[PowerX] enable NVL72 smart provisioning from measured GPU and Grace socket power / 基于实测 GPU 与 Grace socket 功耗为 NVL72 启用智能预配

.prettierignore

@@ -8,3 +8,6 @@
 packages/app/cypress/fixtures/api/collectivex-latest.json
 packages/app/cypress/fixtures/api/collectivex-run-*.json
 packages/app/cypress/fixtures/api/collectivex-runs.json
+# The power-model parity fixture is upstream Python output, regenerated wholesale by
+# packages/app/scripts/generate-power-model-parity.py; pretty-printed it is ~5x larger.
+packages/app/src/lib/power/power-model.parity.json

docs/dashboard-readonly-views.md

@@ -146,18 +146,45 @@ artifacts.
 Power boundary labels are GPU Level Measured, GPU Level Provisioned (TDP), All in
 Provisioned, and All in Measured. The last combines measured GPU power with modeled
 unmeasured components and PUE; it is not a wall-meter measurement. These labels and
-collapsed power-assumption/availability notes do not change metric IDs, API selectors,
+power-assumption notes do not change metric IDs, API selectors,
 or calculations. Profit comparison `powerLabel` display text follows the same names.
-GPU statistics include startup and warmup for all chips in the selected series,
-regardless of chip visibility. They are separate from serving-window power and
-J/token. Run telemetry comes from one source per run: stored series once the run is
-ingested, otherwise its artifacts. Stored `coverage` describes the whole run, even
-when one host is selected: `complete` means the recorded artifact/file inventory
-is present, `partial` means recorded telemetry is missing, and `unknown` means a
-complete inventory was not retained. `missingArtifacts` lists entirely missing
-artifacts; it may be empty when only a file is incomplete.
-GitHub `coverage` is `null` (unassessed). The public view returns private, no-store
-responses and preserves upstream 503 failures.
+All in Measured watts and energy accept validated `single_turn` and AgentX rows
+through the shared chart/API transform, including historical and unofficial rows.
+AgentX uses the power model's agentic workload state without independent workload
+calibration. Telemetry and topology gates still apply; each NVL72 compute tray takes an
+equal share of a modeled rack, with measured Grace-socket power when the row has complete
+Grace-socket telemetry and the modeled Grace profile otherwise.
+
+For All in Measured, `tableRows` retains every GPU-valid observation in the selected
+scope and best-series selection, including axis-clipped and non-frontier points.
+Missing system estimates use `y: null`, `status: "unavailable"`, and
+`unavailableReason`; `measuredGpuWatts` remains available. CSV exports these rows
+with blank missing values. Numeric `series` and `count` are unchanged. Each date
+comparison and unofficial overlay has its own `tableRows`; latest does not pool history.
+
+For GW-year profit, `modeled` selects valid system-power points before building the
+curve at the same target, without extrapolation or another snapshot. `compare` uses
+the same valid-curve throughput for both power budgets, retaining the original
+provisioned estimate when no valid curve covers the target. Provisioned-only keeps
+the original performance curve. Official, comparison and unofficial scopes remain
+independent; both interpolation knots must share a power basis (including measured or modeled NVL72 Grace power), model revision and PUE.
+
+Dense profit charts reserve readable space per bar and scroll within the plot on narrow
+screens; captions and controls stay fixed. This is presentation-only: API selectors,
+calculations, source identities and CSV rows are unchanged. PNG export includes the full
+plot regardless of its current scroll position, so no API or skills contract change is needed.
+Profit charts omit the per-configuration unavailable-estimate list. This is presentation-only;
+the API retains skipped rows and reasons, and calculations and CSV exports are unchanged.
+The GPU statistics table includes startup and warmup for all chips in the selected
+series, regardless of chip visibility. It is separate from serving-window power,
+J/token and selected-time-window calculations. Run telemetry comes from one source
+per run: stored series once ingested, otherwise its artifacts. Stored `coverage`
+describes the whole run, even when one host is selected: `complete` means the recorded
+artifact/file inventory is present, `partial` means recorded telemetry is missing, and
+`unknown` means a complete inventory was not retained. `missingArtifacts` lists entirely
+missing artifacts; it may be empty when only a file is incomplete. GitHub `coverage` is
+`null` (unassessed). The public view returns private, no-store responses and preserves
+upstream 503 failures.
 
 Run-specific recognition labels are also presentation-only. Runs `35879254139`
 and `37181045340` display `UMBP MoRI SGLang` through October 9, 2026 in America/New_York
@@ -285,8 +312,7 @@ point identity. Fewer than three distinct output rates return `fit: null` with
 power, and R² is null when power did not vary.
 
 These analytical results are JSON-only: `format=csv` with any analysis enabled returns
-400, rather than silently exporting only the primary chart. Ordinary CSV retains
-its existing plotted-point contract.
+400, rather than silently exporting only the primary chart. CSV exports plotted points except for All in Measured, which exports `tableRows`.
 
 | Surface                                        | Dashboard control / share parameter               | Read-only API coverage                                                    |
 | ---------------------------------------------- | ------------------------------------------------- | ------------------------------------------------------------------------- |
@@ -342,6 +368,17 @@ those properties.
 @semianalysisai/inferencex-skills 包。所有仪表板路由（含隐藏和功能开关控制的
 视图）均在覆盖表中登记；上表列出各只读接口接受的全部查询参数名。
 接口复用现有计算函数，公开运行与非官方叠加数据保留各自来源。
+整体实测指标的 `tableRows` 保留当前筛选范围和最优曲线选择内所有 GPU 遥测有效的观测点，
+不按前沿或坐标轴显示范围裁剪。估算不可用时 `y` 为 null，`status` 为 `unavailable`，
+`unavailableReason` 给出原因；`measuredGpuWatts` 保留实测 GPU 功耗。CSV 导出同一组行，缺失值留空。
+`series` 和 `count` 保持不变，仍只包含可绘制的数值点；各日期对比和非官方叠加分别返回自己的 `tableRows`，
+Latest 不会合并历史数据。
+
+按 GW 年估算利润时，modeled 先筛选满足系统功耗要求的数据点，再在原目标值上构建曲线，
+不外推，也不借用其他快照。compare 的两种功耗方案使用同一条有效曲线的吞吐量；
+没有有效曲线覆盖目标时，保留原曲线的预配估算。provisioned 单独使用时沿用原性能曲线。
+官方数据、日期对比和非官方叠加各自独立计算；NVL72 Grace socket 遥测要求以及模型版本/PUE 一致性要求同样不变。
+
 私有上传、密钥、提示词、反馈及管理操作不作为公开读取接口。
 OperatorX 的入口受功能开关控制，页面使用专属的 `/api/v1/operatorx/*`
 接口；目前没有发布 `/api/v1/views/operatorx` 契约。

docs/data-transforms.md

@@ -94,8 +94,8 @@ Returns `{ chartData: InferenceData[][], hardwareConfig: HardwareConfig }`.
 | B4 utility modeled (measured → IT → PUE) | `modeledSystemPower.deploymentFacilityWatts ÷ modeledSystemPower.gpuCount` | `B1 J/out × (B4 W ÷ B1 W)`         | `utilityModeledWatts`, `utilityModeledJPerOutputToken`             |
 
 - **N_alloc and total throughput** (`powerBasisNormalization`). Aggregate rows report output per allocated GPU, so `N_alloc` cancels and the builder uses `W ÷ output_tput_per_gpu` without trusting display counts (legacy ingest can encode TP × EP twice). Fixed-sequence disaggregated rows (`disagg && benchmark_type === 'single_turn'`) report output per decode GPU while the deployment also powers the prefill pool, so `total output tok/s = output_tput_per_gpu × num_decode_gpu` and `N_alloc = num_prefill_gpu + num_decode_gpu`; the article's 4P+4D counts eight GPUs, and B3 energy is `(P + D) / D` times `jOutput` on such rows. Other disaggregated benchmark types emit no provisioned energy because it is not verifiable in-app whether AgentX throughput already divides by all GPUs.
-- **B4 source.** Reuses the `SystemPowerEstimate` that `rowToAggDataEntry` attached as `entry.modeledSystemPower`; nothing re-runs `modelSystemPower`. `deploymentFacilityWatts` is modeled IT power (chassis AC plus its scale-out network share) × PUE with PUE applied exactly once inside `estimateChassisPower`, divided by the physical measured GPU count (not `modeledGpuCount`, which over-counts partially allocated chassis: a 4P+4D deployment on two worker hosts models 16 GPUs while measuring 8, and the unit test pins that divisor). Modeled energy scales the producer's same-window `joules_per_output_token` by `B4 W ÷ avg_power_w`, so it inherits B1's token denominator and survives rows whose `output_tput_per_gpu` is missing (the provisioned energies do not; a per-basis point count cannot assume one shared denominator).
-- **Null rules.** A value is emitted only when finite and positive; otherwise the key is omitted (never `{ y: 0 }`), because the metric filters drop points by `metricKey in point` and `remapInferencePoint` falls back to raw throughput when a key exists with an unusable value. B2/B3 W are absent for hardware without registry specs (`getGpuSpecs` returns zeros); their energies are also absent without output throughput or disaggregated counts. B4 requires `modeledSystemPower.status === 'supported'` and B1 watts (`avg_power_w` after `rowToAggDataEntry`'s `power_valid !== 0` admission); B4 energy additionally needs `joules_per_output_token`. Telemetry admission belongs to `modelSystemPower`, which accepts `power_valid === 1` with schema v2 or the validated unversioned single-node producer (`telemetryBasis: 'validated-unversioned-single-node'`), so B4 shares B1's telemetry admission; benchmark types without an upstream workload state (`reason: 'workload'`), unsupported hardware such as GB200/GB300 NVL72 (`reason: 'hardware'`), and telemetry/topology failures withhold it. The public API's `strictV2` row filter is not re-applied in the chart for B1, so it is not re-applied for B4 either (plan §3.1 describes B1 with that filter; the app's chart path is the authority here).
+- **B4 source.** Reuses the `SystemPowerEstimate` that `rowToAggDataEntry` attached as `entry.modeledSystemPower`; nothing re-runs `modelSystemPower`. `deploymentFacilityWatts` is modeled IT power (chassis or rack AC plus its scale-out network share) × PUE with PUE applied exactly once inside `estimateChassisPower` or `estimateRackPower`, divided by the physical measured GPU count (not `modeledGpuCount`, which over-counts partially allocated chassis: a 4P+4D deployment on two worker hosts models 16 GPUs while measuring 8, and the unit test pins that divisor). Modeled energy scales the producer's same-window `joules_per_output_token` by `B4 W ÷ avg_power_w`, so it inherits B1's token denominator and survives rows whose `output_tput_per_gpu` is missing (the provisioned energies do not; a per-basis point count cannot assume one shared denominator).
+- **Null rules.** A value is emitted only when finite and positive; otherwise the key is omitted (never `{ y: 0 }`), because the metric filters drop points by `metricKey in point` and `remapInferencePoint` falls back to raw throughput when a key exists with an unusable value. B2/B3 W are absent for hardware without registry specs (`getGpuSpecs` returns zeros); their energies are also absent without output throughput or disaggregated counts. B4 requires `modeledSystemPower.status === 'supported'` and B1 watts (`avg_power_w` after `rowToAggDataEntry`'s `power_valid !== 0` admission); B4 energy additionally needs `joules_per_output_token`. Telemetry admission belongs to `modelSystemPower`, which accepts `power_valid === 1` with schema v2 or the validated unversioned single-node producer (`telemetryBasis: 'validated-unversioned-single-node'`), so B4 shares B1's telemetry admission; benchmark types without an upstream workload state (`reason: 'workload'`), unsupported hardware (`reason: 'hardware'`), and telemetry/topology failures withhold it. GB200/GB300 NVL72 rows take an equal compute-tray share of `estimateRackPower`, the pinned `power_model` rack evaluated at the measured W/GPU and, when the row carries complete, matching Grace-socket telemetry from the same window, the measured W per Grace socket; otherwise the rack's modeled Grace profile (`unit: 'nvl72-tray'`, `cpuBasis: 'measured' | 'modeled'`). The public API's `strictV2` row filter is not re-applied in the chart for B1, so it is not re-applied for B4 either (plan §3.1 describes B1 with that filter; the app's chart path is the authority here).
 - **Ordering invariant** (unit-tested on real B200 telemetry): B3 ≥ B4 ≥ B1 and B2 ≥ B1 for both W/GPU and J/out on the same point.
 - **Reconstructed prefill energy** (`reconstructedPrefillJPerOutputToken`, `utils/role-energy.ts`). For validated (`power_valid === 1`, schema 2) disaggregated rows, `prefill_joules_per_input_token × (joules_per_output_token ÷ joules_per_input_token)` carries the prefill pool's energy onto the output-token axis: the ratio is the served input:output token count because schema-2 aggregate energy has one numerator. With `decode_joules_per_output_token` it sums back to the deployment's J/out. It feeds only the `i_pcompare=roles` comparison on the energy axis (PowerX Figure 7) and is never a y-axis of its own; aggregate rows and rows missing any of the four inputs omit it.
 - Historical Trends substitutes `output_tput_per_gpu := tput_per_gpu` for legacy rows lacking output throughput; provisioned energies in trends inherit that fallback.

docs/index.md

@@ -13,7 +13,7 @@ Design rationale and non-obvious conventions. See [CLAUDE.md](../CLAUDE.md) for
 - [Pareto Boundary API](./pareto-api.md): Query frontier and hinterland observations, preserve provenance, and distinguish API scope from chart highlights.
 
 - [API Skill Examples](./inferencex-api-examples.md) — Install the public skill, query benchmarks, export measured PowerX data, and explain empty results
-- [PowerX System Power](./powerx-system-power.md) — Pinned chassis model, measured-input guards, assumptions, and reproducible article exports
+- [PowerX System Power](./powerx-system-power.md) — Pinned chassis and NVL72 rack models, NVL72 Grace-socket gate, measured-input guards, assumptions, and reproducible article exports
 - [PowerX Permanent View](./powerx-permanent-view.md) — Power boundaries as gated Measured Energy metrics, `i_metric`/`i_rulers` share links, missing-value states
 - [API Skill Releases](./inferencex-skills-release.md) — Prepare an immutable package, verify clean installations and agent exports, and publish through the package-specific workflow
 - [API Skill Discovery](./inferencex-skills-discovery.md) — Accept or reject implicit skill discovery in fresh Codex and Claude Code projects

docs/powerx-permanent-view.md

@@ -23,19 +23,20 @@ group stays out of the selector otherwise. The boundary metrics are members of t
 
 ## Boundaries
 
-| Basis (`PowerBasis`)  | Selector label               | W / GPU metric                               | J / output token metric                           | Source                                                                |
-| --------------------- | ---------------------------- | -------------------------------------------- | ------------------------------------------------- | --------------------------------------------------------------------- |
-| `gpu-measured`        | GPU measured                 | `y_measuredAvgPower` (+P75/P90, roles, %TDP) | `y_measuredJPerOutputToken` (+ input/total/query) | runner telemetry; existing metrics, unchanged                         |
-| `gpu-provisioned`     | GPU provisioned (TDP)        | `y_gpuProvisionedWatts`                      | `y_gpuProvisionedJPerOutputToken`                 | `HW_REGISTRY.tdp`                                                     |
-| `utility-provisioned` | Utility provisioned (all-in) | `y_utilityProvisionedWatts`                  | `y_utilityProvisionedJPerOutputToken`             | `HW_REGISTRY.power` (all-in kW per GPU)                               |
-| `utility-modeled`     | Utility modeled (PUE)        | `y_utilityModeledWatts`                      | `y_utilityModeledJPerOutputToken`                 | `modelSystemPower` IT power × PUE 1.3 (applied once), ÷ measured GPUs |
+| Basis (`PowerBasis`)  | Selector label              | W / GPU metric                               | J / output token metric                           | Source                                                                                           |
+| --------------------- | --------------------------- | -------------------------------------------- | ------------------------------------------------- | ------------------------------------------------------------------------------------------------ |
+| `gpu-measured`        | GPU Level Measured          | `y_measuredAvgPower` (+P75/P90, roles, %TDP) | `y_measuredJPerOutputToken` (+ input/total/query) | runner telemetry; existing metrics, unchanged                                                    |
+| `gpu-provisioned`     | GPU Level Provisioned (TDP) | `y_gpuProvisionedWatts`                      | `y_gpuProvisionedJPerOutputToken`                 | `HW_REGISTRY.tdp`                                                                                |
+| `utility-provisioned` | All in Provisioned          | `y_utilityProvisionedWatts`                  | `y_utilityProvisionedJPerOutputToken`             | `HW_REGISTRY.power` (all-in kW per GPU)                                                          |
+| `utility-modeled`     | All in Measured             | `y_utilityModeledWatts`                      | `y_utilityModeledJPerOutputToken`                 | `modelSystemPower` IT power × system PUE (1.3 chassis, 1.1 NVL72; applied once), ÷ measured GPUs |
 
 Formulas, the all-GPU normalization (`N_alloc` = prefill + decode GPUs for disaggregated
 rows) and the null rules are specified in
-[Data Transforms → Power boundaries](./data-transforms.md#power-boundaries); the chassis model
-itself in [PowerX System Power](./powerx-system-power.md). The ungated `jOutput` keeps its
-per-decode-GPU normalization; its labels and the boundary metric's `all GPUs` label keep the
-two distinguishable in the selector, the availability list and CSV headers.
+[Data Transforms → Power boundaries](./data-transforms.md#power-boundaries); the chassis and
+NVL72 tray models themselves in [PowerX System Power](./powerx-system-power.md). The ungated
+`jOutput` keeps its per-decode-GPU normalization; its labels and the boundary metric's
+`all GPUs` label keep the two distinguishable in the selector, the availability list and CSV
+headers.
 
 Every boundary metric has `polarity: 'lower'`. Energy metrics therefore get the usual
 lower-is-better Pareto frontier. The three watt keys are listed in `POWER_CURVE_METRICS`