feat(profit): add compare history to /profit-estimator and /profit-estimator-per-gigawatt / 利润估算器：新增对比历史趋势面板

[functionstackx]

Summary

Adds the /inference Compare history panel to /profit-estimator and /profit-estimator-per-gigawatt, so a chip's revenue and profit at the chosen target can be read against the same chip on an earlier run date, or on one specific run of a day that ran twice.

How /inference does it (what this mirrors)

• ChartControls.tsx: ControlPanel legend="Compare history" with a MultiSelect (max 4 chips) and, once a chip is picked, a DateRangePicker whose availableDates are the dates those chips have runs on.
• ChartDisplay.tsx → ComparisonChangelog.tsx: rendered under the controls as soon as a chip is selected. Fed by useComparisonChangelogs (one workflow-info?date= query per available date, bounded to the range once set). Entries are filtered by model DB key, precision, and configKeyMatchesHwKey; range endpoints and the first available date always show. Single-run dates get Git Commit / Workflow Run links and either "Add to chart", "Remove from chart", or a locked "On chart" (range endpoints). Multi-run dates render one block per run (date~r<runId>, labelled #1, #2, …) with their own notes and add/remove. "Add all to chart" pins every entry not yet on the chart.
• InferenceContext.tsx: pinned entries live in selectedDates / i_dates (comma list), cleared with the chips or when the range leaves the available dates. ChartDisplay expands a pinned plain date that has several runs into one entry per run, and numbers runs from dataRunsForDate so legend and changelog agree.
• useChartData.ts: buildComparisonDates = resolveComparisonEntries(selectedDates, range) minus the main run date; a plain date is fetched with benchmarkQueryOptions(model, date, enabled, exact=true), a run entry with (model, '', enabled, false, runId, exactRun=true).
• GPUGraph.tsx: one series per (chip, entry), lighter shade for older entries, labels ${chip} • ${comparisonEntryLabel(entry)}.

What this PR does

• "Compare history" ControlPanel under the custom-cost block: Chip Config MultiSelect (maxSelections=4, agentic configs for the current model at the auto-resolved precision) and a Comparison Date Range DateRangePicker shown once a chip is selected. Same i_gpus / i_dstart / i_dend / i_dates params as /inference, so a comparison URL works on both surfaces.
• Config Changelog: the /inference ComparisonChangelog component, unchanged apart from a new optional analyticsSection prop, rendered under the panel whenever a chip is selected and fed by useComparisonChangelogs(selectedGPUs, range, availableDates, 'agentic_traces'). Same filters, links, lock, add/remove, "Add all", and per-run blocks. Run numbering comes from dataRunsForDate on the changelog's runConfigs; a pinned plain date with several runs expands to per-run entries, as on /inference.
• useProfitHistory resolves entries with /inference's buildComparisonDates and fetches plain dates as exact-date queries and run entries as runId + exactRun queries (calculator view). Loading is folded into the page flag so bars never render half-populated.
• buildProfitHistoryResults dedupes runs per chip per entry, builds groups only for the selected chips, and interpolates on the same monotone Hermite spline the current bars use. Result keys are hwKey[__prec]|entry (| because entries carry their own ~r suffix); rows carry date (the entry) and dateLabel (2026-08-10 #2).
• Chart: today's bars for the compared chips only, plus one bar per entry next to each, ordered chip → entries by comparisonEntrySortValue → today. shadeHistoryColor lightens the chip colour in OKLCH, oldest lightest. X labels, tooltip "Run date", caption (profit-history-note), and the CSV "Run date" column all print the changelog's label for the entry.
• Pricing inputs are unchanged across entries: same target interactivity, utilization, license fee, price source, and TCO tier (or custom $/GPU/hr).
• Chip options are every agentic config for the model at the effective precisions, as on /inference. With the default fixtures that includes H200 fp8: the sparse-precision rule keeps fp8 in play when it has a single curve, so H200 is offered even though its curve stops short of the target on the chart.
• Pinning the run behind a main bar adds no duplicate. The main query is an as-of-date fetch, so each chip's current bar can come from a different run; profitHistoryCurrentRunIds maps each compared chip to the latest run of the current date that covered it (from the changelog's enumeration), dropCurrentRunEntries skips a pinned run only when every compared chip already shows it, and buildProfitHistoryResults skips it per chip. This is the per-chip form of the selectedRunId /inference hands buildComparisonDates.
• A compared chip that only priced on an earlier entry stays in the legend (profitHistoryLegendKeys), and the caption names it as missing on the current date (profitHistoryMissing) rather than dropping its bars.
• When a model switch prunes the whole chip selection, the range and pinned entries are cleared with it, as clearing the chips by hand does.

Analytics

profit_history_gpu_selected ({ gpus, count }), profit_history_date_range_changed ({ startDate, endDate }), and via analyticsSection="profit": profit_comparison_changelog_toggled, profit_changelog_add_all_dates, profit_changelog_add_date / remove_date, profit_changelog_add_run / remove_run. /inference keeps its inference_* names (the prop defaults to inference).

Unofficial-run overlays

N/A. The estimator passes no overlay to useThroughputData, and the history bars come from official runs by date or run id, so there is no overlay branch to serve.

/zh

All new estimator strings are in the component STRINGS table (对比历史趋势, 芯片配置, 选择芯片配置进行对比, 对比日期范围, 选择日期范围, 运行日期); the changelog reuses its own zh table (配置变更日志, 添加到图表, 全部添加到图表, 已在图表中, …). Covered by a Cypress case on /zh/profit-estimator-per-gigawatt.

Tests

• profit-history.test.ts: option filtering and registry ordering; result key parse/format incl. run entries; profitHistoryEntryLabel / profitHistoryEntryDate; row ordering and ranks with same-day runs; fade share and OKLCH shading; buildProfitHistoryResults interpolation, selection filter, and same-day run dedupe; profitHistoryCurrentRunIds / dropCurrentRunEntries (per-chip current run, sparse later run), per-chip skip in buildProfitHistoryResults, profitHistoryLegendKeys (history-only chip kept while comparing), profitHistoryMissing (current date included).
• ProfitEstimatorChart.test.ts: dated rowLabel, splitHistoryLabel, tooltip run date, xLabelLayout with a date line.
• profit-estimator.cy.ts: fixtures gain 2026-07-15 (80% throughput, one run), 2026-08-10 (two runs at 85% and 92%, each with a changelog row), and 2026-08-31 (today, one run); interceptProfitData serves ?date=, ?runId=&exactRun=true, and /api/v1/workflow-info?date=. Cases: panel and option list; Max Range on two chips gives four TCO bars in the right order; URL hydration and clearing; Config Changelog lists three dates with notes and links, "Add to chart" pins i_dates and adds a bar, "Remove from chart" clears it; two-run date shows #1 / #2 blocks, adding run chore: bump the all-minor-patch group with 2 updates #1 adds a ~r<runId> entry and a 2026-08-10 #1 bar, "Add all to chart" adds the rest with revenues rising oldest → newest; hydration from ?i_dates=…~r… with range endpoints locked as "On chart"; /zh labels. Assertions are on the UI (trigger text, Add/Remove buttons, range picker, bar counts), not the address bar: url-state.ts keeps share params in memory and strips them from the address bar after load. Panel interactions wait for the first priced bar, since the options depend on the availability rows and the auto-resolved precision.
• profit-estimator.cy.ts (pre-existing, from feat(app): add MiniMax M3 to the profit estimators with MiniMax list-price defaults / 利润估算器新增 MiniMax M3 并默认采用 MiniMax 官方定价 #1002): the MiniMax → Kimi switch asserted a bare /profit-estimator-per-gigawatt, but modelRoutePathnameRewrite yields /profit-estimator-per-gigawatt/kimi-k3 when leaving a per-model path (the GLM → Kimi case in the same file asserts exactly that). That shard was red on feat(app): add MiniMax M3 to the profit estimators with MiniMax list-price defaults / 利润估算器新增 MiniMax M3 并默认采用 MiniMax 官方定价 #1002's own CI; the expectation now matches the route behaviour.
• Local: typecheck, lint, fmt, check:typography clean; calculator + inference/ui vitest suites pass; profit-estimator.cy.ts passes locally against the fixture build (Cypress 15.21.1, CI=true E2E_FIXTURES=1).

中文说明

为 /profit-estimator 和 /profit-estimator-per-gigawatt 增加与 /inference 相同的"对比历史趋势"面板，包括配置变更日志与固定日期/运行。

• 面板位于自定义成本区块之下：芯片配置多选（最多 4 个，当前模型在自动精度下的 agentic 配置），选中后显示对比日期范围选择器。使用共享的 i_gpus / i_dstart / i_dend / i_dates URL 参数，/inference 上的对比链接可直接在估算器打开。
• 配置变更日志：复用 /inference 的 ComparisonChangelog 组件（仅新增可选的 analyticsSection 属性），选中芯片后显示于面板下方，数据来自 useComparisonChangelogs。保留相同的过滤规则、Git Commit / Workflow Run 链接、范围端点的"已在图表中"锁定、添加/移除、"全部添加到图表"以及多运行日期的逐运行区块（chore: bump the all-minor-patch group with 2 updates #1、chore: bump anthropics/claude-code-action from 04f907ba2a4c77706a56b4078c682d250716cc0c to cd77b50d2b0808657f8e6774085c8bf54484351c #2…）。运行编号取自 dataRunsForDate；固定的单日日期若有多次运行会展开为逐运行条目。
• useProfitHistory 通过 buildComparisonDates 解析对比条目：普通日期按 exact=true 查询，运行条目按 runId + exactRun 查询（calculator 视图）。
• 柱形、提示框、说明文字与 CSV 均使用与变更日志一致的"日期 #n"标签；排序与颜色深浅按 comparisonEntrySortValue。
• 埋点：profit_history_gpu_selected、profit_history_date_range_changed，以及 profit_changelog_* / profit_comparison_changelog_toggled。
• 非官方运行叠加不适用：估算器未向 useThroughputData 传入叠加数据。
• 芯片选项为当前模型在有效精度下的全部 agentic 配置（与 /inference 一致）；固定主柱形所对应的运行不会产生重复柱形（按芯片记录当前运行 id：profitHistoryCurrentRunIds / dropCurrentRunEntries）；仅在历史日期有数据的芯片仍保留在图例中，并在说明文字中标注当日缺失；模型切换清空芯片选择时同时清空日期范围与固定条目。
• 测试：vitest 覆盖运行条目的键、标签、排序与分级，以及当前运行 id、图例键与缺失说明；Cypress 夹具新增双运行日期（含变更日志行）与 workflow-info 拦截，新增变更日志面板、固定日期、逐运行柱形与"全部添加"、URL 还原与锁定、/zh 镜像用例，断言基于界面而非地址栏。修正 feat(app): add MiniMax M3 to the profit estimators with MiniMax list-price defaults / 利润估算器新增 MiniMax M3 并默认采用 MiniMax 官方定价 #1002 中 MiniMax → Kimi 切换用例的路径期望（/profit-estimator-per-gigawatt/kimi-k3）。本地 typecheck / lint / fmt / typography 与 Cypress 用例通过。

---

Note

Medium Risk
Large UI and data-layer change to profit estimation and shared comparison URL state, with many edge cases around run deduplication and chart merging, though behavior mirrors an existing /inference pattern and is heavily tested.

Overview
Adds a Compare history flow to the profit estimator (including per-gigawatt), aligned with /inference: up to four agentic chip configs, a date range, and the shared Config Changelog, all driven by i_gpus, i_dstart, i_dend, and i_dates.

When a comparison is active, the chart shows only the selected chips—today’s bar plus one bar per range endpoint or changelog-pinned date/run—repriced with the same target, token prices, and TCO tier. Historical bars use lighter OKLCH shading, dated axis labels and tooltips, caption notes for missing runs, and a Run date CSV column. New profit-history / useProfitHistory modules fetch benchmarks (reusing buildComparisonDates) and interpolate history through the same path as current bars, including per-chip dedupe of the “current” run so pins do not duplicate today’s bar.

ComparisonChangelog gains an optional analyticsSection for profit-specific tracking. Cypress fixtures and e2e cover the panel, changelog pinning, multi-run days, URL hydration, and /zh labels; one model-switch test now expects /profit-estimator-per-gigawatt/kimi-k3 after the per-model route rewrite.

^{Reviewed by Cursor Bugbot for commit a0ff206. 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	Sep 5, 2026 3:48am UTC

[cursor]

Stale Bugbot comment from a previous run.

[cursor]

Stale Bugbot comment from a previous run.

[cursor]

Stale Bugbot comment from a previous run.

[cursor]

Cursor Bugbot has reviewed your changes using default effort and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 37dc6d4. Configure here.}