Scatter view collapses to a horizontal line after repeated box-zoom + zoom-out

[Darksinian]

Summary

On an interactive scatter chart in a Jupyter notebook, doing a box zoom ~3 times in a row and then zooming out corrupts the view transform: the axis scale goes completely wrong and the point cloud collapses into a horizontal line. The chart does not recover a sane domain on zoom-out.

Repro

Rendered in a notebook (playground.ipynb, editable install of main @ 541a7da):

import numpy as np
import xy
from IPython.display import display

rng = np.random.default_rng(42)
m2 = 20_000
sx = rng.normal(0, 1, m2)
sy = sx * 0.5 + rng.normal(0, 0.6, m2)
cvals = sx**2 + sy**2
svals = np.abs(rng.normal(size=m2))

display(
    xy.scatter_chart(
        xy.scatter(x=sx, y=sy, color=cvals, size=svals, colormap="viridis", size_range=(3, 18), opacity=0.7),
        title="xy — 20k scatter, color + size encoded",
        width=900,
        height=340,
    )
)

Steps:

1. Render the chart above.
2. Box-zoom into the data three times in succession (progressively deeper).
3. Zoom out.

Actual: the scale is completely wrong after the zoom-out; the scatter shrinks/flattens to a horizontal line.

Expected: zooming out walks back to a sane domain; the aspect/scale of the data is preserved at every step.

Notes

• Suspect the accumulated view-transform math on repeated box zooms — possibly the f32 offset re-centering path (dossier §16) or domain bookkeeping in js/src/53_interaction.js (box zoom vs. center-anchored zoom-out).
• The y-domain collapsing to ~zero height ("horizontal line") suggests a degenerate domain is being computed on the zoom-out step rather than inverting the prior box zooms.

Environment: Linux, Chromium-based browser, Jupyter notebook, xy editable install of main @ 541a7da (native kernel backend).

[FarhanAliRaza]

Is this the issue ?

[FarhanAliRaza]

Update: the remaining collapse was reproduced and fixed in #101.

The root cause is anisotropic box zoom: X and Y can be narrowed by very different factors, but Zoom Out previously expanded both geometrically. The less-zoomed axis could therefore grow past its original domain while the other axis remained zoomed, visually flattening the scatter.

The fix caps each axis independently at its home span. In the regression sequence, three box zooms produce X/Y zoom levels of 125× / 1.953×; after four zoom-outs the fixed result is 7.8125× / 1×. Before the fix, Y reached 0.122×—eight times wider than home.

The earlier GPU-precision hypothesis was not required for this failure, so that workaround has been removed to keep the PR focused. The PR includes the 20,000-point notebook reproduction, a Chromium regression, fresh generated bundles, and 187 passing targeted tests.