cascade's last fallback searched the loaded faces and, finding none that
draws the cluster, called `first` while the search's read guard was still
alive (a temporary in the match arm lives to the end of the `let`).
`first` looks up the default family, which loads it through add_fonts,
and add_fonts takes the faces' write lock even when there is nothing to
add: the thread waited on itself, and every other thread then waited on
the loading lock it held.
It shows when a run's chain is empty and nothing loaded draws the
character before the default family has been looked up, so it depends on
which test (or document) gets there first: deckcraft-text's tests hung in
4 of 10 parallel runs here (all threads parked in system_fallback /
try_family, one in add_fonts' write), and in 0 of 15 with this change.
The search now finishes before `first` runs. New test
falling_back_to_the_default_face_does_not_deadlock reaches that path on a
fresh database; it times out on main and passes with the fix.
"Z̴̡̛͓a̸̧̘l̵̩g̶̞o̸̤" in an Aptos run drew its marks as missing-glyph boxes: no face in the chain
covered the whole cluster and the fallback for its base letter neither, so the run's own font was
kept. PowerPoint draws the marks stacked. A loaded plain face that covers the whole cluster is
now taken before falling back to the run's font. w09-scripts s1 8.40 -> 7.86.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The previous change sent every text-presentation pictograph to a symbol font, which also moved
bullets such as "▪" (Wingdings substitutes) from the face they had to a wider one: necc-
presentation-0 s13 18.78 -> 19.72. Now a pictograph is redirected to a plain face only when its
fallback would be the colour emoji font; keycaps and U+FE0E sequences still take a plain face.
necc s12-s14 back to 26.38 / 18.78 / 20.91; w14-xml s7 (pictographs) 9.79 -> 7.49 overall.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Observed on PowerPoint's export of pictographs in Aptos, Arial and Segoe UI:
- characters that default to emoji presentation draw in colour without a selector: ⭐✅☔⌚⏩
and the other BMP Emoji_Presentation characters (DeckCraft only knew the supplementary-plane
ones and drew ⭐✅ as text);
- text-presentation pictographs (⚠ ♥ ✔ ☎ ⬆ ➡ ▶ ⤴) draw as text even where the colour font is the
first installed font that has them: they now come from a plain symbol font, symbol families
first, never the colour one; and a symbol after an emoji no longer inherits its colour face;
- keycap sequences ("1️⃣", "#⃣") draw as text, an outlined key, not the colour keycap.
w14-xml s7 (pictographs) 9.79 -> 6.16, w09-scripts s1 (emoji) 11.56 -> 10.27.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* Emoji: colour glyphs, and emoji take the colour emoji font
Emoji were drawn as black outlines from a symbol font, with ZWJ sequences
(families, flags) falling apart into their parts. Now:
- an emoji cluster (emoji presentation, U+FE0F, ZWJ between pictographs,
skin tones, regional indicators, keycaps) takes an installed colour font
(COLR, CBDT, sbix or SVG tables) before the document's fonts, so its
sequences ligate as in PowerPoint; U+FE0E keeps a plain face;
- COLR v0 and v1 glyphs are painted in their CPAL colours (solid fills,
linear, radial and sweep gradients, transforms, composite modes), with
palette index 0xFFFF in the text colour.
A tiny COLRv0 test font (two squares, CC0, regenerated by
make_colr_font.py) backs the render test.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* Render: lock skrifa for colour glyphs
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* Sprint 1: bidi and Arabic typography engine
fonts: ISO 15924 script and a:latin/a:ea/a:cs slot classification;
shape_range shapes an item with the whole paragraph as HarfRust
pre/post context, explicit script/language/direction, glyphs kept in
visual order; per-grapheme-cluster fallback (slot face, theme script
font, OS cascade families, bundled craft-fonts face); GDEF ligature
carets.
text: UAX #9 levels for the whole paragraph (a:pPr/@rtl), itemization
by level, script, face and shaping style (paint-only changes no longer
split words), logical line breaking at cluster boundaries, per-line
L1/L2 reordering (L4 mirroring by the shaper); RTL start edge and
bullets, justLow kashida; caret Affinity, logical/visual movement,
bidi hit testing and selection as the union of visual boxes.
model/pptx: theme a:font script fonts and algn="justLow" round-trip.
engine: format.rtl (Right-to-Left Text Direction), justifyLow align.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* Fonts: the bundled script face comes before installed fonts; no Microsoft family names
Arabic, Hebrew, Thai and Devanagari clusters try the craft-fonts face first, so every machine
draws the same, then openly licensed or OS-native installed families, then any installed face.
Installed Microsoft faces are still used through that scan, but not named (AGENTS.md §1.1).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Tayeb <tayeb@192.168.11.179>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Brandon Thomas <echelon@gmail.com>
Магда М.Г.