Changelog

All notable changes to the pdfrb gem will be documented in this file.

[Unreleased]

Added — Parity batches 4-9

  • TODO 103 Hyphenation — Knuth-Liang Pdfrb::Layout::Hyphenation with curated English patterns in data/pdfrb/layout/hyphenation_en.txt.
  • TODO 116 PDF/X-6 — Pdfrb::Conformance::PdfX#X6 ISO 15930-11 profile.
  • TODO 117 PDF/VT — Pdfrb::Conformance::PdfVT (VT-1, VT-2) ISO 16612-2 profiles.
  • TODO 118 Multi-revision traversal — Document#each_revision walks the /Prev chain yielding (revision_index, xref, trailer).
  • TODO 119 Table cell spans — TableBox.cell(box, colspan:, rowspan:), hash-form cells, span-aware layout.
  • TODO 120 Multi-page table — MultiPageTableBox splits a table across pages with header-row repeat.
  • TODO 122 Bidi — Pdfrb::Layout::Bidi paragraph-level UAX #9 reordering; TextLayouter reorders Hebrew/Arabic for visual order.
  • TODO 128 Linearization-aware reader — Source::LinearizationReader.detect(io) parses the first indirect object's dict for /Linearized.
  • TODO 129 Type3 font loader — Pdfrb::FontLoader::Type3 with glyph procedures + encoding.
  • TODO 131 Inline images (write) — Canvas#inline_image(dict:, data:) plus Operator::{BeginInlineImage, EndInlineImage}.
  • TODO 133 Custom canvas primitives — polyline, polygon, arc, circle, ellipse, rounded_rectangle, dash=.
  • TODO 135 PDF thumbnail generation — Pdfrb::Task::Thumbnail builds /Thumb image XObjects on every page.
  • TODO 136 PDF/A preflight deep — PREFLIGHT_DEEP rule set on PdfA (JavaScript, PostScript XObjects, encryption, embedded-file bans for A-1/A-2/A-4).
  • TODO 137 PDF/UA tagging deep — list/LI/LBody, Table/TR, TH/Scope structural checks.
  • TODO 138 Tagged PDF validation — Pdfrb::Conformance::TaggedPdf baseline structural rule set.
  • TODO 140 PDF 2.0 AF validation — Pdfrb::Conformance::Pdf2AF App Note 002 validator.
  • TODO 104 Default ICC profile — Color::DefaultProfile emits a usable sRGB v4 ICC profile with desc/wtpt/rXYZ/gXYZ/bXYZ/rTRC/ gTRC/bTRC tags and a real MD5 Profile ID.
  • TODO 139 ICC profile validation — Pdfrb::Color::ICCValidator checks signature, version, device class, color space per ICC.1:2022 §7.
  • TODO 108 Image downsampling — Pdfrb::Image::Downsampler + Task::Optimize#downsample_images! for FlateDecode 8-bpc non-palette images.
  • TODO 130 Color-key masking — ImageLoader::PNG parses tRNS and emits /Mask color-key array.
  • TODO 131 Inline images (read) — Content::Parser recognises the BI/ID/EI sequence and yields an InlineImage invocation.
  • TODO 132 Page-piece dict — Pdfrb::Model::Type::PagePieceInfo per s14.5.
  • TODO 134 Tagged PDF alt-text API — Document::Structure facade for /Alt, /ActualText, /Lang.
  • Layout — cluster-aware default TextShaper, real AFM-based FontFallback, PolygonFrame with proper inscribed-rectangle scan, MultiCellTextLayout, JustificationKashidas.
  • Encryption — StandardSecurityHandler.for_v5 write-side AES-256 R6 handler (file key + /U//O//UE//OE//Perms), V5Writer PKCS#5 algorithms, PublicKeySecurityHandler with real CMS EnvelopedData construction via OpenSSL::PKCS7.
  • Signatures — Pdfrb::Conformance::Pades (B-B, B-T, B-LT, B-LTA profiles), Pdfrb::Conformance::Ltv long-term validation, Pdfrb::Conformance::PdfA4Deep, PdfUA2Deep, DigitalSignature::TimestampClient (RFC 3161 TSA over HTTPS).
  • Images — ImageLoader::TIFF decodes uncompressed RGB/gray; ImageLoader::GIF does full GIF-spec LZW decode with palette.
  • Source — extended Pdfrb::Source::Recovery with trailer-reference recovery, hybrid-xref detection, object-stream rebuild.
  • Tasks — Pdfrb::Task::RegenerateAppearances walks widget annotations and rebuilds /AP /N via Appearance::Generator.

Fixed

  • Stream deduplication in Task::Optimize (was a stub).
  • Writer#write_xref_stream now emits a free entry for every gap in the oid space (was misaligning the reader).
  • Document::Form#flatten! now stamps field appearances into page content via /Do (was a stub).
  • Issue #63: PostScript name extraction for /BaseFont.
  • Issue #64: Resources#empty? handles Hash, Array, Cos::Dictionary, PdfArray.

Changed

  • Release workflow switched to emf2svg-ruby pattern: direct OIDC trusted publishing via rubygems/configure-rubygems-credentials@main
    • plain gem push. Bump job commits version.rb and tags vX.Y.Z.
  • Removed all instance_variable_set/get, send to private methods, respond_to? for type checks.

[0.7.0] — 2026-08-08

Added — Release infrastructure

  • .github/workflows/rake.yml — Metanorma CI generic-rake workflow for tests on every push, PR, tag, and dispatch.
  • .github/workflows/release.yml — Metanorma CI rubygems-release workflow driven by workflow_dispatch with next_version input or repository_dispatch of type do-release.
  • RELEASE.md — documents the publish flow and required secrets.

Added — Image subsystem

  • Pdfrb::Image::Audit — read-only metadata extractor (ImageInfo struct per image XObject) with each_image, all, oversized helpers.
  • Pdfrb::Image::Downsampler — pure-Ruby nearest-neighbour for FlateDecode-encoded 8-bpc non-palette images. PNG predictor round-trip on both decode and encode. JPEG/JP2K skipped (no native deps).
  • Pdfrb::Task::Optimize#downsample_images!(factor:) — applies the downsampler across every image XObject in the document.

Added — Layout bidi

  • Pdfrb::Layout::Bidi — paragraph-level UAX #9 bidi reordering (P2-P3 paragraph level, BD1 type classification, I1/I2 implicit levels, L2 reverse by level, L4 paired-bracket mirroring).
  • Pdfrb::Layout::TextLayouter#layout calls Bidi.reorder before line breaking, so Hebrew and Arabic render in visual order.

Fixed

  • Issue #63: Document::Fonts#add now extracts the PostScript name (name table nameID 6) and uses it in /BaseFont (e.g. ABCDEF+MinionPro-Regular instead of ABCDEF+EmbeddedFont3132).
  • Issue #64: Model::Type::Resources#empty? handles Hash, Array, Cos::Dictionary, and PdfArray values via a single empty_collection? helper.

Changed

  • Pdfrb::Task::Optimize.dedup_streams! now actually rewrites references to duplicate streams (was a stub).
  • Pdfrb::Writer#write_xref_stream emits a free entry for every gap in the oid space, so dedup'd oids don't misalign the reader's per-oid index.
  • Pdfrb::Document::Form#flatten! stamps each field's /AP /N appearance stream into the page content via the /Do operator (was a stubbed no-op).

[0.3.0] — 2026-08-02

Added — Semantic PDF diff

  • Pdfrb::Compare — semantic PDF comparison engine. Compares two PDFs at the structural level: page count, per-page text similarity (Levenshtein), font inventory diffs, image count delta, outline (bookmark) diffs, metadata (/Info) diffs, and structural (Catalog + Trailer key) diffs.
  • Pdfrb::Compare::Report — immutable report with equivalent?, similarity (0..1), summary string, and to_h serialisation.
  • CLI: pdfrb diff LEFT RIGHT — compare two PDFs from the command line. Exits 0 if equivalent, 1 if different.
  • Programmatic API: Pdfrb::Compare.compare(left, right) and Pdfrb::Compare.equivalent?(left, right). Accepts bytes, IO, or Document objects.

Metrics

  • 550 specs (was 537), 0 failures, 4 pending.
  • 0 rubocop offenses.
  • 182 → 185 lib files; 49 → 50 spec files.

[0.2.1] — 2026-08-02

Added

  • Form XObjects (Document::FormXObject) — create reusable form templates with a Canvas; register on pages and draw via /Do operator.
  • TTF subsetting (Font::TrueType::Subsetter) — real glyph subsetting with composite glyph support, loca/glyf/cmap/hmtx rewriting, graceful fallback to full embedding.
  • Outline circular-reference fix — bookmarks now store References (not raw Dictionaries), preventing infinite serialization recursion.
  • Document::Outline round-trips — flat and nested outlines survive write + read correctly.

Metrics

  • 537 specs (was 534), 0 failures, 4 pending.
  • 0 rubocop offenses.
  • 180 → 182 lib files; 47 → 49 spec files.

[0.2.0] — 2026-08-02

Added — P0 feature implementations

  • CMap writer (Font::CMap::Writer) — generates valid /ToUnicode CMap data from glyph-code → Unicode mappings. Supports 1-byte and 2-byte codespaceranges, supplementary Unicode (UTF-16 surrogate pairs), multi-codepoint ligatures, and automatic chunking (≤100 entries per beginbfchar/endbfchar section per PDF spec). Round-trips through Font::CMap::Parser.

  • Document::Files (Associated Files / EmbeddedFiles) — embeds files as /Type /EmbeddedFile streams referenced by /Type /FileSpec dicts, stored in the Catalog's /Names /EmbeddedFiles name tree. Supports MIME types, descriptions, and PDF 2.0 /AF relationship tagging. Round-trips through write + read.

  • XRef stream writer (PDF 1.5+) — emits binary XRef streams instead of classical xref tables. Configurable via config["writer.use_xref_stream"] = true. /W [1 3 1] entry format with FlateDecode compression. Round-trips correctly.

  • Object stream packing (/Type /ObjStm) — packs eligible small objects (non-stream, non-encrypted, < threshold bytes) into compressed object streams. Configurable via `config

    true+config`. Reduces file

    size by 20–50%.

  • Task::Optimize — real implementation (was a no-op stub). Enables FlateDecode compression, XRef stream writing, and ObjStm packing in one call: Pdfrb::Task::Optimize.call(doc, io: out).

  • Document::Outline (bookmarks/outline write-side) — creates /Outlines tree on Catalog with flat and nested entries. Each entry has /Title, /Parent, /First//Last//Next//Prev links.

  • Fixed CMap Parser — regex bug: beginbfchar line matching didn't handle N beginbfchar format (with count prefix). Fixed to match anywhere in the line. Also added surrogate-pair decoding for supplementary Unicode CMaps.

Configuration additions

config["writer.use_xref_stream"]      # bool, default false
config["writer.pack_object_streams"]  # bool, default false
config["writer.object_stream_threshold"]  # int, default 200

Metrics

  • 534 specs (was 503), 0 failures, 6 pending.
  • 0 rubocop offenses.
  • ~85% line coverage.
  • 178 → 180 lib files; 44 → 47 spec files.

[0.1.1] — 2026-08-02

Housekeeping

  • Layout/Composer extracted to sibling arroolio gem. Pdfrb is now pure PDF: bytes ↔ model. No page templates, no flows, no Knuth-Plass, no tables, no SVG. Removed lib/pdfrb/layout/, lib/pdfrb/composer.rb, lib/pdfrb/layout.rb, spec/pdfrb/layout/, TODO.layout/.
  • Removed rexml runtime dependency (was only needed for the FO parser, which moved to Arroolio).
  • Fixed all 12 rubocop offenses. 0 offenses remaining. Disabled RSpec/DescribeClass (legitimate string describes in integration tests).
  • Added SimpleCov (opt-in via COVERAGE=1 env var). Current line coverage: 84.5%. Grouped by Source / Model / Filters / Content / Font / Encryption / Tasks / CLI / Document.
  • Replaced 3 pending specs with skip — clearer exit semantics for strict CI runs.
  • Fixed Model::Cos::Stream#as_parms_list — Hash parms now correctly wraps to [parms] instead of Array(parms) which produced [[:k, :v], ...]. Bug was hidden by tests that didn't exercise DecodeParms as a Hash through the Stream facade.
  • Renamed CMap format dispatchers — parse_format_4 → parse_format_four, etc. (rubocop Naming/VariableNumber).
  • README/CLAUDE.md updated to reflect pure-PDF scope; layout concerns documented as living in arroolio.

Verified

  • 503 examples, 0 failures, 3 pending (conformance specs awaiting fixture corpus).
  • 0 rubocop offenses.
  • gem build pdfrb.gemspec produces pdfrb-0.1.1.gem.
  • End-to-end round-trip: build PDF → write → read → extract text → matches.

[0.1.0] — 2026-08-02

Initial release

Pure-Ruby PDF library: byte-level reader, Arlington-model-driven typed domain model, and serializer.

  • Source: bytes → tokens → COS graph.
  • Model: typed COS values + Type::* semantics, sourced from vendored Arlington PDF Model TSVs (ISO 32000-2:2020).
  • Filter pipeline: FlateDecode, ASCII-Hex, ASCII-85, LZW, RunLength, CCITTFax, DCT, JPX, JBIG2, Crypt.
  • Content: 52 content-stream operators + Canvas drawing API.
  • Encryption: RC4 (pure Ruby), AES (OpenSSL), StandardSecurityHandler (V1-V6 / R2-R6).
  • Font machinery: AFM parser, 5 encoding tables, CMap parser, TrueType file parser (head/hhea/hmtx/cmap/OS-2), Type1 metrics.
  • Image loaders: JPEG, PNG, PDF page import.
  • Tasks: ExtractText (with /ToUnicode CMap), ExtractImages, Merge, Optimize.
  • CLI: version, info, tree, merge, extract-text/images, encrypt, decrypt, optimize, form.
  • Conformance: PDF/A + PDF/UA rule subsets.
  • Round-trip property tested against ISO 32000-2 Annex H examples (12 PDFs).

[Unreleased]

Added — 2026-08-01 (layout engine session)

Full layout engine + XSL-FO pipeline (TODO.layout 01–17).

  • Pdfrb::Layout::FontMetrics (AfmMetrics + TrueTypeMetrics) + GlyphMeasurer — real glyph widths for the 14 standard Type1 fonts (via Adobe AFM, now vendored) and any TTF/OTF (via head/hhea/hmtx/ cmap/OS-2 table parsers). Replaces the 0.5×font_size estimate; matches Adobe metrics to 0.01pt.
  • Pdfrb::Layout::TextLayout — Greedy and Knuth-Plass line breaking with real glyph metrics, alignment (left/right/center/justify), hyphenation hooks, and break-opportunity finder that handles multi-run paragraphs (bold spans, hyperlinks).
  • Pdfrb::Layout::PageFlow + Flowable + Frame — flowing-text engine with split-on-page-break for paragraphs/tables, keep-together, page-break-before/after, and a two-pass static-content render for fo:page-number and fo:page-number-citation.
  • Pdfrb::Layout::Table + TableLayout::Fixed + TableLayout::Auto
    • TableFlowable + TableRenderer — fixed-width and CSS 2.1 auto-width column layout, row/col spans, header row repeat across page breaks.
  • Pdfrb::Layout::PageTemplate + Region + PageSequenceMaster + StaticContent — page geometry with 5 regions (body/before/after/ start/end) and odd/even/first/last/blank page-template selection.
  • Pdfrb::Layout::Leader + FieldRuns (PageNumberField, PageNumberCitationField, PageCountField) — TOC leaders and page-number fields, resolved via the two-pass layout.
  • Pdfrb::Layout::SVG::* — SVG → PDF renderer supporting path data, rect/circle/ellipse/line/polyline/polygon, text, image (with data: URI), nested groups with transforms, and presentational styling. New SVG element = one Element subclass.
  • Pdfrb::Layout::FO::* — XSL-FO parser (REXML-based) + PropertyResolver (inheritance + shorthand expansion) + MasterBuilder + FlowBuilder + Compiler. End-to-end: Pdfrb.render_fo(xml, io:) produces a PDF.
  • Pdfrb.parse_fo(xml) and Pdfrb.render_fo(xml, path_or_io:) top-level entry points.

Fixed — 2026-08-01 (layout engine session)

  • Pdfrb::Font::AFMParser — handle StartCharMetrics 315 (with count), not just bare StartCharMetrics.
  • Pdfrb::Font::GlyphList — accept both canonical hex format and literal-Unicode repackagings of glyphlist.txt.
  • Pdfrb::Filter::FlateDecode — out << unfiltered.pack("C*") so PNG predictor output is concatenated correctly (was appending Array to String).
  • Pdfrb::Model::Cos::Stream#decoded_stream — use type-checked as_filter_list / as_parms_list helpers instead of Array(...), which silently converts a Hash DecodeParms to [[:k, :v], ...] and bypasses the predictor.
  • Pdfrb::Source::ObjectReader#load_from_objstm — unpack the [oid, value] pair from ObjectStreamReader and re-wrap via document.wrap so type-dispatch (Catalog, Pages, etc.) works for compressed objects.
  • Pdfrb::Document::Pages#count — dereference Reference-typed /Count before treating as Integer; fall back to walking the Kids tree.
  • Pdfrb::Document::Pages#walk — robust against Reference-typed /Kids arrays (common in FOP output).

Added — 2026-07-30 (session 2)

Phase 2 — COS layer (TODOs 08–14).

  • Pdfrb::Model::Object — base wrapper with value, oid, gen, document; indirect?, must_be_indirect?, deref, pdf_type.
  • Pdfrb::Model::Reference — frozen value object by (oid, gen), hashable, comparable, serialises as "5 0 R".
  • Pdfrb::Model::PdfArray < Object — Enumerable wrapper around an Array of PDF values, with full mutation API and oid/gen/document support for indirect arrays.
  • Pdfrb::Model::Rectangle — immutable [llx lly urx ury] value with width, height, left/right/top/bottom.
  • Pdfrb::Model::Matrix — immutable [a b c d e f] affine with the PDF-correct column-vector composition (M1 * M2 applies M2 then M1) and translate/scale/rotate/skew factories.
  • Pdfrb::Model::Date — PDF date D:YYYYMMDDHHmmSSOHH'mm' <-> Time.
  • Pdfrb::Model::Cos::NameEncoding — Symbol <-> /... wire format with #xx escape encode/decode.
  • Pdfrb::Model::Cos::StringEncoding — text/binary string helpers: UTF-16BE BOM detection, full PDFDocEncoding table (Appendix D.2), PDFDocEncoding <-> UTF-8 conversion.
  • Pdfrb::Model::Cos::Fields::Field — typed field metadata with resolved_types, valid_value?, Arlington slot.
  • Pdfrb::Model::Cos::Fields converters: Dictionary, Array, String, PDFByteString, Date, Rectangle, Integer — selected by usable_for? and applied in order via Fields.apply.
  • Pdfrb::Model::Cos::Dictionary — base class with define_field, field, each_field, type registry (register_type/lookup_type), []/[]= with full pipeline (default -> dereference -> unbox -> convert), each/each_raw, validate.
  • Pdfrb::Model::Cos::Stream < Dictionary — raw-bytes holder with decoded_stream / encoded_stream= that drives Pdfrb::Filter.
  • Pdfrb::Document — top-level facade with wrap, add, object/dereference, register_listener/dispatch_message.

Phase 5 partial — Filters (TODOs 34–36).

  • Pdfrb::Filter module + registry; Filter.apply(bytes, filters:, parms:, direction:, document:) chains decoders/encoders.
  • Pdfrb::Filter::Base — included interface; subclasses call register_as "Name".
  • Pdfrb::Filter::FlateDecode — zlib + PNG predictor unfiltering (None/Sub/Up/Average/Paeth).
  • Pdfrb::Filter::ASCIIHexDecode — hex <-> bytes with EOD marker >.

Phase 6 — Arlington layer (TODOs 45–53).

  • Vendored data/pdfrb/arlington/latest/*.tsv — 613 TSV files from ~/src/pdfa/arlington-pdf-model/tsv/latest/. Single source of truth for PDF object model.
  • Pdfrb::Arlington::Loader — lazy TSV reader with (name, version) memoisation. object_definition("Catalog") returns a parsed ObjectDefinition.
  • Pdfrb::Arlington::ObjectDefinition / FieldDefinition — typed records for one TSV / one row. FieldDefinition exposes bare types, required/predicate-split required, links, possible-value lists.
  • Pdfrb::Arlington::Type — the 17 type symbols + validity check.
  • Pdfrb::Arlington::PdfVersion — Comparable value object with full .from_tsv_cell parsing (plain, fn:Extension(name, ver), fn:Eval).
  • Pdfrb::Arlington::Predicate::Lexer/Parser/AST/Evaluator/Context — full fn: mini-DSL: tokeniser, recursive-descent parser (logical ops require parens per spec), evaluator with @key, parent::, trailer:: path resolution, and BinOp/UnaryOp/LogicalOp evaluation.
  • Pdfrb::Arlington::Predicate::Functions registry + 6 built-in groups (OCP: one file per category):
    • version.rb — SinceVersion, BeforeVersion, Deprecated, IsPDFVersion, Extension.
    • presence.rb — IsRequired, IsPresent, NotPresent.
    • logical.rb — Eval, Not, And, Or.
    • arithmetic.rb — ArrayLength, StringLength, RequiredValue, BitSet, BitsClear, BitsSet, KeyNameIsColorant, NoCycle.
    • reference.rb — MustBeDirect, MustBeIndirect.
    • file.rb — FileSize.
  • Dictionary.arlington_object "Name" — pulls field metadata from the matching TSV at class-load time and feeds it through define_field. Type mapping from the 17 Arlington symbols to Ruby classes (booleans -> [TrueClass, FalseClass], name -> Symbol, rectangle -> Rectangle, etc.).

Phase 15 partial — fixture corpus sourcing (TODOs 148, 158).

  • claricle/pdf-core-examples forked from pdf-association/pdf-core-examples (private). The ci team is on both claricle/pdfrb and claricle/pdf-core-examples, so claricle-ci (already on ci) can pull examples in CI using CLARICLE_CI_PAT_TOKEN.
  • TODO 148 updated: corpus is fetched (not committed) via rake fixtures:pull; spec/fixtures/pdf-core-examples/ is gitignored.
  • TODO 158 added: GitHub Actions workflow that pulls the corpus before rspec.

Stats

  • 163 specs, 0 failures, 0 rubocop offenses.
  • 39 source files under lib/, 12 spec files under spec/.
  • 613 vendored TSVs under data/pdfrb/arlington/latest/.
  • 158 TODO files in TODO.general-rels/ (added 158 for CI).

Code quality

  • Pure Ruby autoload everywhere; no require_relative in lib/.
  • No respond_to? for type checks, no instance_variable_set/get, no send to private methods.
  • # frozen_string_literal: true on every .rb.
  • Real instances in specs, never double().

Added — 2026-07-30

0.1.0 — initial scaffold.

The pdfrb gem is created as a pure-Ruby PDF library modelled on the sibling postscript gem and HexaPDF, with the PDF object model driven directly by the vendored Arlington PDF Model TSVs.

  • Project skeleton: gemspec, Gemfile, Rakefile (default = spec + rubocop), .rspec, .rubocop.yml (Cimas base + rspec/performance/ rake plugins), .gitignore, .gitattributes (TSV EOL=LF).
  • Pdfrb::PdfConstants — PDF lexical constants (header magic, EOL, whitespace, delimiters, version enumeration, keywords) per ISO 32000-2:2020 s7.2.
  • Pdfrb::Error hierarchy — typed errors for every failure mode: ParseError (with source_position), LexError, SyntaxError, MalformedPdfError (with recovered?), SerializeError, FilterError (with filter_name), EncryptionError, UnsupportedVersionError, ObjectReferenceError (with oid/gen), ValidationError (with field/predicate names).
  • Pdfrb::Configuration — per-Document frozen settings hash with deep-merge. Defaults for malformed-PDF recovery, auto-decrypt, invalid-string handling.
  • Pdfrb::DataDir — resolves paths under the gem's bundled data/pdfrb/ (Arlington TSVs, AFM, ICC, etc.).
  • Top-level lib/pdfrb.rb autoload tree for all planned namespaces (Source, Model, Arlington, Filter, Content, Encryption, DigitalSignature, Font, FontLoader, ImageLoader, Layout, Task, Serializer, Writer, Importer, Revision(s), XrefSection, Document, Composer, CLI). Lazy: file not loaded until constant accessed.
  • Rakefile arlington:refresh task to re-vendor TSVs from ~/src/pdfa/arlington-pdf-model/tsv/latest/.
  • CLAUDE.md onboarding doc.
  • TODO.general-rels/ work-breakdown structure (157 entries across 16 phases).

Architecture

Mirrors postscript's layered split (Source / Model / Serializer) extended for PDF's needs. The differentiator vs HexaPDF is that field metadata is loaded at runtime from the vendored Arlington TSVs (single source of truth) rather than hand-coded via define_field.

Code quality

  • Pure Ruby autoload everywhere; no require_relative in lib/.
  • No respond_to? for type checks, no instance_variable_set/get, no send to private methods.
  • # frozen_string_literal: true on every .rb.
  • Real instances in specs, never double().