[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

[edwingao28]

Summary

• GB200/GB300 NVL72 rows get All in Measured power: measured W/GPU and ACPI Grace-socket W feed the InferenceX power_model rack at PUE 1.1; rows without complete socket telemetry use the rack's modeled Grace (cpuBasis: modeled).
• lib/power/power_model/ ports power_model/src/power_model file for file at the InferenceX#3887 merge commit 1e509932e: one .ts per .py, same class and field names; cli.py and __main__.py skipped (manifest). Closed-form profiles and their generator are removed; system-power-model.ts is a thin adapter.
• The port rejects what pydantic rejects and formats what Python formats: wrong-family parts, chassis provenance, %g tie rounding, repr escaping, duplicate detail keys and validator error classes match the reference (findings table below).
• Profit keys its power basis on measured vs modeled Grace, never interpolating across them; CSV adds cpu_basis and the port digest. Guide.
• Depends on [PowerX] unify telemetry inspection and simplify power controls / 统一遥测查看并简化功耗控件 #1220; base merged through master at 23a6f9fe. Grace collection: InferenceX#3296.

AI model disclosure

• Model/version: claude-opus-5-5 with delegated subagents; GPT-6 (variant unverified); claude-fable-5-1
• Role: Opus: power_model port, parity fuzz, tests, docs; GPT-6: earlier implementation, review, tests; Fable: Bugbot fixes, revision pin, fuzz follow-up fixes, Chinese review, PR body

Validation

Check	Scope	Result
Parity fixture (power-model.parity.json)	1,572 Python-generated cases: every system, workload, scale-out state, socket input and rejection path, plus 240 seeded random inputs; totals, component trees, provenance, configuration	all within 1e-12
Old closed form vs port	4,860 chassis and rack inputs	agree to 5e-16
GB300 replay	594.191 W/GPU, 98.066 W/socket	1,013.8 facility W/GPU (1,063.0 with scale-out), matches the upstream CLI
Local at 872c360d	vitest src/lib/power	1,745 tests pass
CI at 872c360d	Claude review, Cursor Bugbot, Vercel	pass; no test CI runs on this base, no GPU qualification

Differential fuzz against Python 1e509932e: seeded random inputs through the public API and every component class; errors compared by class and message.

Layer	Cases	Numbers compared	Outside 1e-12
Public API create_power_model(...).estimate_breakdown(...)	30,000	1,505,361	0
UBB / HGX boards and NVSwitches	20,000	333,909	0
Bianca boards, DC converter, Grace, LPDDR5X, EPYC	225,000	2,899,806	0
Generic components, fans, PSU, rack power, NICs	87,500	1,128,727	0
Networking, systems, rack scale, catalog	38,850	1,981,977	0
Total	401,350	7,849,780	0; max relative 2.2e-15, all from fan x**3 (libm pow vs x*x*x)

Behavioural findings from the fuzz, each fixed in fe8caa48 with a test in power-model-validation.test.ts:

Finding	Fix
Structural typing accepted wrong-family parts (B300 board + ConnectX-7 NIC, MI300 chassis + Hopper board, GB200 tray + GB300 board, GB300 rack + GB200 tray)	instanceof guards with pydantic's model_type message
HGX chassis dropped a caller-supplied provenance	forwarded to the system node
formatG rounded exact decimal ties half-up; Python's %g rounds half to even	exact decimal expansion, ties to even
Enum error text lacked Python repr escaping	py_repr
detail() returned the first duplicate key; dict() keeps the last	last wins
ValueError inside a model validator escaped unwrapped	surfaced as ValidationError

中文说明

• GB200/GB300 NVL72 行启用 All in Measured 功耗：实测 W/GPU 与 ACPI Grace socket 功耗输入 InferenceX power_model 的机架模型（PUE 1.1）；socket 遥测不完整的行使用机架建模的 Grace（cpuBasis: modeled）。
• lib/power/power_model/ 按文件一一移植 power_model/src/power_model（InferenceX#3887 合并提交 1e509932e）：每个 .py 对应一个 .ts，类名与字段名相同；跳过 cli.py 与 __main__.py。删除闭式 profile 及其生成器，system-power-model.ts 只是薄适配层。
• 移植版与 pydantic 的拒绝行为和 Python 的文本格式一致：异族部件、chassis provenance、%g 平局舍入、repr 转义、重复 detail 键及校验器错误类型均与参考实现相同（见上表）。
• Profit 按实测/建模 Grace 区分功耗基准，不跨基准插值；CSV 新增 cpu_basis 与移植版摘要。
• 依赖 [PowerX] unify telemetry inspection and simplify power controls / 统一遥测查看并简化功耗控件 #1220；base 已合并至 master（23a6f9fe）。

验证： parity 夹具 1,572 个 Python 生成用例（含 240 个固定种子随机输入）全部在 1e-12 内；差分 fuzz 401,350 个用例、7,849,780 个数值无一超出 1e-12；872c360d 本地 vitest src/lib/power 1,745 项通过；CI（Claude review、Cursor Bugbot、Vercel）通过，此 base 不运行测试 CI，无 GPU 验证。

---

Note

High Risk
Touches core facility-power modeling, profit GW-year capacity math, and public inference CSV/table contracts for All in Measured; incorrect rack/Grace gating or curve selection would skew economics and API rows.

Overview
Replaces the closed-form chassis power profiles with a module-for-module TypeScript port of upstream InferenceX power_model, pinned via generate-power-model-parity.py and power-model.parity.json. modelSystemPower / All in Measured now covers GB200/GB300 NVL72 racks (estimateRackPower): equal compute-tray share of a modeled 18-tray rack at PUE 1.1, using validated Grace-socket telemetry when complete or modeled Grace otherwise (cpuBasis); GB200/GB300 are no longer blocked as unsupported hardware.

PowerX / views API: All in Measured tableRows keep every GPU-valid row (including unavailable system estimates with y: null and reasons); CSV exports those rows with blanks instead of only plotted series. Boundary labels align with GPU Level Measured, All in Provisioned, All in Measured; chart copy documents dual PUE and Grace input rules.

Profit Estimator (gated): Modeled/compare builds throughput from power-valid points only at the same target (no extrapolation across invalid knots); compare pairs use that curve for both bars and fall back to provisioned when no valid curve exists. UI drops the per-SKU unavailable disclosure list; dense compare charts get horizontal scroll (PNG export still full width). CSV/tooltips add power basis and model revision; NVL72 rows label measured vs modeled Grace and never mix bases in interpolation.

Docs, offline export (cpu_basis, port digest), and Cypress/vitest coverage expanded for AgentX multinode, NVL72 empty states, profit curves, and tooltip links to the pinned model.

^{Reviewed by Cursor Bugbot for commit 5edf102. Bugbot is set up for automated code reviews on this repo. Configure here.}

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
inferencemax-app	Ready	Preview	Oct 15, 2026 11:50pm UTC

[github-actions]

Claude finished @edwingao28's task in 4m 14s —— View job

---

Review complete — 🟢 LGTM, no blocking issues found

I reviewed the full diff against origin/feat/powerx-db-ingest, traced the new logic paths, and verified the Profit Estimator renders live at http://localhost:3000/profit-estimator-per-gigawatt (GB300/GB200 NVL72 SKUs plot correctly, 0 console errors).

Todo

• Gather context (diff vs base, changed files)
• Check CI status
• Review core logic (modeled-system-power, system-power-model, profit-power, benchmark-mapper, metric-keys)
• Review UI changes (ProfitEstimator, tooltips) + runtime verification
• Review tests & docs coverage
• Review Chinese copy
• Post findings

What I verified

• Rack topology invariants (system-power-model.profiles.json): gb200/gb300 both = 4 GPUs/tray × 18 trays = 72 GPUs, 2 Grace/tray. Consistent with unitGpuCount, the socketCount === units.length × graceSocketsPerComputeTray gate, and the "rack AC ÷ 72 GPUs" normalization (chassisAcWattsPerGpu reduces to rackAc/72 regardless of units.length). ✅
• Fail-closed scrubbing (benchmark-mapper.ts): CPU-side keys are governed by cpu_power_valid independently of power_valid; scrubWithheldPowerMetrics correctly excludes the CPU subset from the GPU-side delete and vice-versa; absent verdict → withheld (no legacy rows predate it). Consistent across metric-keys.ts, power-publication.ts manifest, and the estimator's admission gate. ✅
• Planning gate (profit-power.ts): only fully-measured units (chassisBasis === 'full') become planning figures; interpolation refuses to mix two bases between knots (powerSourceKey guard); partial/extrapolated trays are correctly excluded from pricing. ✅
• Module vs. gpu-plus-grace basis: module reading is never scaled (covers whole tray incl. idle GPUs); gpu-plus-grace adds the sourced regulator-loss allowance on the GPU share only. Per-tray means divide deployment totals by units.length consistently. No division-by-zero (all units branches guarantee ≥1 element before modelTray). ✅
• Both large JSON files parse; the offline exporter's per-hardware default PUE selection (1.3 air-cooled / 1.1 DLC) matches the dashboard's defaultSystemPue, so article figures track chart hovers. ✅
• CSV/tooltip/label plumbing: column counts stay aligned when includeBasis is on (3 headers ↔ 3 basisColumns); provisioned rows (compare mode) fall back cleanly. useMemo/useCallback dep arrays updated for the new locale/powerBasisNotes reads. ✅
• Test coverage: new functions/paths all have colocated tests (profit-power.test, modeled-system-power.test, system-power-model.test, benchmark-mapper.test, tooltip-utils.test, api-documentation.power.test, metric-keys.test, plus the Cypress basis-control spec). No missing-test gaps. ✅

Chinese copy

The diff touches user-visible zh strings in ProfitEstimatorChart.tsx, ProfitEstimatorDisplay.tsx, and tooltipUtils.ts (功耗口径 / 实测 + 估算 / 稳压损耗由模型估算 / tray assumptions, etc.). All are semantically faithful and read as natural technical Chinese, with product/hardware/profile/socket/tray terms correctly kept in English per the copy guide. No fidelity or register issues — nothing requiring maintainer confirmation.

Notes (non-blocking)

• The tests-unit / tests-e2e / lint CI workflows haven't reported on this PR head yet (only Vercel + Claude route). The PR body states they pass locally; worth confirming the merge-gate checks are green before merge.
• Runtime verification of the measured NVL72 path is not possible yet — no live row carries the CPU leg, so those rows report cpu-telemetry unavailable as the PR documents. The estimator's official-frontier-only scope makes the ?unofficialrun= overlay rule correctly N/A.

This is a carefully structured, well-tested change. No changes requested.
· feat/powerx-nvl72-smart-provisioning

[github-actions]

Claude finished @edwingao28's task in 4m 31s —— View job

---

Claude Code Review — finalizing

• Gather context (diff, CI status, dev server)
• Review NVL72 rack model & planning gate (profit-power.ts, modeled-system-power.ts)
• Review gpu-metrics-point API route + measured-metric config + power-compare utils
• Runtime: Profit Estimator (NVL72 bars), boundaries chart, roles chart, power timeline — all render real data, 0 console errors
• ChartDisplay timeline routing + boundary disclosures verified
• Deep static review of large files (PowerTimeline, ScatterGraph, gpu-metrics route) — awaiting subagent
• Post findings

[github-actions]

Claude finished @edwingao28's task in 1m 48s —— View job

---

Re-review — 🟢 LGTM, no blocking issues found

Re-reviewed against origin/feat/powerx-article-parity, focused on the single new commit since the prior LGTM at 5edf1029.

• Gather context (new commits: 55e4706a revision pin + 57b80149 parent-sync merge)
• Review new/changed logic
• Check CI status
• Runtime verification
• Chinese copy pass (N/A — see below)
• Post findings

What changed since the last review

One production-relevant commit: 55e4706a "chore: pin power model port to InferenceX 69f12085e" — a pure revision-pointer update, no model math or logic touched. It moves modelRevision / sourceUrl / the generator REVISION from 5cfc833d… to the #3887 merge commit 69f12085e… while keeping sourceTree identical at ceec5a03…, so the upstream power_model tree is unchanged and no numerics move. The other new commit, 57b80149, is a parent-sync merge whose files cancel against the base.

Verified

• No numeric drift. Diffed power-model.parity.json across 5edf1029…55e4706a by parsing both revisions: the only changed key is modelRevision — every one of the ~1,572 parity cases is byte-identical. Consistent with the unchanged sourceTree. ✅
• No stale references. git grep 5cfc833d across packages/ + docs/ returns nothing; all three sites (generator REVISION, manifest.json, parity.json) move to 69f12085e together. ✅
• Pin stays internally consistent. power-model-parity.test.ts:159-160 asserts parity.modelRevision === manifest.modelRevision and manifest.sourceUrl contains manifest.modelRevision; all three were updated in lockstep, so the guard holds. ✅
• User-visible surface correct. SYSTEM_POWER_MODEL_REVISION = manifest.modelRevision and the tooltip renders estimate.modelRevision.slice(0, 12) → 69f12085e6ad, linking to SYSTEM_POWER_MODEL_SOURCE_URL (= the new commit URL). It's a plain git SHA (no app-sha256: prefix), so slice(0,12) is the whole label — no broken prefix-strip. ✅
• Runtime. /profit-estimator-per-gigawatt?c_power=modeled and /inference?i_metric=y_utilityModeledWatts&i_seq=agentic-traces both render real data, 0 console errors (1 benign pre-existing preload warning). The measured NVL72 path still can't be exercised live (no row carries the Grace-socket leg; PR depends on [PowerX] unify telemetry inspection and simplify power controls / 统一遥测查看并简化功耗控件 #1220) as documented, so it stays covered by the parity fixture + nvl72-power.fixture.ts unit/API tests.
• Chinese copy. 55e4706a touches only revision strings / a Chinese commit-body line — no user-visible /zh website string changed. N/A.

Notes (non-blocking)

• The PR body still references 1e509932e as the pinned commit in places, while the actual pin is now 69f12085e. Cosmetic — worth refreshing so reviewers aren't confused by the stale SHA.
• CI on this head: only the Claude route reports on this feature-branch PR; the authoritative tests-unit / tests-e2e / lint merge-gate workflows run on the shared-tree PR per the PR body. Confirm those are green before merge — that's the authoritative signal, not this comment.

No new issues introduced by the revision pin. No changes requested.
· feat/powerx-nvl72-smart-provisioning

[github-actions]

Claude finished @edwingao28's task in 3m 27s —— View job

---

Re-review — 🔴 One blocking issue: UI/API parity broken by the latest commit

Re-reviewed against origin/master, focused on the single new production commit since the prior LGTM at 55e4706a.

• Gather context (new commit 8a04fb18; CI status)
• Review new/changed logic
• Runtime verification (inference All-in-Measured, Best on)
• Chinese copy pass (N/A — see below)
• Post findings

What changed since the last review

One production commit: 8a04fb18 "fix: keep table-only SKUs in Best per SKU when other SKUs are ranked" — it addresses the open Cursor Bugbot "Best filter drops table-only SKUs" finding. A new helper withUnrankedSkus (best-series-per-sku.ts) is added and wired into the UI bestHwTypes memo (InferenceContext.tsx:1322) so that, under "Best per SKU", a SKU with GPU-valid table rows but no rankable estimate (e.g. an NVL72 row whose all-in estimate is unavailable) stays visible instead of being hidden by a ranked SKU elsewhere in the selection.

🔴 BLOCKING — the matching API change is missing (parity regression)

The fix was applied to the UI path only. The public views-API filter in packages/app/src/lib/views-api/series.ts:422 still reads:

.filter((point) => !best || bestHwKeys.size === 0 || bestHwKeys.has(point.hwKey))

Both paths build their rows from the same allInMeasuredTableData, so for identical inputs the dashboard table (now via withUnrankedSkus) and /api/v1/views/inference tableRows (still via bestHwKeys.has) now return different row sets under Best — the API drops exactly the table-only SKUs the UI keeps. This is the read-only-coverage parity rule in AGENTS.md ("Reuse the dashboard's pure transformation functions so the API and UI cannot silently calculate different results"), and it's precisely the case prior reviews warned about: "it would need an explicit retain-unrankable-rows carve-out in both the UI and the API with a matching regression test, to preserve parity." Changing a frontend-view filter without the matching API change is 🔴 BLOCKING on PR review.

Details + a drop-in fix are in the inline comment on series.ts:422. A views-API regression test (Best on + one ranked SKU + one table-only SKU) should land in the same change (extensions.test.ts / route.test.ts).

Verified (the rest is clean)

• UI helper logic is correct. withUnrankedSkus adds a candidate's hwKey only when its baseSku has no winner in best, so SKUs that already won keep only their winner (no losing series re-added) and chart-only unranked series are not promoted. The three new unit cases in best-series-per-sku.test.ts pin exactly this. ✅
• No crash / stale-closure risk. The memo deps are unchanged and correct; tableCandidates is derived once and reused. Runtime: /inference?i_metric=y_utilityModeledWatts&i_seq=agentic-traces&i_best=1 renders real data, 0 console errors (1 benign pre-existing preload warning). ✅
• Chinese copy: 8a04fb18 adds no user-visible /zh strings (only code + tests + comments). N/A.

Notes (non-blocking)

• CI now runs the authoritative merge gates on this head (base is master): Lint & Format green; Tests (Unit/E2E/Skills) in progress at review time — confirm green before merge.
• The measured NVL72 path still can't be exercised live (no live row carries the Grace/module leg) as documented; the table-only scenario above is reachable by construction via the retain-measured-rows path and the NVL72 fixtures.

Once the API series.ts filter is brought in line with the UI (plus a parity test), this is good to go.
· feat/powerx-nvl72-smart-provisioning

[github-actions]

Claude finished @edwingao28's task in 2m 18s —— View job

---

Re-review — 🔴 BLOCKING: the UI/API parity issue from the last review is still open

Re-reviewed against origin/master, focused on what changed since the prior 🔴 BLOCKING review at 8a04fb18.

• Gather context (new commits since prior review, CI status)
• Re-check the prior 🔴 BLOCKING finding (UI/API parity in series.ts)
• Review new/changed logic
• Runtime verification
• Post findings

What changed since the last review

Only one new commit: cd65abb1 "test: align NVL72 E2E expectations with modeled Grace sockets" — E2E-test-only (powerx-compare.cy.ts +2, profit-estimator.cy.ts +9/-4). No production code changed, so the blocking parity finding from the last review is not addressed.

🔴 BLOCKING — still open: series.ts:422 drops table-only SKUs the UI keeps

Commit 8a04fb18 fixed the UI Best-per-SKU set to retain table-only SKUs via withUnrankedSkus(best, graph.data, tableCandidates) (InferenceContext.tsx:1322), but the matching API change was never made. The views-API tableRows filter still reads:

.filter((point) => !best || bestHwKeys.size === 0 || bestHwKeys.has(point.hwKey))

Both paths build rows from the same allInMeasuredTableData, so under "Best per SKU" with at least one ranked SKU present, the dashboard table now keeps GPU-valid NVL72 rows (all-in estimate unavailable) while /api/v1/views/inference tableRows drops them — a silent UI/API divergence. Per AGENTS.md "Read-only coverage for public views", a changed frontend-view filter without the matching API change is 🔴 BLOCKING. Details + a drop-in fix (reuse withUnrankedSkus) and the required views-API regression test are in the inline comment on series.ts:422.

Verified (otherwise clean)

• UI side is correct. withUnrankedSkus adds a candidate's hwKey only when its baseSku has no winner in best, so ranked SKUs keep only their winner and table-only SKUs stay visible — matches the new best-series-per-sku.test.ts cases. The bestHwTypes memo deps are correct. ✅
• Runtime. /inference?i_metric=y_utilityModeledWatts&i_seq=agentic-traces&i_best=1 renders real data, 0 console errors (1 benign pre-existing preload warning). The measured NVL72 path still can't be exercised live (no row carries the Grace/module leg); the table-only scenario is reachable by construction via the retain-measured-rows path + NVL72 fixtures. ✅

Notes (non-blocking)

• CI on this head (base master): Lint & Format green; Tests (Unit/E2E/Skills) in progress at review time — confirm green before merge.

Once the API series.ts filter is brought in line with the UI (plus a parity test), this is good to go. Everything else in the PR was cleared in prior comprehensive reviews.

[github-actions]

Claude finished @edwingao28's task in 2m 16s —— View job

---

Re-review — 🔴 One blocking issue: stale contract digest fails Tests (Unit)

Re-reviewed against origin/master, focused on the prior 🔴 BLOCKING UI/API parity finding and the new commit fc1cd29c that addresses it.

• Gather context (new commit fc1cd29c)
• Verify the series.ts Best-filter parity fix + regression test
• Check CI status
• Runtime verification (inference All-in-Measured, Best on)
• Post findings

The prior UI/API parity finding is fixed ✅

fc1cd29c brings the views-API tableRows Best filter in line with the dashboard: series.ts:427-433 now computes tableHwKeys = withUnrankedSkus(bestHwKeys, mapped, allInRows) and filters on it, reusing the same helper the UI calls in InferenceContext.tsx:1322 (withUnrankedSkus(best, graph.data, tableCandidates)). Both derive best from bestSeriesPerSku(…, direction) over the same ranked point set and fall through to "keep everything" when !best/bestHwKeys.size === 0, so the two paths can no longer diverge. The new regression test (series.test.ts, "keeps table-only all-in SKUs under best=true like the dashboard") pins exactly the scenario the earlier reviews required — ranked winner h200_trt only in series, but both gb200_trt (table-only) and h200_trt in tableRows. Good.

🔴 BLOCKING — the digest for the edited series.ts was not bumped

Tests (Unit) is red on this head. fc1cd29c changed src/lib/views-api/series.ts but left its sourceSha256 in api-route-catalog.ts:1399 stale, so the contract-sync guard fails:

src/lib/views-api/series.ts changed. … update its shared-source SHA-256 digest in src/lib/api/api-route-catalog.ts.
Expected: "d904952b…"  (recorded)
Received: "8697fb60…"  (actual)

I recomputed the hash locally — the current series.ts is 8697fb60…, confirming the recorded digest is stale. Per AGENTS.md "API reference and contract synchronization", confirm whether the human reference / OpenAPI / Chinese copy need updates (the tableRows best=true behavior now retains table-only SKUs), then bump the digest. Drop-in fix in the inline comment on api-route-catalog.ts:1399.

Runtime verification ✅

/inference?i_metric=y_utilityModeledWatts&i_seq=agentic-traces&i_best=1 renders real data, 0 console errors (1 benign pre-existing preload warning).

CI status

• Lint & Format ✅ · Tests (Unit) 🔴 (the digest guard above) · Tests (E2E) / Tests (Skills) in progress.

Once the series.ts digest in api-route-catalog.ts is refreshed (and any docs it flags are reviewed), Tests (Unit) should go green and the PR is good to go — the parity fix itself is correct and tested.

[github-actions]

Claude finished @edwingao28's task in 1m 36s —— View job

---

Re-review — 🟢 LGTM, both prior BLOCKING findings resolved

Re-reviewed against origin/master, focused on the single new commit since the prior 🔴 BLOCKING review at fc1cd29c.

• Gather context (new commit 2d194554; CI status)
• Verify the stale-digest BLOCKING finding is resolved
• Verify the UI/API parity fix + regression test
• Check CI status
• Post findings

Both prior 🔴 BLOCKING findings are fixed

1. UI/API parity (series.ts) — fixed in fc1cd29c. The views-API tableRows Best filter now computes tableHwKeys = withUnrankedSkus(bestHwKeys, mapped, allInRows) (only when best && bestHwKeys.size > 0) and filters on it (series.ts:425-434), reusing the same withUnrankedSkus helper the UI calls in InferenceContext.tsx:1321. Both paths derive best from bestSeriesPerSku(…, direction) over the same ranked set and fall through to "keep everything" when !best/bestHwKeys.size === 0, so the two can no longer diverge. The regression test (series.test.ts:161 "keeps table-only all-in SKUs under best=true like the dashboard") pins exactly the scenario the earlier reviews required. bun vitest run src/lib/views-api/series.test.ts → 27 passed locally. ✅

2. Stale contract digest — fixed in fc1cd29c/2d194554. I recomputed sha256(series.ts) locally = 8697fb60…, which now matches the recorded sourceSha256 in api-route-catalog.ts:1399. bun vitest run src/lib/api/api-route-catalog.test.ts → 14 passed. The Tests (Unit) gate that was red on the prior head should now go green. ✅

What 2d194554 adds (the only new commit)

A docs-only follow-up to the guard message: it bumps the series.ts digest and documents the best=true table-only-SKU behavior in views-api/docs/inference.ts (tableRows description + the operation summary, EN+ZH). docs/inference.ts is the documentation registry (not a digested shared-contract source), so no further digest bump is required. The added Chinese — （没有可排名估算的 SKU 保留其全部观测点） — is semantically faithful and reads as natural technical Chinese, with SKU/optimal/best kept in English per the copy guide. No fidelity/register issues.

CI status

• Lint & Format ✅ · Tests (Unit / E2E / Skills) in progress at review time (base is master, so the authoritative gates run here) — the two targeted suites I ran locally (api-route-catalog, series) both pass, so the digest guard failure from the last review is cleared.

Once Tests (Unit) reports green on this head, this is good to merge — the parity fix is correct and tested, and the digest is now in sync. Everything else in the PR was cleared in prior comprehensive reviews.