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
Touches shared chart zoom and renderer paths used by all D3 charts; behavior is guarded by new tests but incorrect identity/replay logic could affect zoom state or tooltips.

Overview
Skips redundant identity zoom replay on chart rebuild so charts are not fully re-rendered twice when the user has not zoomed.

useChartZoom.setupZoom now calls zoom.transform only when the stored transform or the SVG node’s internal __zoom state is non-identity; user zoom and defaultZoomK replay behavior is unchanged. useD3ChartRenderer explicitly dismisses pinned tooltips on identity rebuilds, matching the old side effect of the identity replay’s zoom emit.

Adds useChartZoom.setup.test.ts (jsdom + real d3-zoom) covering no emit at identity, zoom preservation, ref sync, defensive node/ref desync, and non-identity defaultZoomK.

^{Reviewed by Cursor Bugbot for commit 1c344a9. 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 12, 2026 4:25am