Bump to 1.3.0. Establish a single source of truth for category->pip-slot
mapping: each data row is one pip placed at its category's factor-level
index, with ndots shared between the plotted layout and the legend. This
removes the paste0() id join and the reliance on scaled dots values.

Correctness fixes:
- character dots decoded to the wrong pip position (row-order dependent)
- n=6 plot layout did not match the legend design
- panel/legend slot-count desync (absent levels, drop=FALSE, faceting)
- non-injective dots palette crashed draw_panel
- paste0(cat,x,y) id collisions drew phantom pips / cross-contaminated aes
- pip diameters fed to GeomPoint size aesthetic (~24% too small, overflow);
  pips now drawn as true circleGrob mm circles
- mapped size renormalised per panel instead of globally
- min_fill inverted the size encoding for pip_scale < 0.25
- NA/invalid sizes mishandled (silent drop vs full-size resurrection)

Robustness / API:
- ndots required and validated (was an opaque base-R error)
- forward ... to layer(); restore unknown-parameter warnings
- validate pip_scale in (0,1]; clamp the pip-centre shift
- per-tile tile_df aesthetics (mapping alpha/linewidth no longer crashes)
- per-axis pip packing; proportional make_offsets padding
- graceful fallback for non-default coordinate systems
- clear error for >6 dot categories at build time

Docs / data:
- geom_dice() @examples uses long-format data (comma strings were never split)
- regenerate sample_dice_data1/2 to the documented 160-row structure

Add tests/testthat/test-logic-fixes.R covering every fixed issue.
R CMD check --as-cran clean; full testthat suite passing.