perf(charts): skip identity zoom-transform replay on chart rebuild

[Oseltamivir]

Context

Second of the INP fixes (field CrUX INP 273 ms p75, failing CWV). The Firefox profile of the live inference tab shows every chart rebuild renders everything twice: once directly (axes, grid, all layers at base scales) and once more via setupZoom's unconditional svg.call(zoom.transform, stored) — which synchronously dispatches a zoom event whose handler re-renders axes + grid + every layer again. The profiled hot path:

layout effect → setupZoom → zoom.transform → emit('zoom') → onZoom
  → renderAxes + renderGrid + updateLayerOnZoom(all layers) + user onZoom

When the user has never zoomed (identity transform — the overwhelmingly common case) the replay repositions nothing: it just repeats the full render for pixel-identical output, doubling every ~250–490 ms rebuild long task.

Changes

• useChartZoom.setupZoom — replay the stored transform only when it (or the node's internal __zoom state, as a defensive desync check) is non-identity. At identity, attaching the behavior already leaves d3's internal state consistent with the freshly drawn base-scale DOM.
• useD3ChartRenderer — the identity replay had one observable side effect: its emit dismissed a pinned tooltip on every rebuild. That dismissal is now performed explicitly when the replay is skipped, so observable behavior is unchanged.

Zoom preservation is untouched (docs/d3-charts.md "Zoom Transform Preservation" / pitfalls "D3 Zoom Transform Loss"): non-identity transforms — user zooms and defaultZoomK charts — replay exactly as before, verified by tests.

Impact

Halves the main-thread cost of every chart rebuild while not zoomed: initial mounts, data loads, comparison-date loads, legend/precision toggles, metric switches — across all D3Chart consumers (inference scatter, GPU timeline, evaluation/reliability bars, trends, calculator). Stacks with #444.

Tests

New useChartZoom.setup.test.ts (jsdom, real d3-zoom on a real SVG):

• identity stored transform → no zoom emit on setup (regression test: fails against old implementation)
• non-identity transform → replayed exactly once with the right k/x/y on re-setup (zoom preservation)
• zoomTransformRef stays in sync after replay
• node/ref desync → defensive replay normalizes the node
• defaultZoomK ≠ 1 → initial transform still applied

Full unit suite passes (2037 tests). pnpm typecheck / lint / fmt clean.

Applies uniformly to official and ?unofficialrun= overlay rendering — both go through the same rebuild path; overlay rooflines/points reposition via the same non-identity replay when zoomed.

---

Note

Medium Risk
Changes the shared D3 chart zoom setup path used on every rebuild; risk is mitigated by focused tests and preserved behavior for non-identity zoom and tooltip dismissal.

Overview
Stops redundant full chart re-renders when zoom is at identity (the common case) by not calling zoom.transform during setupZoom unless the stored transform or the SVG node’s internal __zoom state is non-identity.

Non-identity zoom (user pan/zoom, defaultZoomK) still replays exactly as before. useD3ChartRenderer explicitly dismisses a pinned tooltip on rebuild when that identity replay is skipped, matching the old side effect of the zoom handler firing.

Adds useChartZoom.setup.test.ts covering identity skip, zoom preservation, ref sync, defensive node/ref desync, and defaultZoomK initial apply.

^{Reviewed by Cursor Bugbot for commit 0561024. 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	Jun 16, 2026 12:43am