A zoomed bin grid does not match the unzoomed grid when the viewport covers the data #124

Problem

A zoomed binned trace (histogram, histogram2d, geo_histogram2d) bins on a lattice that starts at 0, with bin width span / n (snap_range in flexviz/trace/bin_grid.py). The unzoomed grid starts at the data minimum. So a viewport that covers the same range as the unzoomed grid gives another grid:

• The grid almost always holds n + 1 bins.
• The edges move, and discrete data can get an empty bin.

Example: a heatmap with y data 1..7 and y_bins=7. Unzoomed, it has 7 rows, one for each value. A zoom to [1, 7] gives 8 rows on [0.857, 7.714], and the row from 5.14 to 6.0 is empty.

Where users see it

• A manual zoom to the data range (on main today).
• The first zoom from the unzoomed view: all edges move, also when the range stays almost the same.
• A lock on a link group whose members show different ranges (fix(adapters): aggregate linked members at the range a lock pins #116). The lock writes the group viewport for every member, also for the clicked figure, because the validator requires equal viewports in a group. So the clicked heatmap re-bins although its display did not change.

Cause

A zoomed request does not know the data extent of the column. The lattice needs an anchor that does not move under a pan, and 0 is the only anchor that needs no data.

Possible fix

Give zoomed binned requests the data extent of the column. Then do one of these:

• Start the lattice at the data minimum. Pans stay stable, and a viewport that covers the data gives the unzoomed grid.
• Treat a viewport that covers the data extent as unzoomed.

Cost and scope

• One extra min/max pass per zoomed binned request. Earlier measurements: about 25–33 ms on in-memory data, about 100 ms zoomed on a 100M-row scan (geo).
• It touches the 1-D and 2-D grids, the cube target domain of a zoomed histogram, and the client hover arithmetic (perf(adapters): resolve linked-hover bins from the edge triples by arithmetic #119).
• Skip viewport requests that cannot change the result: span-covered pans and wheel bursts #34 proposes to send the domain per axis on the delta. The same domain can serve both issues.

Found in the review of #116.