Stop the PNG exporter from deleting text it cannot draw

CHANGELOG.md

@@ -89,6 +89,15 @@ in the README).
   never appended report exactly what they did before.
 
 ### Fixed
+- **The native raster exporter silently deleted every character outside its
+  glyph atlas.** No glyph and no advance, so `Zürich` came out of `to_png()` as
+  `Zrich`, a `format="€,.0f"` axis lost its symbol on every tick, and a CJK tick
+  label exported as blank space — while the same figure's SVG rendered all of
+  them correctly. The atlas now carries Latin-1 Supplement and Latin Extended-A
+  (every Latin-script language) plus non-ASCII currency, and anything still
+  unbaked — CJK, Cyrillic, emoji — renders as the U+FFFD replacement box with
+  its advance reserved, so the limitation is visible rather than silent (§28).
+  Costs ~48 KB of baked coverage (+4% on the core dylib).
 - **The categorical palette cycled per trace instead of per series.** A box is
   four traces and a stem is two, so four box series under a four-color
   `xy.theme(palette=...)` all wore `palette[0]`, and eight box series drew two

scripts/gen_font.py

@@ -21,6 +21,15 @@
 
 # Printable ASCII. Axis labels/titles/legends mostly stay within this set.
 FIRST, LAST = 0x20, 0x7E
+# Latin-1 Supplement + Latin Extended-A letters: every language written in the
+# Latin script that xy is likely to label a chart in. Without these a `Café` or
+# `Zürich` tick label lost the accented letter *entirely* — no glyph, no
+# advance, no warning — while the same chart's SVG rendered it correctly.
+# Ranges, not a hand-typed string, so the coverage is auditable.
+_LATIN = "".join(chr(c) for c in [*range(0x00C0, 0x0180)] if chr(c).isalpha())
+# Currency beyond ASCII `$`. `xy.y_axis(format="€,.0f")` is a first-class
+# feature, so its symbol has to survive to_png().
+_CURRENCY = "¢£¤¥₣₤₦₩₪₫€₭₮₱₲₴₹₺₽₿"
 # Extra codepoints for the pyplot shim's TeX-subset → unicode conversion
 # (greek, super/subscripts, math operators) plus typography mpl emits (−, µ).
 EXTRA = sorted(
@@ -30,6 +39,13 @@
         "⁰¹²³⁴⁵⁶⁷⁸⁹⁺⁻⁼⁽⁾ⁿⁱ"
         "₀₁₂₃₄₅₆₇₈₉₊₋₌₍₎ₐₑₕᵢₖₗₘₙₒₚᵣₛₜᵤᵥₓ"
         "×·±∓≤≥≠≈∞√°→←∂∇∫∝∈−–—‘’“”…µ"
+        "§¶†‡•‰≡⌀"
+        + _LATIN
+        + _CURRENCY
+        # U+FFFD is the fallback the rasterizer substitutes for anything still
+        # missing, so a character xy cannot draw shows up as a visible box
+        # instead of vanishing (§28: never silent).
+        + "�"
     )
 )
 PX = 16  # render size; runtime scales this coverage bitmap to the requested size.

spec/api/styling.md

@@ -698,12 +698,20 @@ so small labels are slightly less refined than a browser's.
 **Native text coverage.** The atlas is a generated grayscale DejaVu Sans sheet
 (`src/font.rs`, regenerated by `scripts/gen_font.py`) baked at a 16 px base
 cell and blitted and scaled at runtime, so glyph coverage is a fixed set: ASCII
-32–126 plus 110 enumerated extras (Greek, common math operators, arrows,
-super/subscripts, typographic quotes and dashes). A codepoint outside that set
-— CJK, Cyrillic, Arabic, emoji, most accented Latin — has no glyph and is
-**silently skipped**, with no tofu box and no advance reserved, so a title,
-tick label, legend entry, or annotation containing one renders shortened rather
-than raising. Use `engine=xy.Engine.chromium` for full Unicode text.
+32–126 plus 329 enumerated extras — the Latin-1 Supplement and Latin Extended-A
+letters (every language written in the Latin script), non-ASCII currency
+symbols, Greek, common math operators, arrows, super/subscripts, and
+typographic quotes and dashes. `format="€,.0f"` and a `Zürich` tick label
+therefore survive `to_png()` intact.
+
+A codepoint outside that set — CJK, Cyrillic, Arabic, emoji — renders as
+U+FFFD, the replacement box, and reserves that glyph's advance. It used to be
+**silently skipped**, with no box and no advance, so `Zürich` exported as
+`Zrich` and a CJK label exported as blank space while the same figure's SVG was
+correct; §28 asks that a decision the engine makes on the user's behalf be
+visible, and a box is the visible form of "this renderer cannot draw that".
+Zero-width and control characters are still dropped — they have nothing to
+show. Use `engine=xy.Engine.chromium` for full Unicode text.
 
 The atlas bounds the native **raster** formats (PNG, JPEG, WebP) only. The
 other two native formats carry their own text contracts: SVG emits real

spec/design/rust-engine.md

@@ -88,12 +88,12 @@ src/                                          # 15,423 lines shipped, 8 modules
                         #       byte-identical to the serial path) and the fused
                         #       raster→PNG encode via `png`/fdeflate.
                         #   No unsafe; owns the crate's one third-party dep.
-  font.rs       (1039)  # generated by scripts/gen_font.py — do not hand-edit.
+  font.rs       (2503)  # generated by scripts/gen_font.py — do not hand-edit.
                         #   Baked DejaVu Sans grayscale coverage atlas for
-                        #   raster.rs: 205 glyph records (advance, w, h, left,
+                        #   raster.rs: 424 glyph records (advance, w, h, left,
                         #   top, coverage offset/len) at BASE_PX=16 plus the
                         #   row-major coverage bytes, and EXTRA_CODEPOINTS, the
-                        #   sorted table of the 110 non-ASCII codepoints. Data
+                        #   sorted table of the 329 non-ASCII codepoints. Data
                         #   only — no shaping, no FreeType, no unsafe. Coverage
                         #   limits are a product constraint, not a detail: §2.1.
   css.rs         (788)  # tiered CSS value/color validation behind `xy_css_check`.