feat(overview): add 30-day cost comparison / 总览页新增 30 天成本对比

[edwingao28]

Summary

• Add a shareable 30-day change mode to /overview and /zh/overview.
• Compare each platform with its latest validated result from 30–60 days earlier, matched by model, scenario, hardware, SLO, and engine scope.
• Preserve tier, engine, locale, and comparison state in URLs; render finite changes as heat-map percentages, missing baselines as ∞, and missing/stale current results as —.

Unofficial overlays: N/A — Overview is a server-rendered summary and has no ?unofficialrun= path.

Validation

• typecheck, lint, fmt, production build
• Unit: 3,628 passed
• Cypress component: 195 passed
• Cypress integration (Chrome): 585 passed
• Overview E2E (Firefox): 20 passed

中文说明

• 为 /overview 与 /zh/overview 新增可分享的「30 天变化」模式。
• 按模型、场景、硬件、SLO 与引擎范围匹配各平台，并与 30–60 天前最近一次有效结果对比。
• 在 URL 中保留档位、引擎、语言与对比状态；可比结果显示热力百分比，缺少历史基线显示 ∞，当前无结果或没有更新结果显示 —。

非官方运行叠加不适用：Overview 是服务端渲染的摘要页，不存在 ?unofficialrun= 数据路径。

验证

• typecheck、lint、fmt、production build
• 单元测试：3,628 项通过
• Cypress 组件测试：195 项通过
• Cypress 集成测试（Chrome）：585 项通过
• Overview E2E（Firefox）：20 项通过

---

Note

Medium Risk
Introduces a second benchmark fetch and historical matching rules that drive published cost deltas; incorrect window or pairing logic would misstate trends, though changes are display-only with broad test coverage.

Overview
Adds a 30-day change comparison mode on the inference cost overview (?compare=30d), alongside the existing vs B200 view.

Users switch via a new comparison control; tier, engine scope, locale, and comparison mode stay in sync through canonical URLs. In history mode the matrix drops the B200 “Reference” column treatment and shows per-platform cost deltas versus each platform’s latest validated result from 30–60 days earlier (same model, scenario, hardware, SLO, and engine scope). Comparable cells keep heat-map percentages; missing baselines or stale current evidence show cost only (no misleading ∞ in history).

Backend: resolves comparison mode from compare, derives a snapshot date from current overview-visible rows, loads an as-of benchmark snapshot for the baseline window, and merges via assembleOverviewHistoricalPageData (with cached getCachedBenchmarksAsOf). Fixtures and unit/E2E coverage exercise the new path; E2E date assertions ignore .sr-only text so screen-reader baseline dates don’t fail “no visible dates” checks.

^{Reviewed by Cursor Bugbot for commit 9dabdea. 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 (UTC)
inferencemax-app	Ready	Preview	Aug 5, 2026 8:53pm

[github-actions]

Claude finished @edwingao28's task in 7m 57s —— View job

---

PR Review: 30-day cost comparison on /overview

• Read changed files & analyze diff
• Check CI status
• Runtime verification (localhost + Playwright)
• Post review findings

🟢 LGTM — no blocking code issues found.

I verified the change at runtime against the read-only DB (/overview, /overview?compare=30d, and /zh/overview?tier=75&engine=all&compare=30d). The feature works as intended.

Runtime verification ✅

• 30-day mode renders finite deltas as a heat-map (−9%, −45%, +14%, …), ∞ for a missing 30–60-day baseline, and — for stale/no-newer current results — matching the spec and the legend copy.
• The Compare switcher toggles between vs B200 and 30-day change; the default (vs B200) mode is unregressed.
• URL state is preserved across switchers: the "vs B200" link keeps ?tier=75&engine=all, the engine switcher keeps ?tier=75&compare=30d, aria-current tracks the active mode. zh strings render (对比 B200, 30 天变化, legend).
• 0 console errors on all three routes.

Correctness spot-checks ✅

• getCachedBenchmarksAsOf blob cache key is benchmarks-as-of:<keys>:<date> — the as-of date is part of the key, so different target dates don't collide with each other or with the absolute-latest benchmarks slot (benchmark-data.server.ts:25).
• The costPerMtok === null early-return in displayedComparison is unreachable from CellValue (guarded by the member.costPerMtok === null check at overview-scorecard.tsx:392), so the vs-B200 badge behavior is unchanged — no regression.
• assembleOverviewHistoricalPageData matches by model|scenario|hardware and applies engineScope to both current and baseline assemblies; the no_newer_result / no_baseline branches are exhaustive.

Tests ✅ — 86 overview unit tests pass locally; new E2E (overview.cy.ts) and fixtures added for both modes.

ℹ️ Note on the red Tests (Unit) check (not caused by this PR): the only failure is src/etl/trace-replay-ingest.test.ts > streams a file-backed trace… timing out at 5000ms — an ETL trace-replay test untouched by this PR (a known flaky timeout). The PR's own unit suites (overview-data, overview-data.server, overview-links) all pass. Recommend a re-run of the unit job to get it green before merge.
· feat/overview-30-day-comparison

[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.}

Want higher recall? High effort reviews run extra passes and find more bugs. A team admin can switch effort levels in the Cursor dashboard.

^{Reviewed by Cursor Bugbot for commit b6ff87f. Configure here.}

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: b6ff87fd42

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[github-actions]

Claude finished @edwingao28's task in 2m 4s —— View job

---

Re-review: 30-day cost comparison on /overview

• Review new commit 61eda29 (post-dates prior review at b6ff87f)
• Verify Cursor Bugbot a11y finding (sr-only on cost delta badge)
• Verify Codex P2: history window from scoped rows
• Verify Codex P2: gate historical changes on tier evidence date
• Runtime verification (localhost + Playwright)
• Post review findings

🟢 Re-review: LGTM — commit 61eda29 resolves all three outstanding findings; no new blocking issues.

Since the prior review (b6ff87f), commit 61eda29 "align historical evidence and accessibility" cleanly addresses every open comment. I verified each fix statically and at runtime against the read-only DB.

Cursor Bugbot — cost-delta badge screen-reader text ✅ Fixed
CostDeltaBadge now renders a <span className="sr-only">{aria}</span> sibling and drops the unreliable aria-label on the role-less span, matching the working CellMissing pattern (overview-scorecard.tsx:366-369). Confirmed in the live DOM: visible text -38% is aria-hidden, the sr-only span reads "38% cheaper than this platform's Jul 1 result", and no aria-label remains on the wrapper.

Codex P2 — history window built from unscoped rows ✅ Fixed
overviewSnapshotDate now takes engineScope and derives the snapshot date from overviewEngineRows(rows, engineScope) instead of the raw fetch (overview-data.ts:166-184); the server passes the active scope through (overview-data.server.ts:48). A newer out-of-scope ATOM/TRTLLM row can no longer push the 30-day target forward for the community matrix.

Codex P2 — delta gated on frontier latestDate instead of tier evidence ✅ Fixed
New overviewTierEvidenceDate(read) returns read.evidenceDate?.to ?? read.config?.latestDate and is now used for the no_newer_result cutoff and both baseline labels (overview-data.ts:345-346, 689-720). Since these branches only run when costPerMtok !== null (an in-range interpolated read), evidenceDate.to is always populated — the SLO-specific tier date now drives the — decision, so a stale tier read is no longer mistaken for a newer result.

Runtime verification ✅ (/overview?compare=30d)

• Heat-map renders finite deltas (−9%, −45%, −13%, +14%, −12%, −21%, +3%), ∞ for missing 30–60-day baselines, — for no-newer/stale current results — matching the legend copy.
• 0 console errors.

CI ✅ — all checks green on the latest commit (Typecheck & unit, Lint/format, Component, E2E Chrome ×4 + Firefox, CodeQL, Cursor Bugbot). The previously-flagged flaky trace-replay-ingest timeout was relaxed in this commit and the unit job is now green. New unit coverage (overview-data.test.ts +27) and E2E (overview.cy.ts) exercise the scoped-window and tier-evidence paths.

Overlay path is correctly N/A — Overview is server-rendered with no ?unofficialrun= branch.
· feat/overview-30-day-comparison

[github-actions]

Claude finished @edwingao28's task in 2m 39s —— View job

---

Re-review: 30-day cost comparison on /overview

• Check for new commits since prior review (61eda29)
• Review new commit 9dabdea "promote comparison tabs"
• Check CI status
• Runtime verification (localhost + Playwright)
• Post findings

🟢 Re-review: LGTM — commit 9dabdea is a clean presentation + simplification change; no blocking issues.

Only one new commit since the prior LGTM (61eda29). It does two things, both verified statically and at runtime:

1. Compare switcher promoted to a centered tab row
The OverviewComparisonSwitcher moved out of the card header into a primary <nav> tab row (overview-page.tsx:99-105), restyled as underline tabs. URL state is preserved: on /overview?tier=75&engine=all&compare=30d the vs B200 link resolves to /overview?tier=75&engine=all and aria-current tracks the active mode. Confirmed the nav is a direct child of [data-testid="overview-page"] with justify-center, matching the updated E2E assertions.

2. History mode: current-cost-only when no valid 30-day comparison
displayedComparison now returns a badge only for status === 'comparable' && costDeltaPct !== null (overview-scorecard.tsx:246-252); no_baseline / no_newer_result fall through to null, so those cells render just the cost with no badge and an untinted background. The DisplayedComparison['status'] type is correctly narrowed via Exclude<…, 'no_newer_result'>, and the now-dead ∞/— history branches, aria strings, and legend copy were removed consistently (EN + zh).

Runtime verification ✅

• /overview?compare=30d: comparable cells render heat-map deltas (-38%, -67%, -9%, -45%, +14%, +3%, +12%, -21%), platforms without a baseline (e.g. DeepSeek V4 8K/1K B300/GB200/GB300) show cost only — no ∞, no — badge. Legend reads "Platforms without a valid 30-day comparison show current cost only."
• /overview (default vs-B200): unregressed — B200 stays the reference column (0 deltas), other cells show ±%, and ∞ still correctly appears for genuinely-missing B200 baselines (the doc comment was rescoped to "Missing B200 evidence" to match).
• Genuinely-empty cells still render — via the separate CellMissing path (no data for this scenario / no exact @50 result) — unaffected.
• 0 console errors on all routes.

CI — Tests (Unit) ✅ and Lint & Format ✅ green on the latest commit; Tests (E2E) in progress. The E2E spec (overview.cy.ts) was updated in this same commit to assert the new no-badge/transparent-cell behavior and the centered switcher, so the new code path is covered.

Overlay path remains correctly N/A — Overview is server-rendered with no ?unofficialrun= branch.
· feat/overview-30-day-comparison