Stability of layer data attributes

[teunbrand]

This PR aims to fix #3175.

It avoids dropping attributes in some places and actively restores them in other places.
Attributes should now be stable from the moment aesthetics are computed up to the Geom$draw_*() methods.

[teunbrand]

I cannot figure out why coverage is failing and R CMD check passes.

[yutannihilation]

I too cannot figure out, but I confirmed this error is reproducible on my local. As it looks too cryptic, maybe we can just ignore the failing tests for now? cf. https://covr.r-lib.org/#exclusions

══ Failed tests ════════════════════════════════════════════════════════════════
── Failure ('test-layer.R:177:3'): attributes on layer data are preserved ──────
attr(ld, "test") (`actual`) not equal to "preserve me" (`expected`).

`actual` is NULL
`expected` is a character vector ('preserve me')
── Failure ('test-scales.R:71:3'): identity scale preserves input values ───────
`d1` (`actual`) not equal to `d2` (`expected`).

     names(actual) | names(expected)     
 [1] "shape"       - "size"          [1] 
 [2] "size"        - "alpha"         [2] 
 [3] "alpha"       - "x"             [3] 
 [4] "x"           - "y"             [4] 
 [5] "y"           - "colour"        [5] 
 [6] "colour"      - "fill"          [6] 
 [7] "fill"        - "shape"         [7] 
 [8] "PANEL"       | "PANEL"         [8] 
 [9] "group"       | "group"         [9] 
[10] "stroke"      | "stroke"        [10]

[ FAIL 2 | WARN 1 | SKIP 258 | PASS 1593 ]

[teunbrand]

I think I've made some debugging process. I think for some reason, ScalesList$transform_df() correctly tells the plot to skip on identity transformations in most contexts, but (for reasons beyond my current understanding) we do not skip this in the coverage test. The clue for this comes from #6076, where identity scales protest in check_transformation(), though it has no business checking identity transformations.

(df = df) at ggplot2/R/scales-.R:102:5
 11.                 └─ggplot2 (local) transform_df(..., self = self) at ggplot2/R/ggproto.R:188:17
 12.                   └─base::lapply(df[aesthetics], self$transform) at ggplot2/R/scale-.R:520:5
 13.                     └─ggplot2 (local) FUN(X[[i]], ...)
 14.                       └─ggplot2 (local) transform(..., self = self) at ggplot2/R/ggproto.R:[188](https://github.com/tidyverse/ggplot2/actions/runs/11951275132/job/33694697269?pr=6194#step:5:189):17
 15.                         └─ggplot2:::check_transformation(...) at ggplot2/R/scale-.R:640:3

So this laid bare that I hadn't included scale transformations as preserving attributes, and now this is included :)

[thomasp85]

Can you please add a test for being able to modify the attribute during computations

[teunbrand]

Added a test. I should note that within the 'layer - panel - group' hierarchy attributes are unstable. So if you want to append some attributes, you should do it at the layer-level. This is because if you do it at e.g. the panel level and one panel returns a different attribute value than another, there is no principled way of merging attributes.