Centralize type-specific logic via self-declared flags (re: #543)

[grantmcdermott]

Draft / proof-of-concept — opening move toward #543.

This is a first, deliberately small step toward the consolidation discussed in #543: moving hard-coded type == "..." logic branches out of the main pipeline (tinyplot.R, facet.R, legend.R, lim.R) and into flags that each plot type declares for itself via its data_<type>() constructor. It directly implements the design floated in that thread:

• @grantmcdermott (comment, point 1): "keep all type-specific behaviour in the dedicated constructors (data_<type> and draw_<type>), which is the principled design choice. But we couldn't figure out a way to do that at the time."
• @grantmcdermott (point 2): a par_overrides-style escape hatch on the settings container that types plug into.
• @katrinabrock: the missing_xi_means_empty_plot proposal (≈ our draw_empty_facet), and the idea that settings could be a nested named list of semantic type attributes (e.g. x_optional) rather than 1:1 flags.

Mechanism

A new settings$type_axes_hints slot (NULL by default): a named list of behaviour flags that a type's data_*() function writes via the existing env2env() write-back (same pattern as group_offsets/offsets_axis). The pipeline reads isTRUE(type_axes_hints[["<flag>"]]) instead of matching the type string, so the generic layout/legend code no longer needs to know specific types exist.

What this PR does (spineplot only, as the pilot)

Converts the six scattered type == "spineplot" checks to five self-declared flags:

Flag	Replaces (was hard-coded in)	Behaviour
self_axes	tinyplot.R dynmar block (#635)	reserve tick-row margin despite xaxt/yaxt = "n"
rhs_axis	facet.R ×2, legend.R (outer-legend bump)	reserve right margin for a secondary axis
proportional_lim	lim.R ×2	[0,1] limits; don't expand to axis breaks
legend_fill_from_col	legend.R (pt.bg)	legend swatch fill from col
draw_empty_facet	tinyplot.R empty-facet guard	draw with zero per-group rows

type_axes_hints is threaded through the exported draw_legend() (new NULL-default param + @param), build_legend_env(), and the multi-legend path.

Safety

Pure refactor — behaviour is byte-identical. All five flags are TRUE for spineplot and NULL for every other type, so each site evaluates exactly as before. Verified cmp-identical PNG renders across: simple, breaks/flip, facet, by-group, outer-legend (the RHS bump path), and the xlab/ylab = NA cases from #635.

Still to do (scope for discussion before expanding)

This PR intentionally migrates one type to prove the pattern and invite design feedback before a broader sweep. Open questions / follow-ups:

• Naming + shape. Is type_axes_hints the right umbrella, or should hints be grouped differently (e.g. a broader type_hints/settings$hints nest covering non-axis concerns too)? Per @katrinabrock, nesting keeps settings readable as this grows.
• Migrate neighbours. draw_empty_facet still co-exists with a hard-coded c("histogram","hist","rect","segments") list — those types should self-declare it too (this is exactly @katrinabrock's missing_xi_means_empty_plot). Likewise the else if (type == "ridge") legend-fill branch sits right next to the converted legend_fill_from_col check and is a natural next migration.
• Other scattered type == / type %in% checks. Inventory beyond spineplot: the boxplot-flip and ridge whtsbp checks (duplicated across tinyplot.R + facet.R), the categorical-axis type lists in facet.R, density/hist/barplot flags. Decide which are genuine "type properties" worth a flag vs. which are fine as-is (cf. @grantmcdermott point 3: branching at the primitive level — points/lines/rect — may be acceptable).
• Custom-type story (Custom types: Clarify how name impacts behavior, consider alternatives to hardcoded types in tinyplot::tinyplot #543 point 4). Once flags live on the type, document them in the types vignette so custom-type authors get the same toolbox, and revisit registering convenience strings for custom types.
• Dual conventions interim. Until the above land, the codebase has two idioms (string checks + hints). Worth a tracking note so the inconsistency is intentional.

[vincentarelbundock]

Overall, I think this is a very good path to go on. I agree we should remove the type specific logic from the main functions and put as much as we can in the type-specific functions.

The one thing that worries me a bit is not specific to the new flags you just adopted, but relates to our naming conventions. I feel that we haven't been super strict in naming variables internally, and that can make things harder to read and maintain than they should be. For example, group_offsets vs offsets_axis, where we don't follow a clear pattern, like verb_noun_adjective.

This is of course a different issue, but I wanted to raise it here because this PR will likely lead to the creation of several new internal variables.

[katrinabrock]

@grantmcdermott Let me know if you are looking for any kind of feedback from me.

---

Also, not related exclusively to this issue, I've recently made public the project I (and @rhirk) have been working on that's making use of tinyplot: https://github.com/livingingroups/trackframe

The custom types are here:

• https://github.com/livingingroups/trackframe/blob/21c9be7924d871ef7a5a09c6d25285ec4865b2d9/R/plot_type_arrows.R
• https://github.com/livingingroups/trackframe/blob/21c9be7924d871ef7a5a09c6d25285ec4865b2d9/R/plot_type_arrows_scaled.R

They don't suffer too badly from this issue, but @rhirk had previously built the type_arrows functionality based on l rather than segments (here) and I was a bit surprised and confused by how that name = argument impacted the behavior which prompted me to raise #543 in the first place