[thebhef]

Two plot-panel defects with one cause between them: two mechanisms were
each deciding where a marker goes, and neither was choosing well.

Points: On looked like extrapolation. It thinned markers to a flat
500 across the visible range on a uniform index stride. A uniform stride
over a min/max envelope aliases onto one leg of it, so the dots hugged
one side of a line that was oscillating through both and most extrema
carried none at all — the markers were on the line, the wrong ones were
being chosen. The cap is gone: on marks every served sample. Nothing
is unbounded by that, because the serve already is, to a few points per
canvas pixel column.

An enum lane is now an ordinary series with an overlay on top. It is
served, plotted and marked exactly as a numeric stepped series is (the
host half landed in the previous commit), and the tiles are an overlay
computed from runs of equal consecutive served values — shaping
already-paged data for the renderer, not re-deriving a model fact. The
lane's private marker pass is deleted, and with it the reason it existed:

• the tiles draw in the drawAxes hook, under the series layer,
  so uPlot's markers land on a tile instead of beneath 65-85 % of
  opacity. Only the labels are held back for the draw hook — a tile's
  label is the only place the held value's name appears, and the one
  part of a lane the waveform may not cross;
• a code is normalised into the tile band rather than the whole lane
  band. The tile is the centred 60 % of the lane, so a table's extreme
  codes used to plot in the gap between lanes — for a two-code enum,
  code 0 sat below its own tile entirely — and a marker on the plotted
  value landed off the tile it belongs to;
• a lane's marker ink is measured against the tile the way its label's
  already is: the accent wherever it clears 3:1, otherwise the extreme
  opposite the theme's background. One ink per lane, because whether the
  accent survives a tinted tile is a property of the theme rather than
  of the value. The single-enum ribbon keeps the plain accent — its line
  plots the real code against a real y scale, so its markers are as
  often off the ribbon as on it.

What a lane loses is its exemption from uPlot's density rule under
auto; it now falls under the same rule, and the same minimum-sample
floor, as a slow line on a shared axis. That is the point of unifying
them.

The categorical fetch flag goes with the host's reducer, so
DecimatedRequest no longer carries a render mode and the fetch memo
key no longer keys on one.

Reviewers: the draw-order change is the load-bearing part — tiles before
the series layer, labels after — and PlotArea.draw.test.ts drives both
passes the way uPlot runs them. The new cases there pin a marker per
served sample with no stride, a lane's markers coming from the shared
points.filter, no marker left on a cell the extrapolation blanking
takes, and a held state arriving as one tile at the window width and
point budget the categorical serve was introduced against.

Co-Authored-By: Claude Fable 5.1 noreply@anthropic.com