Time Series

New Distribution chart — a quantity spread over a fixed set of named bins, read against a reference distribution. It is a new component rather than an extension of TimeSeries because three of that component's properties are chronological and cannot be parameterised out: its axis keys are sorted (['Thu','Fri','Sat'] sorts into a week that does not exist), a line asserts the metric passed through every value between two samples when between two hourly aggregates there is nothing to pass through, and delta() first-to-last is arithmetic performed on a circle. An unmeasured bin **hatches**. Everywhere else a null can simply be dropped, but for bars a height of zero and a bar never drawn are the same picture — this is null ≠ zero at its hardest case. bin.note prints a second axis reading under the first, replacing the timezone toggle a caller would otherwise build; a toggle hides half the truth and is missing from every screenshot. Also: error states on every chart that fetches — TimeSeries, Sparkline, MetricTable, NetworkGraph, Distribution — resolved through resolveDataState so the precedence is the array in data-state-model rather than five hand-rolled ladders. In a dashboard "the fetch failed" and "there is no data" are different statements. Delta and SeriesTable get none: neither has a fetch lifecycle. StatItem gains tone, carried by three signals rather than a hue alone — the rail, the value colour, and an sr-only judgement.

TimeSeries can plot more than one metric, and it finally draws an x axis. compare takes further series — { points, label, unit } — drawn against the same axes. It is purely additive: points is still the required single-series prop it always was, so nothing that already calls this component changes. Each line gets a --chart-N hue **and** a distinct dash pattern, and the new legend swatch repeats the dash, so two lines stay two lines in a greyscale print and to a red-green colour-blind reader. Five series are drawn (the size of the palette); any beyond that stay in the data table and the legend says how many. All series share one y domain, and that is not configurable. A second y axis lets an author slide two unrelated metrics until they appear to cross wherever the argument needs them to; a series two orders of magnitude smaller will render as flat here, which is the true statement about it. The x axis is a baseline, a tick per labelled slot, and up to five HTML labels below the plot — HTML because SVG text scales with the viewBox, and at a 320 viewport this plot is 288px wide against a 900-unit box, so text-xs inside it paints at 4px. Below sm the middle labels drop and the two ends stay. The crosshair readout now names and values every series at the cursor, from the one aria-live <output> that the arrow keys already fed. There is deliberately no second, hover-only tooltip to fall out of step with it.

The visualization layer ships: a token-driven chart set built on the VISUALIZATION_PHILOSOPHY row-first doctrine — MetricTable as the centrepiece, TimeSeries as the thing a row is promoted into, plus a network graph, sparklines and deltas. Every chart carries the DS-wide loading skeleton opt-in.