Essenfont is assembled, not drawn. Five stages take 38 donor fonts and produce one OpenType Collection with five plane subfonts, all reproducible from the committed manifest in essenfont/essenfont.

The whole pipeline at a glance

   donor fonts (TTF / OTF / TTC / WOFF2 / CBDT)
            │
            ▼  Fontisan::FontLoader
   CpMap  ─── per-codepoint { donor_label, gid } ─── ucode's universal-set manifest
            │
            ▼  Essenfont::Otc::PlanePartitioner
   Blueprint ── 5 Partitions (BMP / SMP / SIP / TIP / SSP), each ≤ 65,484 glyphs
            │
            ▼  Essenfont::Otc::StitcherSession + Fontisan::Stitcher
   5 stitched subfonts (outlines + propagated CBDT/CBLC)
            │
            ▼  Fontisan::Collection::Builder  (table dedup by SHA256)
   Essenfont-Regular.otc  ── validated by Fontisan::Collection::Reader

The five stages, in detail

  1. Read donors — Open each of the 38 donor files via fontisan’s FontLoader. Each donor declares its license, its sha256, and the codepoint ranges it covers. The build refuses anything outside the OFL-compatible policy before a single glyph is read.

  2. Build the codepoint map (CpMap) — For every assigned Unicode 17 codepoint, decide which donor owns it. ucode’s universal-set manifest is the authoritative per-codepoint donor assignment. PUA, surrogate, and reserved codepoints are filtered out. CBDT-only color-bitmap donors are kept out of the outline map.

  3. Partition by plane — Group the 159,866 codepoints into named subfonts, each under the 65,535-glyph cap. The partitioner groups by Unicode plane — BMP, SMP, SIP, TIP, SSP — which is MECE, stable across version bumps, and discoverable in O(1) via cp >> 16.

  4. Stitch subfonts — Copy each donor’s outline glyphs into the right subfont via fontisan’s Stitcher. Outlines are stitched first, so cmap.build’s first-wins policy binds codepoints to real outline GIDs. CBDT/CBLC color-bitmap tables are propagated afterward via raw-byte table_data reads.

  5. Pack the collection — Pass the compiled subfonts to Collection::Builder, which writes the OTC. Tables are deduplicated by SHA256 across the five faces. After writing, Fontisan::Collection::Reader opens the file back up and asserts face count, per-face glyph count ≤ 65,535, and cmap union size ≥ 99% of the input.

Why partition by plane: the 65,535-glyph cap

TrueType and OpenType address glyphs by a 16-bit ID, so a single face can hold at most 65,535 glyphs (one reserved for .notdef). Unicode 17 has 159,866 assigned codepoints — nearly 2.5× the cap.

Subfont Range Carries
:plane_0 (BMP) U+0000–U+FFFF World scripts — Latin, Cyrillic, Greek, Arabic, Hebrew, CJK
:plane_1 (SMP) U+10000–U+1FFFF Historical scripts, emoji, music, math symbols
:plane_2 (SIP) U+20000–U+2FFFF CJK Extensions B / C / D / E / F / I
:plane_3 (TIP) U+30000–U+3FFFF CJK Extension C + J, Tangut, Khitan Small Script
:plane_14 (SSP) U+E0000–U+EFFFF Language tags (~100 codepoints)

Why planes (not script families, not blocks)?

  • MECE. Every codepoint lives in exactly one plane — no gaps, no overlap.
  • Stable. Plane assignment never changes under a Unicode version bump.
  • O(1) lookup. cp >> 16 gives the plane number directly.
  • Semantically meaningful. Each plane is a coherent unit to ship in one face.

The default cap is PlanePartitioner::DEFAULT_CAP = 65,484. If a future Unicode version balloons one plane past the cap, the partitioner sub-splits by block: plane_2_a, plane_2_b, etc.

Outline format: glyf vs CFF2

  • glyf (TrueType) — quadratic Béziers, hinting, universally supported.
  • CFF2 (OpenType 1.8+) — cubic Béziers, variation stores, ~35% smaller via subroutinization.

CFF1 is not shipped — CFF2 strictly dominates it.

How color emoji survives into the OTC

Noto Color Emoji glyphs are CBDT/CBLC (color bitmaps), not outlines. Two problems solved:

  1. CBDT donors filtered from outline map — prevents empty .notdef placeholders.
  2. CFF2 preserved on round-trip — table bytes are read raw via table_data and spliced directly.

Per-block WOFF2 subsetting for this website cannot yet carry CBDT (fontisan subsetter drops unrecognized tables), so color-emoji blocks render in monochrome on the site. See Coverage & gaps.

Reproducibility

git clone https://github.com/essenfont/essenfont
cd essenfont
bundle install
bundle exec ucode fetch fonts

bundle exec ruby scripts/build.rb
bundle exec ruby scripts/verify.rb release/Essenfont-Regular.otc

Donor binaries are never committed to git. Each donor’s sha256 is verified before use.

Next: Donor strategy →