fix(inference): rename J/total-token title to avoid input-metric heuristic

[arygupt]

Summary

Hot-fix on top of #393: rename the Measured J per Token chart title so it doesn't accidentally trigger the input-metric X-axis swap.

What was wrong

The inference scatter chart has a heuristic in ChartControls.tsx:

const isInputMetric = title.toLowerCase().includes('input');
if (isInputMetric) {
  // swap X-axis from interactivity → P99 TTFT
}

This is correct behavior for true input-only metrics (input throughput, input cost) — but my title for the new J/total-token metric was "Measured Joules per Token (input + output)", which contained the word "input" and tripped the heuristic incorrectly.

Visible symptom: selecting "Measured J per Token" auto-changed the X-axis to P99 TTFT instead of staying on interactivity.

Fix

Renamed the title from "Measured Joules per Token (input + output)" → "Measured Joules per Token (incl. prompt)". Same meaning, no "input" trigger word.

Why not patch the heuristic instead

The substring check is intentional — it's how the codebase routes metrics into the right X-axis. Changing it would require either:

• adding a more specific flag to chart config (_isInputMetric: true)
• or rewriting the routing logic

Both are bigger changes than this hot-fix needs. Renaming one title string keeps the diff minimal and preserves the existing convention.

Test plan

• pnpm typecheck clean
• pnpm test:unit 1944/1944 passing (no test changes needed)
• After deploy: select "Measured J per Token" → X-axis stays on interactivity (not P99 TTFT)

---

Note

Low Risk
Copy-only chart config label changes with no logic, API, or data-path impact.

Overview
Renames the Measured J per Token chart title in inference-chart-config.json from "(input + output)" to "(incl. prompt)" for both interactivity and e2e chart definitions.

The wording still describes total-token energy (prompt + generation), but drops the substring "input" so it no longer matches the existing isInputMetric title heuristic that swaps the interactivity chart X-axis to TTFT. Measured J per Token should keep the default interactivity X-axis instead of jumping to P99 TTFT.

^{Reviewed by Cursor Bugbot for commit ec9ceca. 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, Comment	May 27, 2026 1:45am

[claude]

Claude finished @arygupt's task in 53s —— View job

---

Claude PR Review

🟢 LGTM — no blocking issues found.

• Read PR context
• Inspect the diff in inference-chart-config.json
• Verify the heuristic in ChartControls.tsx
• Check for any other places that reference the old title

Verification

• The heuristic at packages/app/src/components/inference/ui/ChartControls.tsx:214-220 does title.toLowerCase().includes('input'). The new title "Measured Joules per Token (incl. prompt)" no longer contains the substring input, so the X-axis swap will not be triggered. ✅
• grep for "input + output" across packages/ returns no remaining references — nothing else (tests, snapshots, docs) hard-codes the old string. ✅
• The metric key (measuredJPerTotalToken), short label (Measured J per Token (J/tok)), and roofline configuration are unchanged, so display ordering, legends, and roofline math are unaffected. ✅
• No tests needed: pure config string rename, not new functionality. The PR description confirms pnpm typecheck and pnpm test:unit are clean.

Note (non-blocking, out of scope)

The includes('input') substring heuristic is fragile by design — any future metric title that happens to contain the word "input" (e.g. "input-sensitive accuracy") would also trip the X-axis swap. The PR author correctly chose to avoid touching it in this hot-fix; if it bites again, the right follow-up is an explicit _isInputMetric flag in inference-chart-config.json (or on ChartDefinition), so axis routing isn't dependent on display text.
· Preview