Mermaid compatibility history

How each native family stage was brought up and verified against the pinned Mermaid 12.1.0 build (commit 21f72f07). The living reference is the compatibility checklist; these records keep the evidence each stage was accepted on.

Agent flow stage

The 19 pinned bodies in tests/fixtures/mermaid/agentflow/ render natively. Global declarations release nodes and flows from enclosing ownership; nested flow children retain their direct container. Inline flow metadata applies after its body closes. Collapsed flows hide descendants, redirect crossings to the outermost summary and preserve genuine authored self-links. Expanded headers have compact source-linked selection regions.

Metadata mappings merge shallowly, remove prototype keys recursively, preserve quoted/block-scalar punctuation and retry multiline trailing commas without removing content commas. Parsing rejects object tags and aliases and bounds map items and nesting. The pinned contract validator's console warnings, exact js-yaml scalar/alias equivalence and automatic ELK placement are non-goals. Named edge metadata is retained separately from vertices; named edges take their CSS classes, as pinned; edge animation is browser CSS, a non-goal. Removed shapes use the rounded task fallback; unknown shape names produce located errors.

Gantt stage

The 43 verbatim pinned bodies in tests/fixtures/mermaid/gantt/ render natively. Twelve deterministic cases captured from the pinned upstream database check fractional/calendar durations, month-end clamping, millisecond/second epochs, implicit starts, inclusive dates, multiple dependencies and excluded-day completion. Logical completion and drawn completion remain separate, so a dependent task can start after an excluded-day gap. Includes override exclusions, and Friday/Saturday weekends are supported.

Task bars and milestones share a proportional axis. Vertical markers consume no ordinary task row. Section headings, task bars and markers retain source origins; short horizontal bars take precedence over vertical markers in inline row maps. Calendar resolution, exclusion shading and source-map assembly yield to the event loop and can be abandoned before publishing a partial schedule. Task, date, tick, calendar-iteration and scene bounds limit retained work; overly dense tick requests fall back to automatic calendar-aligned ticks.

The today marker draws as pinned .today does, 2px in todayLineColor, and todayMarker CSS styles it with commas for semicolons: stroke, stroke-width and a non-zero stroke-dasharray apply, and a declaration a browser would drop leaves the default. Compact mode and the gantt configuration are covered under Configuration. dateFormat reads the dayjs customParseFormat tokens YYYY to ZZ, SSS, X and x; the rarer Do, S and SS read as literal text. axisFormat covers every d3-time-format directive. Browser callbacks are descriptive metadata. The ordinary native palette and terminal-adapted spacing are used.

Use-case stage

All twenty pinned e2e/platform/dev-diagrams/diagrams/use-case bodies parse and export natively. The reference fixture captures the pinned executable parser's actors, use cases, groups, members, relations and notes. References can be refined by later declarations; conflicting kinds, generated IDs, labels, shapes, stereotypes, metadata and boundary ownership produce located errors.

Native glyphs cover normal, hollow, awesome and business actors, ellipse and rectangular use cases, business slashes, package boundaries, association endpoints, hollow generalization triangles, include/extend links and note connectors. JSON cards preserve source property order, nested paths, scalar-array rows and ECMAScript numeric display. Graphics use shared SVG/PNG primitives; cell previews retain readable labels and semantic source selection.

config.usecase supports actor/use-case font sizes and weights, node/rank spacing and diagram padding, with explicit zero surviving overlays. Spacing converts upstream units to terminal scene units. Font-family requests use the bundled font fallback; useMaxWidth remains governed by the preview/export caller's viewport. Backtick labels on actors, use cases, notes and relationships draw their Markdown runs, as pinned; exact upstream routing is a non-goal. Unregistered icons draw the pinned unknown-icon tile and keep the authored icon name in the legend.

The parser, lexer, metadata, symbol, relationship, JSON and style limits have focused regressions. Eighteen hand mutations were killed, including role checks, resource limits, generated UTF-16 IDs, JSON exponents, overlay precedence, named edge styles and marker stroke widths.

Architecture stage

All eighteen pinned architecture bodies match semantic data captured with the pinned Langium grammar and value converters. Full-document SVG exports include the pinned seed and randomize configuration examples. IDs, declared-parent ordering, duplicate names, endpoint roles, group traversal and alignment members are validated before layout. Titles retain the grammar's quote escaping and raw entity text; native drawing decodes entities safely.

The native constraint grid aligns opposing ports, applies authored row/column hints and separates colliding peers. Groups are placed as opaque boxes, then expanded into nested frames. Service and group-boundary ports feed orthogonal routes with optional arrowheads at either end; self-loops use exterior paths. Contradictory alignments relax collision placement to keep glyphs separate.

Built-in server, database, disk, internet and cloud glyphs, text icons, missing icons and junction dots share native SVG/PNG primitives. Label masks keep links from obscuring text. Selection targets service/junction bounds and compact group headers, rather than the full group frame. Source data stays unchanged by layout.

Nonzero seeds produce reproducible native tie ordering; seed zero uses the injected random generator. Explicit false and zero survive document overlays. This grid is an adaptation to terminal cells, not the upstream fcose solver. The fcose solver, its tuning options and obstacle-aware routing are non-goals; other icons draw the unknown tile, as pinned. Twenty meaningful hand mutations guard the stage; architecture's own focused tests cover every line.

Event-modeling stage

Thirty-three pinned documentation, rendering and parser examples match semantic data captured from the executable Langium grammar and event-modeling database. The fixtures preserve time/reset frame IDs (including significant zero padding), role aliases, qualified names, explicit sources, data blocks and references, notes, and given/when/then clauses. Source provenance is recorded alongside them.

Inference follows the pinned database: an ordinary frame without explicit inputs links from the nearest preceding frame in another lane. A reset frame suppresses that inference. Explicit input references disable inference even when unresolved; resolved inputs deduplicate in source declaration order, and forward frames do not become available early. Qualified lane identities retain the pinned source-order allocation. Native placement additionally keeps returning cards apart so their text remains readable in terminal cells.

Native graphics show role colors, ordered lane bands, source-linked frame IDs, data compartments and input arrows. Referenced data uses the first declaration and overrides inline data. Opaque JSON, HTML, URI, Markdown and embedded-DSL examples remain text, never executable content. Notes, declared entities and criteria remain source-linked readable legends rather than extra graph nodes. Synthesized lane headers are excluded from selection.

Lexing yields trivia as well as visible fields, so long comment streams remain cancellable. Located errors and bounds cover duplicate frames, malformed fields, 64 KiB tokens, 131,072 lexical work steps, 8,192 fields per declaration, 4,096 declarations and 99 qualified lanes per role band. Twenty meaningful hand mutations guard inference, reset/explicit-input behavior, ordering, resource limits, data precedence, locations, role colors, non-overlap and selection.

Pixel configuration (padding, rowHeight), browser typography and exact upstream geometry are non-goals. These are not implied by corpus acceptance or full line coverage.

Venn stage

The eighteen pinned rendering examples in tests/fixtures/mermaid/venn/ are compared with venn-reference.json, captured from the pinned Jison grammar and original database. Front matter is stripped for these body comparisons. Geometry regressions verify area ratios, requested pair overlap, clipping area/containment, determinism, source coordinates across CRLF comments, UTF-16 identifier ordering, relative geometry after title translation, styles, degenerate regions and bounds. Eighteen targeted mutations were killed, including the clipping interpolation mutation after adding an intersection-area assertion. PNG examples were inspected. The combined Mermaid run passed 2,483 tests / 23,183 assertions with full Mermaid source line coverage; fifteen IDE render/export cases passed 105 assertions, and all seven static gates passed. A clean final full repository verification remains pending with the rest of the compatibility work.

Railroad stage

The 86 verbatim accepted documentation, rendering and parser bodies in tests/fixtures/mermaid/railroad/ are compared with railroad-reference.json. railroad-provenance.json records each upstream source. The executable reference uses the four pinned Langium grammars, original token builders/value converters and original parser-to-renderer tree transformations. Undefined/Infinity values serialize to JSON null, represented by a nullable repetition maximum natively. All bodies render through native rows and SVG. Additional tests cover finite counts, escaped strings, metadata-named rules, reserved constructors, title-token prefix behavior, comments/CRLF, exception/lookahead transforms, branch non-overlap, optional bypasses, return loops, source links and malformed/excessive input. Twenty meaningful mutations were killed; the finite-maximum mutation was rerun without its earlier warning-producing harness variant. PNG tracks were inspected. The combined Mermaid run passed 2,604 tests / 23,903 assertions with full Mermaid source line coverage; nineteen IDE render/export cases passed 133 assertions. All seven static gates passed. Full repository verification is recorded separately.

The combined composer verify after railroad passed with 100% core source line coverage (32,040/32,040) and own-plugin line coverage (24,844/24,844).

Modern flowchart shape stage

flowchart-shapes.json records the 53 ordinary shape definitions and their 150 unique names/aliases from the pinned rendering-elements/shapes.ts registry. All map to native silhouettes and internal marks, including documents/stacks, open data stores, storage cylinders/buckets, windows, persons, braces, tags and label-free symbols. Curves are bounded sampled terminal geometry. The same silhouettes drive edge intersections and pointer picking. Labels on label-free symbols remain in semantic search/source data. Unknown authored shapes produce located errors; Porta's earlier round/ellipse extensions are retained. Authored fill/stroke, width and dash presence propagate to internal marks.

Regressions compare alias SVGs, check fill:none, decorative boundaries, stacked layers, empty corners, cylindrical cap alignment/ring clearance, circular person heads and long styled labels. All 53 PNGs were reviewed as a contact sheet and braces/cylinder captions corrected. Twenty-three meaningful mutations were killed. The combined Mermaid run passed 2,767 tests / 28,091 assertions with full Mermaid source line coverage. Twenty IDE render/export cases passed 140 assertions, and all seven static gates passed. The preceding full repository verify through the railroad stage is green; a final full verify remains pending after later work.