[Power] role-local measured energy axes + per-worker power drilldown | 前端：新增角色级实测能耗坐标轴与每 worker 功耗下钻

[edwingao28]

Summary

This PR makes role-specific measured power data usable in the InferenceX frontend:

• Adds Measured Prefill Joules per Input Token and Measured Decode Joules per Output Token to the Measured Energy y-axis group.
• Adds a pinned-tooltip Measured Worker Power table for official results, unofficial-run overlays, and GPU/date comparisons when a row contains workers[].
• Shows a specific bilingual empty-state explanation when the selected dataset does not report role-level prefill/decode energy, instead of suggesting that the chart is generically broken.
• Keeps metric switching non-destructive: configurations without the selected role metric leave the rendered chart and legend, then return when the user switches back to a metric they report.

Data availability

Current role-energy coverage is sparse. The preview has real role-level points for selected DeepSeek V4 Pro 8K/1K FP4 and GLM-5.2 AgentX FP8 runs, but most historical rows do not contain the two role-energy scalars.

Current AgentX rows also do not contain workers[], so the AgentX worker drilldown is waiting for producer-side emission. The same tooltip path can be verified today on fixed-sequence historical disaggregated rows, including DeepSeek R1 MI355X results. The UI does not derive role energy from worker watts when the producer has not published the role-energy fields.

Validation

• bun run typecheck ✅
• bun run lint ✅
• bun run fmt ✅
• App unit suite: 281 files / 4,727 tests ✅
• Focused component coverage: ScatterGraph and GPUGraph, 35/35 tests ✅
• English desktop and Simplified Chinese 375 px empty states rendered and checked ✅
• GitHub CI re-runs the complete unit and browser matrix on this commit.

中文说明

本 PR 让 InferenceX 前端可以展示并解释按角色拆分的实测功耗数据：

• 在“实测能耗”纵轴分组中新增“每输入 token 实测 Prefill 能耗”和“每输出 token 实测 Decode 能耗”。
• 当数据行包含 workers[] 时，在点击固定的 tooltip 中展示“各 worker 实测功耗”表格；官方结果、非官方运行叠加层和 GPU/日期对比均支持。
• 如果当前数据集未提供 Prefill/Decode 各角色的能耗数据，空状态会明确说明缺少这类数据，不再让用户误以为图表出现了通用故障。
• 切换指标不会破坏芯片选择：缺少当前角色指标的配置会暂时从图表和图例中移除，切回其支持的指标后会自动恢复。

数据覆盖情况

目前按角色拆分的能耗数据仍然较少。预览环境中，部分 DeepSeek V4 Pro 8K/1K FP4 和 GLM-5.2 AgentX FP8 运行包含真实数据点，但大多数历史数据行尚未提供这两个角色能耗字段。

当前 AgentX 数据行也尚未包含 workers[]，因此 AgentX 的 worker 下钻仍需等待生产端输出。现阶段可以使用固定序列的历史分离式推理数据验证同一套 tooltip 路径，例如 DeepSeek R1 的 MI355X 结果。如果生产端未发布角色能耗字段，前端不会根据 worker 功耗自行推算角色能耗。

验证

• bun run typecheck ✅
• bun run lint ✅
• bun run fmt ✅
• App 单元测试：281 个文件 / 4,727 项测试 ✅
• ScatterGraph 与 GPUGraph 定向组件测试：35/35 ✅
• 已实际检查英文桌面端和 375 px 简体中文空状态 ✅
• GitHub CI 将在本次提交上重新运行完整的单元测试和浏览器矩阵。

[vercel]

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

Project	Deployment	Actions	Updated (UTC)
inferencemax-app	Ready	Preview	Aug 31, 2026 10:24pm

[github-actions]

Claude finished @edwingao28's task in 3m 5s —— View job

---

Review: PR #935 — role-local measured energy axes + per-worker power drilldown

• Read the diff against feat/measured-power-tiering
• Review new registry axes + chart field mapping
• Review generateWorkerPowerHTML (escaping, unit conversion, caps)
• Review coverage filtering / legend behavior
• Check test coverage
• Check CI status
• Runtime verification (Playwright)

🟢 LGTM — no blocking issues found

Static review + runtime verification both clean. This is a tight, well-scoped, well-tested addition.

What I verified statically

• Registry wiring — both new keys (measuredPrefillJPerInputToken, measuredDecodeJPerOutputToken) carry bilingual labels/titles, polarity: 'lower', METRIC_EXPLANATIONS entries with the tier note, and membership in MEASURED_ENERGY_METRIC_CONFIG_KEYS (so the tier ring / tooltip line / footer key apply automatically). tokenMetricTypeForConfigKey correctly classifies them input/output via substring — no output substring collision on the prefill key.
• Transform — role-local energy and workers[] are gated on measuredPowerValid (benchmark-transform.ts:237-242, :253), so power_valid: 0 scrubs them consistently with the role-watts fields. Gating role-local joules on power_valid only (not hasWholeDeploymentEnergySemantics) is correct — the role-prefixed fields have unambiguous role-local semantics and never had the whole-deployment drift problem. This is the shared transform, so the overlay/?unofficialrun= path is covered too.
• Tooltip XSS — w.role and w.hosts (artifact JSONB) are HTML-escaped before interpolation (escapeHtml, tooltipUtils.ts); the remaining interpolated values are typed numbers. isPinned is in scope in all three generators (official/overlay/GPU-compare) and correctly gates rendering; 8-row cap + +N more overflow logic is correct (no line at exactly 8).
• avg_mem_used_mb / 1024 → GiB — the MiB-despite-_mb assumption is explicitly commented as a re-check item for PLAN-09's producer. Fine to ship as documented.

Runtime (Playwright, real read-only DB)

• Both new options render under the Measured Energy group in the y-axis selector.
• Selecting Measured Prefill Joules per Input Token produces no console errors; the chart coverage-filters to an empty state (the default agentic model in prod has no role-local energy yet — the expected graceful path). The synthetic Cypress fixtures exercise the populated render.

Tests — comprehensive: transform pass-through + scrub, chart-field emission/absence/zero-preservation, Pareto corners + bilingual labels, token-type classification + share-link round-trip, tooltip table across all three generators (pinned/unpinned, empty/absent, 8-row cap, ZH, HTML escaping), and a new AgentX e2e spec covering both axes, non-destructive coverage filtering, the worker drilldown, and graceful absence.

Note: stacked PR — base is feat/measured-power-tiering (#934), which must land first.
· feat/worker-power-drilldown

[github-actions]

Claude finished @edwingao28's task in 1m 57s —— View job

---

Re-review: PR #935 — role-local measured energy axes + per-worker power drilldown

• Read the diff against feat/measured-power-tiering
• Identify new commits since the last review (ce14eec)
• Review new registry axes + chart field mapping
• Review generateWorkerPowerHTML (escaping, units, caps)
• Review coverage filtering / legend behavior
• Check test coverage
• Check CI status
• Runtime verification (Playwright)

🟢 LGTM — no blocking issues found

The only change since my last review (ce14eec) is the merge commit 393d5b36 ("merge measured-power-tiering, resolve tooltip conflicts"). I focused this pass there.

What the merge changed in scope

• The conflict resolution in tooltipUtils.ts cleanly integrates master's tooltip localization (Chinese dates via formatTooltipDate, DPA: 是/否, worker 数, GitHub Actions 运行记录, runLinkHTML(runUrl, locale)) alongside this PR's generateWorkerPowerHTML. All three generators (official / overlay / GPU-compare) still call generateWorkerPowerHTML(d, isPinned, locale) with isPinned in scope; the escaping, 8-row cap + +N more, and MiB→GiB conversion are unchanged from the version I already approved.
• No conflict markers, and the feature files carried by this PR (metric-registry.ts, chart-utils.ts, types.ts) are untouched by the merge — the two new axes (measuredPrefillJPerInputToken / measuredDecodeJPerOutputToken), their MEASURED_ENERGY_METRIC_CONFIG_KEYS membership, and buildMeasuredPowerChartFields mapping remain exactly as reviewed.
• The Chinese tooltip strings introduced by the merge (数据并行注意力 (DPA), worker 数, etc.) originate from the base/master side, not this PR's authored work; they read naturally and are consistent with the established register — nothing to flag.

Runtime (Playwright, real read-only DB)

• /inference renders with 0 console errors after the merge — the build is healthy, no regression from the conflict resolution. The two new axes and worker-drilldown behavior were already verified in the prior LGTM pass; the merge introduces no UI-surface change requiring re-verification (worker table needs a pinned point with populated workers[], which prod AgentX rows still don't ship — the synthetic Cypress fixtures cover that path).

CI — Claude (route) pass, Vercel pass; tests-* / lint / Cursor Bugbot still pending at review time (nothing failing). No new code was added without tests.

Nothing blocking. Reminder for the merge order: this is a stacked PR — base feat/measured-power-tiering (#934) must land first; do not squash into master before it.
· feat/worker-power-drilldown

[github-actions]

Claude finished @edwingao28's task in 1m 15s —— View job

---

Re-review: PR #935 — role-local measured energy axes + per-worker power drilldown

• Diff against feat/measured-power-tiering (13 files, +668/-1 — matches PR)
• Identify new commits since last review (393d5b36)
• Re-check core feature files (registry axes, chart fields, tooltip)
• Check CI status
• Runtime verification (Playwright)

🟢 LGTM — no blocking issues found

The only new commit since the last LGTM (393d5b36) is the base merge 225163cb ("chore: merge updated measured-power base"), which pulls the refreshed feat/measured-power-tiering base in. The PR's own diff against the base is unchanged — 13 files, +668/−1, identical to what was already reviewed. The merge resolved cleanly: no conflict markers, and this PR's authored files (metric-registry.ts, chart-utils.ts, types.ts, tooltipUtils.ts) are untouched by the merge itself.

Independently re-verified in this pass

• Registry axes — measuredPrefillJPerInputToken / measuredDecodeJPerOutputToken carry bilingual labels/titles, polarity: 'lower', and membership in MEASURED_ENERGY_METRIC_CONFIG_KEYS (so the tier ring / tooltip line / footer key apply). No output substring collision on the prefill key nor input on the decode key — tokenMetricTypeForConfigKey classifies each correctly.
• Chart fields (chart-utils.ts:575) — role-local joules are mapped straight through chartMetric(...); the comment correctly notes rowToAggDataEntry already scrubbed them on power_valid, so no double-gating.
• Tooltip XSS — w.role and w.hosts (artifact JSONB) are escapeHtml-escaped before interpolation; remaining interpolated values are typed numbers. generateWorkerPowerHTML(d, isPinned, locale) is gated on isPinned && workers?.length in all three generators (official / overlay / GPU-compare), with the 8-row cap + +N more overflow.
• avg_mem_used_mb / 1024 → GiB — MiB-despite-_mb assumption stays explicitly commented as a re-check item for the feat/agentx-power-artifacts producer. Fine to ship as documented.

Runtime — /inference renders with 0 console errors after the merge; build is healthy, no regression from the base update. The two new axes and worker-drilldown paths were runtime-verified in the prior passes (worker table needs a pinned point with populated workers[], which prod AgentX rows don't yet ship — the synthetic Cypress fixtures cover it).

CI — Claude route in progress at review time; tests-* / lint / Bugbot not yet reported. Nothing failing. No new untested code was introduced by the merge.

Nothing blocking. Merge-order reminder still stands: this is a stacked PR — base feat/measured-power-tiering (#934) must land first; do not squash into master before it.
· feat/worker-power-drilldown

[github-actions]

Claude finished @edwingao28's task in 1m 8s —— View job

---

Re-review: PR #935 — role-local measured energy axes + per-worker power drilldown

• Identify new commits since last review (225163cb)
• Review any new/changed feature code
• Check CI status
• Runtime verification (Playwright)
• Post findings

🟢 LGTM — no blocking issues found

The only new commit since the last LGTM is the merge 88558843 ("chore: merge latest master"). It's a clean merge — git show 88558843 carries no conflict-resolution hunks, and this PR's own delta is unchanged: 13 files, +668/−1, byte-identical to what the three prior LGTM passes reviewed (git diff 225163cb..HEAD is empty). None of the authored feature files (metric-registry.ts, chart-utils.ts, types.ts, tooltipUtils.ts) were touched by the merge.

Nothing new to review substantively. Carrying forward the previously verified conclusions:

• Registry axes measuredPrefillJPerInputToken / measuredDecodeJPerOutputToken — bilingual labels/titles, polarity: 'lower', MEASURED_ENERGY_METRIC_CONFIG_KEYS membership; tokenMetricTypeForConfigKey classifies each correctly with no substring collision.
• Chart fields — role-local joules pass straight through chartMetric(...), no double-gating (already scrubbed on power_valid in rowToAggDataEntry).
• Tooltip XSS — w.role / w.hosts escapeHtml-escaped; generateWorkerPowerHTML gated on isPinned && workers?.length across all three generators; 8-row cap + +N more.
• avg_mem_used_mb / 1024 → GiB — MiB-despite-_mb assumption remains explicitly commented as a re-check for the feat/agentx-power-artifacts producer. Fine to ship as documented.

CI — Lint & Format ✅, Tests (Unit) ✅, Code Quality ✅; Tests (E2E) + Claude in progress at review time, nothing failing.

Runtime — /inference renders with 0 console errors after the master merge; build healthy, no regression from the base update. The worker table needs a pinned point with populated workers[] (prod AgentX rows still don't ship it) — the synthetic Cypress fixtures cover that path.

Merge-order reminder still stands: this is a stacked PR — base feat/measured-power-tiering (#934) must land first; do not squash into master before it.
· feat/worker-power-drilldown

[github-actions]

Claude finished @edwingao28's task in 1m 43s —— View job

---

Re-review: PR #935 — role-local measured energy axes + per-worker power drilldown

• Identify new commits since last review (225163cb)
• Review core feature files (registry, chart fields, tooltip)
• Check CI status
• Runtime verification
• Post review

🟢 LGTM — no blocking issues found

New commits since the last LGTM (225163cb): a master merge (88558843), a docs/comment relabel (c760e8bb), and the removal of the synthetic AgentX e2e spec (e5e04d39). The PR's authored feature files (metric-registry.ts, chart-utils.ts, types.ts, tooltipUtils.ts logic) are otherwise unchanged from the version already reviewed.

What the new commits changed

• c760e8bb — pure comment/text relabel: swaps internal PLAN-02 / gap G2 / PLAN-09 planning labels for human-readable prose in three files (certified-power-filter.cy.ts, tooltipUtils.ts MiB→GiB comment, benchmark-transform.test.ts). No behavior change.
• e5e04d39 — removes cypress/e2e/agentic-measured-power.cy.ts (−166) and its timings.json entries. Note this is a reversal of the e2e coverage the PR body still describes. It's a defensible call — the spec depended entirely on synthetic fixtures (prod AgentX rows don't ship workers[] yet), and the feature logic remains well-covered at the unit level (tooltip-utils.test.ts across all three generators incl. escaping/8-row cap/ZH; chart-utils.test.ts emission/absence/zero; benchmark-transform.test.ts pass-through + power_valid scrub; metric-registry.test.ts classification + share-link round-trip). Non-blocking, but worth updating the PR description so it no longer advertises the removed spec.

Re-verified

• Registry/chart fields — measuredPrefillJPerInputToken / measuredDecodeJPerOutputToken carry bilingual labels/titles, polarity: 'lower', and MEASURED_ENERGY_METRIC_CONFIG_KEYS membership. No output substring on the prefill key nor input on the decode key, so tokenMetricTypeForConfigKey classifies each correctly. Chart fields map straight through chartMetric(...); the comment correctly notes rowToAggDataEntry already scrubbed on power_valid (no double-gating).
• Tooltip XSS — w.role and w.hosts (artifact JSONB) are escapeHtml-escaped; remaining interpolated values are typed numbers. generateWorkerPowerHTML(d, isPinned, locale) gated on isPinned && workers?.length in all three generators, 8-row cap + +N more, MiB→GiB /1024.

Runtime — /inference renders with 0 console errors after the merge; build healthy.

CI — Lint & Format ✅; Tests (Unit) / Tests (E2E) in progress at review time, nothing failing.

Nothing blocking.
· feat/worker-power-drilldown

[github-actions]

Claude finished @edwingao28's task in 1m 34s —— View job

---

Re-review: PR #935 — role-local measured energy axes + per-worker power drilldown

• Identify new commits since last review (88558843)
• Review changed feature code
• Check CI status
• Runtime verification
• Post findings

🟢 LGTM — no blocking issues found

Two new commits since the last LGTM, both low-risk:

• c760e8bb (docs: replace internal PowerX planning labels) — comment-text only. Swaps internal PLAN-02/PLAN-09/G2 planning labels for reader-facing prose in certified-power-filter.cy.ts, tooltipUtils.ts, and benchmark-transform.test.ts. No functional change — the avg_mem_used_mb / 1024 → GiB line and all logic are byte-identical apart from the comment.
• e5e04d39 (test: remove synthetic AgentX power E2E) — deletes cypress/e2e/agentic-measured-power.cy.ts (166 lines) and its timings.json entry. Intentional: the spec relied on synthetic workers[] fixtures because prod AgentX rows don't yet ship worker telemetry (producer lands with feat/agentx-power-artifacts).

Note (non-blocking): removing that spec leaves the worker drilldown with no E2E coverage, but it retains comprehensive unit coverage — tooltip-utils.test.ts exercises generateWorkerPowerHTML across all three generators (official / overlay / GPU-compare), pinned/unpinned, empty/absent, the 8-row cap, ZH strings, and HTML escaping, which satisfies the AGENTS.md overlay-path test requirement. Worth re-adding a real-data E2E once feat/agentx-power-artifacts populates workers[] in fixtures.

Carried-forward verification (unchanged by these commits):

• Registry axes measuredPrefillJPerInputToken / measuredDecodeJPerOutputToken — bilingual labels/titles, polarity: 'lower', MEASURED_ENERGY_METRIC_CONFIG_KEYS membership; tokenMetricTypeForConfigKey classifies each correctly (no input/output substring collision).
• Chart fields (chart-utils.ts:575) — role-local joules pass straight through chartMetric(...), correctly not double-gated (already scrubbed on power_valid in rowToAggDataEntry).
• Tooltip — w.role / w.hosts (JSONB) escapeHtml-escaped; other interpolated values are typed numbers; gated on isPinned && workers?.length.

CI — Lint & Format ✅, Tests (Unit) ✅, Code Quality ✅, PR checks ✅; Tests (E2E) + Claude in progress at review time, nothing failing.

Runtime — /inference renders with 0 console errors after these commits; build healthy, no regression.

Merge-order reminder still stands: this is a stacked PR — base feat/measured-power-tiering (#934) must land first; do not squash into master before it.
· feat/worker-power-drilldown

[github-actions]

Claude finished @edwingao28's task in 39s —— View job

---

Re-review: PR #935 — role-local measured energy axes + per-worker power drilldown

• Identify new commits since last review (e5e04d39)
• Review changed code
• Check CI status
• Runtime verification
• Post findings

🟢 LGTM — no blocking issues found

The only new commit since the last LGTM is 0c5c96fb (docs: simplify PowerX implementation comments). It is comment/text-only — no functional change:

• certified-power-filter.cy.ts — trims the file header and several inline explanatory comments; all cy.* assertions are byte-identical.
• tooltip-utils.test.ts — removes a section banner and two inline comments; test bodies unchanged.
• tooltipUtils.ts — condenses the WORKER_ROWS_LIMIT and generateWorkerPowerHTML doc comments. The gating condition (!isPinned || !Array.isArray(d.workers) || d.workers.length === 0), escaping, 8-row cap, and MiB→GiB /1024 conversion are all unchanged.

git diff confirms no logic lines were touched. Nothing to review substantively.

Carried-forward verification (unchanged by this commit):

• Registry axes measuredPrefillJPerInputToken / measuredDecodeJPerOutputToken — bilingual labels/titles, polarity: 'lower', MEASURED_ENERGY_METRIC_CONFIG_KEYS membership; tokenMetricTypeForConfigKey classifies each correctly (no input/output substring collision).
• Chart fields — role-local joules pass straight through chartMetric(...), not double-gated (already scrubbed on power_valid in rowToAggDataEntry).
• Tooltip XSS — w.role / w.hosts (artifact JSONB) escapeHtml-escaped; other interpolated values are typed numbers.

CI — all six runs pending/in-progress at review time (freshly triggered by this push); nothing failing.

Merge-order reminder still stands: this is a stacked PR — base feat/measured-power-tiering (#934) must land first; do not squash into master before it.
· feat/worker-power-drilldown

[github-actions]

Claude finished @edwingao28's task in 47s —— View job

---

Re-review: PR #935 — role-local measured energy axes + per-worker power drilldown

• Identify new commits since last review (e5e04d39)
• Review changed code
• Check CI status
• Post findings

🟢 LGTM — no blocking issues found

Two new commits since the last LGTM, both comment/narration only — no behavior change:

• 0c5c96fb (docs: simplify PowerX implementation comments) — trims comments in certified-power-filter.cy.ts, tooltip-utils.test.ts, and tooltipUtils.ts. In tooltipUtils.ts it drops the WORKER_ROWS_LIMIT inline note and condenses the generateWorkerPowerHTML docblock; the guard (!isPinned || !Array.isArray(d.workers) || d.workers.length === 0), 8-row cap, +N more, escaping, and MiB→GiB conversion are byte-identical.
• 3c80df4e (docs: remove redundant PowerX test narration) — trims comments in quickFilters.test.ts, tooltip-utils.test.ts, and benchmark-transform.ts. The benchmark-transform.ts change only reworded the comment above the measuredPowerValid ? … : undefined hoists — the gating logic is unchanged.

Confirmed via git show that neither commit touches any executable line. The core feature files (metric-registry.ts, chart-utils.ts, types.ts, and the generateWorkerPowerHTML logic itself) are unchanged from the versions approved in the prior passes.

Carried-forward verification (unchanged):

• Registry axes measuredPrefillJPerInputToken / measuredDecodeJPerOutputToken — bilingual labels/titles, polarity: 'lower', MEASURED_ENERGY_METRIC_CONFIG_KEYS membership; tokenMetricTypeForConfigKey classifies each correctly (no input/output substring collision).
• Chart fields — role-local joules pass straight through chartMetric(...), not double-gated (already scrubbed on power_valid in rowToAggDataEntry).
• Tooltip XSS — w.role / w.hosts (JSONB) escapeHtml-escaped; remaining interpolated values are typed numbers; gated on isPinned && workers?.length across all three generators.

CI — Lint & Format ✅, Tests (Unit) ✅, Code Quality ✅, PR checks ✅; Tests (E2E) + Claude in progress at review time, nothing failing.

Nothing blocking. Merge-order reminder still stands: this is a stacked PR — base feat/measured-power-tiering (#934) must land first; do not squash into master before it.

[github-actions]

Claude finished @edwingao28's task in 2m 11s —— View job

---

Re-review: PR #935 — role-local measured energy axes + per-worker power drilldown

• Identify new commits since last review (3c80df4e)
• Review changed feature code
• Check CI status
• Runtime verification (Playwright)
• Post findings

🟢 LGTM — no blocking issues found

The only new commit since the last LGTM is 62095602 (fix: explain unavailable role-local energy). It's a focused, well-tested addition:

• New helper isRoleLocalMeasuredEnergyConfigKey (metric-registry.ts) — matches exactly the two role-local keys (y_measuredPrefillJPerInputToken / y_measuredDecodeJPerOutputToken), disjoint from the whole-deployment measured-energy keys. Covered by a new registry test.
• Empty-state copy — both ScatterGraph.tsx and GPUGraph.tsx swap noDataHint for a role-energy-specific bilingual message when a role-local axis is selected and no series carry role energy. This is the expected path today (prod AgentX rows don't yet ship role-local energy). It's a heuristic keyed on the selected axis rather than the reason data is empty, so an unrelated empty selection on that axis would also show this hint — acceptable UX tradeoff, not a bug.
• Tests — new component specs for both graphs (EN + ZH) assert the role-energy explanation renders and the generic hint does not; the metric-registry unit test locks the helper's classification.
• Chinese copy — 当前数据集未提供 Prefill/Decode 各角色的能耗数据。请选择其他模型、场景、精度、日期或实测能耗指标。 reads naturally, preserves the source meaning, and correctly keeps Prefill/Decode in English per convention. Nothing to flag.

Runtime (Playwright, real read-only DB) — loaded /inference?i_metric=y_measuredPrefillJPerInputToken; the new empty state renders ("This dataset does not report role-level prefill/decode energy…") with 0 console errors, and the share-link param round-trips to the axis. Carried-forward verification of the two axes, chart-field mapping, and generateWorkerPowerHTML escaping/8-row-cap/gating from the prior passes is unchanged.

CI — Lint & Format ✅, Tests (Unit) ✅, PR checks ✅; Tests (E2E) / Code Quality / Claude in progress at review time, nothing failing.

Nothing blocking. Merge-order reminder still stands: this is a stacked PR — base feat/measured-power-tiering (#934) must land first; do not squash into master before it.
· feat/worker-power-drilldown