Fallback fonts: probe glyph coverage with ttf-parser before building the fontdue face

The first string containing a glyph outside Sora stalled the caller for over a second, even on a desktop i7. `system_fonts::lookup_handle` walked the fallback chain building a `fontdue::Font` for every slot just to consult its `cmap`, and fontdue parses the whole face up front: reaching a dingbat such as U+2733 meant paying for `NotoSansCJK-Regular.ttc` (19 MB) although the glyph is not even there and only DejaVu, the last slot, carries it. In forge that showed as the app switcher taking more than a second to appear the first time a window title carried such a symbol; the top bar never triggered it because its text is all Latin.
Each slot now has two lazily resolved stages: the file bytes, read once, and the fontdue face. Coverage is answered from the bytes alone with `rustybuzz::ttf_parser::Face::parse(...).glyph_index(ch)`, which only touches the tables it is asked for, and the fontdue face is built solely for the slot that actually owns the glyph. A CJK codepoint still pays the CJK parse, since it has to be rasterised, but a symbol that lives in DejaVu no longer drags the collection into memory. The sticky negative results are kept at both stages, and the shaping-side `bytes` in `FontHandle` now share the `Arc` with the coverage stage instead of a second copy.

src/system_fonts.rs

@@ -90,43 +90,61 @@ const FALLBACK_FONT_CANDIDATES: &[ FallbackFontSpec ] =
 	FallbackFontSpec { path: "/usr/share/fonts/truetype/dejavu/DejaVuSans.ttf",            face: 0 },
 ];
 
-/// Per-slot lazy state. `Vec` length matches
+/// Per-slot lazy state, two stages so probing a slot for coverage never
-/// [`FALLBACK_FONT_CANDIDATES`]; each `OnceLock` resolves
+/// pays the fontdue parse: `bytes` is the file read once, `font` is the
-/// independently so the act of looking up a Devanagari codepoint
+/// fontdue face built only when a glyph in this slot is actually needed.
-/// doesn't drag the CJK pack into memory. `None` inside a resolved
+/// `None` in a resolved stage is a sticky negative result.
-/// slot means the file was missing or fontdue rejected it — a sticky
+struct Slot
-/// negative result so subsequent misses skip the slot in O(1).
-fn slots() -> &'static [ OnceLock<Option<FontHandle>> ]
 {
-	static SLOTS: OnceLock<Vec<OnceLock<Option<FontHandle>>>> = OnceLock::new();
+	bytes: OnceLock<Option<Arc<Vec<u8>>>>,
+	font:  OnceLock<Option<FontHandle>>,
+}
+
+fn slots() -> &'static [ Slot ]
+{
+	static SLOTS: OnceLock<Vec<Slot>> = OnceLock::new();
 	SLOTS.get_or_init( ||
 	{
 		( 0..FALLBACK_FONT_CANDIDATES.len() )
-			.map( |_| OnceLock::new() )
+			.map( |_| Slot { bytes: OnceLock::new(), font: OnceLock::new() } )
 			.collect()
 	} )
 }
 
-/// Try to load and parse the fallback font at slot `idx`. Each
+fn slot_bytes( idx: usize ) -> Option<Arc<Vec<u8>>>
-/// `OnceLock` wraps `Option<FontHandle>` so a missing or malformed
+{
-/// file is recorded as `None` and never re-attempted. The raw bytes
+	slots()[ idx ].bytes.get_or_init( ||
-/// are preserved inside the handle so the same `Arc<Vec<u8>>` can be
+	{
-/// handed to rustybuzz for shaping without re-reading the file.
+		std::fs::read( FALLBACK_FONT_CANDIDATES[ idx ].path ).ok().map( Arc::new )
+	} )
+		.clone()
+}
+
+/// Whether slot `idx` owns a glyph for `ch`, answered from the `cmap`
+/// alone: ttf-parser only touches the tables it is asked for, so this
+/// stays cheap even for a multi-megabyte CJK collection.
+fn slot_covers( idx: usize, ch: char ) -> bool
+{
+	let Some( bytes ) = slot_bytes( idx ) else { return false };
+	let face = FALLBACK_FONT_CANDIDATES[ idx ].face;
+	rustybuzz::ttf_parser::Face::parse( &bytes, face )
+		.ok()
+		.and_then( |f| f.glyph_index( ch ) )
+		.is_some()
+}
+
+/// Build (once) the fontdue face for slot `idx`. The raw bytes are
+/// preserved inside the handle so the same `Arc<Vec<u8>>` can be handed
+/// to rustybuzz for shaping without re-reading the file.
 fn slot_handle( idx: usize ) -> Option<FontHandle>
 {
-	let slot = &slots()[ idx ];
+	slots()[ idx ].font.get_or_init( ||
-	slot.get_or_init( ||
 	{
-		let spec  = &FALLBACK_FONT_CANDIDATES[ idx ];
+		let bytes = slot_bytes( idx )?;
-		let bytes = std::fs::read( spec.path ).ok()?;
+		let face  = FALLBACK_FONT_CANDIDATES[ idx ].face;
-		let opts  = FontSettings { collection_index: spec.face, ..FontSettings::default() };
+		let opts  = FontSettings { collection_index: face, ..FontSettings::default() };
 		let font  = Font::from_bytes( bytes.as_slice(), opts ).ok()?;
-		Some( FontHandle
+		Some( FontHandle { font: Arc::new( font ), bytes, face } )
-		{
-			font:  Arc::new( font ),
-			bytes: Arc::new( bytes ),
-			face:  spec.face,
-		} )
 	} )
 		.clone()
 }
@@ -136,14 +154,6 @@ fn slot_handle( idx: usize ) -> Option<FontHandle>
 /// `Arc<Font>` for the rest of the process. Returns `None` if no
 /// installed fallback covers the codepoint — the caller then paints
 /// the primary font's `.notdef` rather than dropping the glyph.
-///
-/// Side effect: walking the chain may load and cache a slot even if
-/// it doesn't end up covering `ch` (`lookup_glyph_index` reads the
-/// `cmap` table, which requires the font to be parsed). That's
-/// acceptable — the slot is cached on the first encounter regardless,
-/// and most coverage gaps in early slots are the small Noto Sans
-/// scripts (Devanagari, Arabic, …) whose total weight is a fraction
-/// of the CJK pack everyone was paying for unconditionally.
 pub fn lookup( ch: char ) -> Option<Arc<Font>>
 {
 	lookup_handle( ch ).map( |h| h.font )
@@ -225,13 +235,7 @@ pub fn primary_handle() -> FontHandle
 /// without re-reading the font file.
 pub fn lookup_handle( ch: char ) -> Option<FontHandle>
 {
-	for idx in 0..FALLBACK_FONT_CANDIDATES.len()
+	( 0..FALLBACK_FONT_CANDIDATES.len() )
-	{
+		.find( |&idx| slot_covers( idx, ch ) )
-		let Some( handle ) = slot_handle( idx ) else { continue };
+		.and_then( slot_handle )
-		if handle.font.lookup_glyph_index( ch ) != 0
-		{
-			return Some( handle );
-		}
-	}
-	None
 }