Position stacking is computed globally instead of per-facet-panel

[cpsievert]

Summary

When combining stacked bars with faceting, position stacking accumulates globally across all facet panels instead of resetting to 0 within each panel. This causes bars in later panels to "stack on top of" values from earlier panels.

Reproducible example

Using the built-in penguins dataset:

SELECT * FROM ggsql:penguins WHERE sex IS NOT NULL
VISUALISE island AS x, species AS fill
DRAW bar
FACET sex

Observed stacking values

Facet	Island	Species	pos2end (bottom)	pos2 (top)
female	Biscoe	Adelie	0.0	22.0
female	Biscoe	Gentoo	44.0	102.0
male	Biscoe	Adelie	22.0 ❌	44.0
male	Biscoe	Gentoo	102.0 ❌	163.0
female	Torgersen	Adelie	0.0	24.0
male	Torgersen	Adelie	24.0 ❌	47.0

Expected

Each facet panel should stack independently from 0:

Facet	Island	Species	pos2end (bottom)	pos2 (top)
female	Biscoe	Adelie	0.0	22.0
female	Biscoe	Gentoo	22.0	80.0
male	Biscoe	Adelie	0.0 ✅	22.0
male	Biscoe	Gentoo	22.0 ✅	83.0

(Note: the "expected" Gentoo values above are illustrative — the exact counts would differ since the current stacking also seems to use global cumulative counts rather than per-panel counts.)

Root cause

In src/plot/layer/position/stack.rs:273, the cumulative sum only partitions by the group column (x position):

col(&stack_col)
    .cum_sum(false)
    .over([col(&group_col)])   // ← only groups by pos1 (x), not by facet columns
    .alias("__cumsum__"),

It should also partition by facet columns so stacking resets within each panel. In ggplot2, position adjustments are computed per-panel (after faceting).

Inconsistency with ggplot2

In ggplot2, this works correctly because position adjustments happen per-panel:

library(ggplot2)
library(palmerpenguins)
ggplot(na.omit(penguins), aes(x = island, fill = species)) +
  geom_bar() +
  facet_wrap(~sex)
# Each panel stacks independently from 0

[teunbrand]

Is this a problem just for stacking or for more position adjustments? I think the proposed solution to add the panel columns to the grouping columns has the right shape, but if there are more issues it might need to adjust the scope.

[thomasp85]

huh, I was sure I had already added panel info to that. Thanks for flagging it

[cpsievert]

Investigated — dodge and jitter are both affected by the same issue.

Dodge: compute_group_indices(&df, &layer.partition_by) includes facet columns, so with fill=["X","Y"] and facet=["F1","F2"], it computes n_groups=4 instead of n_groups=2. This makes adjusted_width = 0.9/4 = 0.225 instead of the correct 0.9/2 = 0.45.

Jitter: Same path when dodge=true — inflated group count produces incorrect dodge offsets.

Filed #254 and fixed in this PR alongside the stacking fix. Added a shared non_facet_partition_cols() helper that filters facet columns from partition_by using spec.facet, used by both dodge and jitter.