Binary (TTF/OTF) import is lossy — design a proper import path (components, glyph set, names, variations)

[kostyafarber]

Summary

The binary reader (crates/shift-backends/src/binary/reader.rs) is currently a "draw what skrifa gives us" importer. That's fine as a preview, but as an import path into an editor it silently discards most of the structure we care about. Found while dogfooding with a subset Space Grotesk VF (see the shift-dogfood repo): the TTF import can't exercise component decomposition, anchors, or variable editing because none of that data survives.

#48 already tracks the fvar slice of this. This issue is the umbrella for the rest: what a real binary import path should preserve.

What gets lost today

1. Components are flattened to contours.
font_from_skrifa() (binary/reader.rs:147) draws every glyph through skrifa's OutlinePen with DrawSettings::unhinted(...) (reader.rs:179), which resolves composite glyphs into outlines. A font with 652 composites imports as 100% contours — there is nothing to decompose, and editing a base glyph no longer propagates. Instead we should read glyf composite records directly (component glyph id + transform) and build Component IR, the same shape the designspace reader produces.

2. Unencoded glyphs are skipped entirely.
The import loop iterates char_map.mappings() (reader.rs:173), so anything without a cmap entry — a.alt, .case/.sc variants, and unencoded marks used as composite bases — never imports. Should iterate the full glyph set (maxp/glyf), not the charmap.

3. Glyph names ignore the post table.
Names are synthesized as the literal character or uniXXXX (reader.rs:193). A post format 2.0 font carries production names (eacute, acutecomb) — we should use them, falling back to AGL-style names only for post format 3. Literal-character names (é) are also going to be hostile to anything keyed by glyph name (features, kerning groups, .shift round-trips).

4. All variation data is dropped.
Both metrics and outlines are read at LocationRef::default() only (reader.rs:154, reader.rs:179). #48 covers reading fvar axes; the bigger design question for a variable-first editor is what to do with gvar: reconstruct editable masters from the deltas (min/default/max per axis, plus intermediates where gvar has them), or import static-at-default and tell the user. Either is defensible — silently static is not.

5. No anchors (expected, worth documenting).
Anchors don't exist in compiled fonts; they're baked into GPOS mark-attachment at build time. Deriving anchors from GPOS mark/mkmk lookups is possible and would make imported binaries much more editable, but it's inference, not recovery — fine to declare out of scope, in which case the import UX should say so.

Scope notes

• OTF/CFF has no composites at all (CFF charstrings are flat, modulo legacy seac), so item 1 is TTF/glyf-only; the CFF path stays outline-based.
• Dispatch is shared: both TTF and OTF route through BytesFontAdaptor (font_loader.rs:111-112).
• Once 1–4 land, a shift convert font.ttf out.shift CLI subcommand becomes worth exposing — the pieces (binary::read_font_file + ShiftSourcePackage::save_font) already exist, it's just not worth wiring up while the reader is this lossy.

Suggested order

1. Full glyph set + post names (2, 3) — small, unblocks everything keyed by name
2. Preserve glyf composites as Component IR (1)
3. fvar axes (TTF/OTF loader does not read fvar — variable font axes are lost #48), then the gvar/masters design decision (4)
4. Decide and document the anchor policy (5)

Acceptance sketch

Importing the shift-dogfood subset TTF (397 glyphs, 169 composites, wght 300–700) should yield: 397 glyphs including unencoded ones, eacute as a live component of e + acutecomb with production names, and either editable wght masters or an explicit "imported static at default location" notice.