SKILL.md:L20-L35 Debug order for a new checkpointrecommendations[1]: SKILL.md:L10-L12 and L37-L66recommendations[2]: SKILL.md:L84-L106 Multimodal feature validation order; Layer-reduction smoke testsapplicability.scope: SKILL.md:L8-L26 When to use this skill vs ascend-performance-analysisapplicability.prerequisites: SKILL.md:L10-L13 and L48-L80 Steps 1 and 2applicability.input_contract: SKILL.md:L71-L122 Steps 2 and 3; L304-L314 perf_summary.py usageapplicability.produced_artifact: SKILL.md:L98-L191 Steps 3 and 4applicability.failure_modes: SKILL.md:L274-L291 Known limitations and caveatsrecommendations[0]: SKILL.md:L71-L137 Steps 2 and 3recommendations[1]: SKILL.md:L146-L222 Steps 4-6recommendations[2]: SKILL.md:L282-L298 Known limitations and caveatsapplicability.scope: SKILL.md:L29-L33 Relationship to profiling; L91-L100 Staged escalationapplicability.prerequisites: SKILL.md:L29-L55 Relationship to profiling; Hard gates; Numeric goalapplicability.input_contract: SKILL.md:L37-L48 and L71-L87 and L182-L194 and L227-L233applicability.produced_artifact: SKILL.md:L182-L196 Iteration discipline; L227-L233 Verification checklistapplicability.refusal_point: SKILL.md:L29-L67 and L182-L207applicability.failure_modes: SKILL.md:L78-L87 and L200-L220recommendations[0]: SKILL.md:L37-L48 and L182-L194 and L227-L233recommendations[1]: SKILL.md:L71-L100 Metric contract; Staged escalationrecommendations[2]: SKILL.md:L211-L225 Known limitations (from source)applicability.scope: SKILL.md:L15-L25 How to use this casebook; L29-L179 Bottleneck classesapplicability.prerequisites: SKILL.md:L15-L23 How to use this casebookapplicability.input_contract: SKILL.md:L15-L23 and L186-L208 Adding a new caseapplicability.produced_artifact: SKILL.md:L15-L25 and L186-L208applicability.refusal_point: SKILL.md:L207-L208 Adding a new caserecommendations[0]: SKILL.md:L15-L25 How to use this casebookrecommendations[1]: SKILL.md:L10-L13 and L58-L60 and L91-L93 and L128-L130applicability.scope: SKILL.md:L14-L27 Accepting an automated optimization stage; L44-L46 Five-layer analysis pipelineapplicability.prerequisites: SKILL.md:L176-L184 Profiling collection prerequisitesapplicability.input_contract: SKILL.md:L16-L23 and L48-L118 and L121-L139applicability.produced_artifact: SKILL.md:L14-L27 and L142-L174applicability.failure_modes: SKILL.md:L66-L118 Five-layer analysis pipelinerecommendations[0]: SKILL.md:L44-L69 Five-layer analysis pipelinerecommendations[1]: SKILL.md:L81-L118 Layers 2 and 3recommendations[2]: SKILL.md:L14-L25 and L155-L174applicability.scope: metadata.json#/applicability/scope (existing value preserved)applicability.hardware: metadata.json#/applicability/hardware (existing value preserved)applicability.prerequisites: metadata.json#/applicability/prerequisites (existing value preserved)applicability.input_contract: src/inferrsi/data/skill_library/methods/video-diffusion-optimization/SKILL.md#1-freeze-the-goal-contractapplicability.produced_artifact: src/inferrsi/data/skill_library/methods/video-diffusion-optimization/SKILL.md#7-final-reportapplicability.failure_modes: SKILL.md:L18-L24 Ascend-Specific Layout And Profiler Referencesrecommendations[0]: SKILL.md:L10-L16recommendations[1]: SKILL.md:L18-L23 Ascend-Specific Layout And Profiler Referencesrecommendations[2]: SKILL.md:L14 and L24-L27applicability.scope: SKILL.md:L95-L213 Reduction, broadcast, elementwise, matmul and conversion routingapplicability.input_contract: SKILL.md:L30-L91 Universal four-element template; Cross-class rulesapplicability.produced_artifact: SKILL.md:L30-L48 Universal four-element template; L224-L225 Known limitationsapplicability.failure_modes: SKILL.md:L57-L84 Cross-class rules; L191-L202 MatMul routingrecommendations[0]: SKILL.md:L50-L74 Cross-class rulesrecommendations[1]: SKILL.md:L43-L48 and L62-L70recommendations[2]: SKILL.md:L75-L91 and L179-L230applicability.scope: SKILL.md:L43-L77 The chained procedure, Stages 1-3applicability.prerequisites: SKILL.md:L79-L87 Stage 4; L108-L117 Communication-specific rulesapplicability.input_contract: SKILL.md:L45-L77 and L108-L117applicability.produced_artifact: SKILL.md:L65-L106 Stages 3-6applicability.failure_modes: SKILL.md:L50-L77 and L98-L106 and L119-L129recommendations[0]: SKILL.md:L45-L63 Stages 1 and 2recommendations[1]: SKILL.md:L65-L96 Stages 3-5recommendations[2]: SKILL.md:L98-L129 Stage 6; Metric separation rulesrecommendations[3]: SKILL.md:L143-L180 Unresolved conflicts; Later-revision safeguards; OS bridgeapplicability.scope: SKILL.md:L19-L31 SearchSpace mapping; L43-L100 Controller proceduresapplicability.prerequisites: SKILL.md:L102-L116 Hierarchy, delegation, permissionsapplicability.input_contract: SKILL.md:L43-L116 Controller procedures; Hierarchy, delegation, permissionsapplicability.produced_artifact: SKILL.md:L35-L41 Prime rule; L115-L136 Verification contractapplicability.failure_modes: SKILL.md:L35-L41 and L59-L69 and L89-L100recommendations[0]: SKILL.md:L43-L57 CPU: quota versus contentionrecommendations[1]: SKILL.md:L59-L87 Cpuset; Memoryrecommendations[2]: SKILL.md:L35-L41 and L78-L87 and L129-L136recommendations[3]: SKILL.md:L89-L116 IO; Hierarchy, delegation, permissionsapplicability.scope: SKILL.md:L10-L12 Trigger and evidence gateapplicability.input_contract: SKILL.md:L14-L16 Trigger and evidence gateapplicability.produced_artifact: SKILL.md:L35-L37 Required output and proposed verificationapplicability.refusal_point: SKILL.md:L16 and L41recommendations[0]: SKILL.md:L18-L20 Decision procedurerecommendations[1]: SKILL.md:L22-L25 Decision procedurerecommendations[2]: SKILL.md:L25-L31 Decision procedure; Portability gateapplicability.scope: SKILL.md:L8-L31 Intro; Versioned Ascend native descriptor-pool caseapplicability.input_contract: SKILL.md:L33-L86 and L160-L198applicability.produced_artifact: SKILL.md:L21-L23 and L33-L75 and L200-L207applicability.failure_modes: SKILL.md:L48-L75 and L149-L158recommendations[0]: SKILL.md:L56-L75 Asynchronous transfer: three independent obligationsrecommendations[1]: SKILL.md:L33-L54 Pinned host memory: ownership before capacityrecommendations[2]: SKILL.md:L77-L86 and L143-L158 and L200-L207applicability.scope: SKILL.md:L8-L33 When to Use; Using Analysis Resultsapplicability.prerequisites: SKILL.md:L19-L33 and L87-L95 and L378applicability.input_contract: SKILL.md:L27-L33 and L59-L95 and L127-L146 and L195-L239applicability.produced_artifact: SKILL.md:L232-L239 Phase 6; L361-L370 Final Summary Outputapplicability.failure_modes: SKILL.md:L27-L50 and L107-L111 and L220-L230 and L275recommendations[0]: SKILL.md:L27-L33 and L127-L157recommendations[1]: SKILL.md:L107-L111 and L189-L230recommendations[2]: SKILL.md:L37-L50 and L220-L239 and L266-L275applicability.scope: SKILL.md:L8-L28 When to Useapplicability.prerequisites: SKILL.md:L30-L34 Prerequisites; L83-L101 Input validationapplicability.input_contract: SKILL.md:L12-L15 and L119-L128 Root Cause Phaseapplicability.produced_artifact: SKILL.md:L93-L115 and L168-L175 and L290-L296applicability.failure_modes: SKILL.md:L123-L128 and L281-L286 Pitfallsrecommendations[0]: SKILL.md:L51-L53 and L75-L79 and L123-L128 and L236-L242recommendations[1]: SKILL.md:L61-L75 and L184-L190 and L290-L296recommendations[2]: SKILL.md:L123-L128 and L281-L286applicability.scope: SKILL.md:L32-L41 Scope; L45-L108 CPU affinityapplicability.prerequisites: SKILL.md:L153-L170 Structured environment checksapplicability.input_contract: SKILL.md:L45-L170 and L172-L185applicability.produced_artifact: SKILL.md:L153-L170 and L204-L211 Verification checklistapplicability.failure_modes: SKILL.md:L65-L97 and L114-L151recommendations[0]: SKILL.md:L45-L69 CPU affinityrecommendations[1]: SKILL.md:L75-L108 and L114-L131 and L191-L202recommendations[2]: SKILL.md:L133-L170 Device-busy preflight; Structured environment checksapplicability.scope: metadata.json#/applicability/scope (existing value preserved)applicability.architectures: metadata.json#/applicability/architectures (existing value preserved)applicability.excluded_mechanisms: metadata.json#/applicability/excluded_mechanisms (existing value preserved)applicability.input_contract: SKILL.md:L20-L26 Establish the actual contractapplicability.produced_artifact: SKILL.md:L51-L63 Return a reviewable hypothesisapplicability.failure_modes: SKILL.md:L28-L42 Establish the actual contract; Choose the next discriminating checkrecommendations[0]: SKILL.md:L28-L32 Establish the actual contractrecommendations[1]: SKILL.md:L34-L42 Choose the next discriminating checkrecommendations[2]: SKILL.md:L58-L63 Return a reviewable hypothesisapplicability.scope: SKILL.md:L32-L42 Prime rule; L44-L91 Workload, concurrency and cache proceduresapplicability.prerequisites: SKILL.md:L106-L116 and L139-L159 Proposed safe procedureapplicability.input_contract: SKILL.md:L44-L54 and L69-L104 and L147-L162applicability.produced_artifact: SKILL.md:L32-L42 and L93-L116 and L147-L162applicability.refusal_point: SKILL.md:L32-L54 and L118-L127recommendations[0]: SKILL.md:L44-L67 Workload matching; Requested versus achieved concurrencyrecommendations[1]: SKILL.md:L69-L91 Cache-state accounting; Startup versus steady-state runsrecommendations[2]: SKILL.md:L93-L145 Output schema; Device attribution; Saturation-evidence gate; Process attributionrecommendations[3]: SKILL.md:L147-L162 Proposed safe procedureapplicability.scope: SKILL.md:L8-L16 Scope and evidence contract; Related capabilitiesapplicability.input_contract: SKILL.md:L18-L20 Required evidenceapplicability.produced_artifact: SKILL.md:L34-L40 Countersignals, verification and escalationapplicability.refusal_point: SKILL.md:L20-L24 and L36recommendations[0]: SKILL.md:L22-L26 Decision procedure and available actionsrecommendations[1]: SKILL.md:L28-L32 Decision procedure and available actionsrecommendations[2]: SKILL.md:L30-L38 Decision procedure; Countersignals, verification and escalationapplicability.scope: SKILL.md:L8-L16 Scope and evidence contract; Related capabilitiesapplicability.input_contract: SKILL.md:L18-L20 Required evidenceapplicability.produced_artifact: SKILL.md:L34-L40 Countersignals, verification and escalationapplicability.refusal_point: SKILL.md:L36-L40 Countersignals, verification and escalationrecommendations[0]: SKILL.md:L22-L24 Decision procedure and available actionsrecommendations[1]: SKILL.md:L26-L28 Decision procedure and available actionsrecommendations[2]: SKILL.md:L30-L32 Decision procedure and available actionsKnowledge authorship declaration