What earns a concept

A concept sits in the vocabulary only when real charts demand it and no composition of the rest can draw it. Both clauses. Everything else is a preset or an example.

Why#

Chart libraries die of vocabulary. Someone asks for a chart, and it lands as a new type. The type grows its own options. A few years on, the library is a catalog of a hundred near-duplicates. barh beside bar. Heatmap beside Image beside Matrix. None of them compose. All of them disagree in small ways. Each one is a maintenance bill that does not end. The cost to the user is worse than the maintenance cost. A vocabulary that large cannot be learned. It can only be searched.

The opposite failure is purism. Refuse a concept the grammar genuinely cannot express, and people hand-roll the hard part. The hand-rolled versions get the statistics wrong. That is how a plotting ecosystem ends up with a hundred incorrect quartile implementations.

The idea#

One membership test. Two clauses, both required.

• Real charts demand it. Charts people actually draw, not charts a completeness argument names.
• No composition of the existing marks, stats, and scales reproduces its rendered output.

The eight marks pass. A violin is composable from Area once the KDE exists, but the KDE itself is not composable from marks, so the stat earns a seat and the violin stays a preset. Cells earns one mark, not three. Value grids, rgb images, and categorical regions are one geometry with three color readings. A heatmap is not a mark at all. It is Cells under a colormap, and the preset says so.

The same test shapes the statistics layer. One Reducer vocabulary serves bins, groups, and windows, so "rolling p95" needs zero new concepts. The layer still refuses a fake uniform algebra. Online accumulators, reducers, and batch transforms keep distinct execution contracts. Pretending every statistic merges is how libraries ship wrong parallel medians.

A new concept must also pay for itself across features. The band scale earned its seat by serving bar charts, confusion matrices, and attention maps with one mechanism. A concept that serves one chart is that chart's implementation detail, not vocabulary.

Consequences#

• The mark family is closed, and declared complete. Adding a mark is a design event, judged by the test, never a convenience.
• A feature request gets a composition first. "From the grammar, no preset" in the gallery is the test passing in public.
• Options do not multiply. An option must be a mark channel, a stat parameter, a scale option, or a theme entry, or it does not ship.
• Statistical correctness lives in one place. One type-7 quantile implementation serves the box plot, the reducers, and the Q–Q plot.
• Removing is a contribution. A concept whose charts the grammar learns to compose is retired at the next major version.

Not this#

• No mark per chart type, and no Heatmap mark beside Cells.
• "We might need it" is not clause one. "It would be elegant" is not clause two.
• No uniform Stat trait that pretends every statistic streams and merges.
• The mark enum does not open so dependents can register marks. New geometry lands in the crate, under this test.

See Presets are packaging for where refused concepts go instead, and Vision rule 2.

Spelled today#

mark::Mark is the closed enum over Line, Points, Bars, Area, Cells, Range, Rule, Text; its docs declare the family complete. stat::Reducer is the shared aggregation vocabulary; the execution contracts are documented per type in stat (accumulators with merge laws, reducers, batch transforms). scale::Scale is the closed axis specification. The refusals are recorded in the README's "What it will not be." This section may rot; the rest must not.