feat(render): a reduced column can be graded by dwell, not drawn as one flat band

A pixel column reduced from a dense envelope is honest about where the signal
went and silent about how long it was there: its min and max are identical
whether its samples fill the span or pile at its two ends, so a ramp and a pulse
are one band. Style::intensity_grading grades it instead -- bright where the
signal sits, faint where it only passes through -- by asking the gather for a
level finer than the pixel grid and reading the distribution from those
sub-buckets. Off by default: a change of look, not a correction of one.

The measure is dwell per row, each segment depositing one unit spread over the
rows it spans; counting +1 per row instead grades a sine's fast middle as
brightly as its rails, which is the pattern inverted. The accumulator is one
column wide (a full w*h density image is 33 MB at 4K) with each span written as
two delta writes, and normalisation is against the column's own peak -- a
frame-wide scale fades the trace out exactly where it moves fastest. Alpha is
quantised into tiers, one LineBatch and one draw call each, because the line
pipeline's colour is a uniform and a per-segment alpha is inexpressible.

The finer ask is floored at level 1 whenever the flat path would have used the
pyramid: level_for() answers 0 -- raw samples -- as soon as a sub-column falls
under one bucket, so an unfloored ask turns a moderately dense frame into an
O(visible) gather (measured at 172 samples a pixel: raw=1154 lod=0 and 6.4 ms a
frame against raw=710 lod=453 and 2.3 ms). Grading is a way of drawing a
reduction and may not undo it.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

README.md

@@ -37,6 +37,18 @@ run touched a GPU, so read it as an upper bound and a record of the software pat
   window's worth, the scroll offset is bounded by `retained - span`, so a pan or a zoom-in near an
   edge cannot leave the window's own span over data the scope never had (a window *wider* than the
   data keeps the older placement -- right edge on the newest sample);
+- a trace whose gathered window is denser than the axis is reduced **onto the pixel grid** --
+  one snapped vertical segment per integer column, lit from that column's own extremes --
+  because an LOD envelope staged as a polyline puts three or four near-vertical hairlines per
+  column at as many sub-pixel phases, and the beat between the bucket pitch and the pixel pitch
+  is a crawling moiré; one segment per column leaves one phase, so there is nothing to beat
+  (`FrameStats::column_traces`, and `docs/data-and-performance.md` for the four rules it keeps);
+- and such a column can be drawn **graded by dwell** instead of as one flat band
+  (`Style::intensity_grading`, off by default, `FrameStats::graded_traces`): bright where the
+  signal sits, faint where it only passes through, the way a digital-phosphor scope reads. It is
+  what puts back the information the reduction throws away — a column's min and max are the same
+  whether its samples fill the span or pile at its two ends — so the gather is asked at a level
+  finer than the pixel grid and the distribution read from those sub-buckets;
 - sample **markers** when a window is sparse enough that its samples are the shape
   (`Style::marker_below_samples`, 50 by default, `marker_px` across, in the trace's own colour);
 - grid with independent major/minor line classes (`Style::grid_major_visible`,

docs/data-and-performance.md

@@ -166,6 +166,116 @@ series therefore stages O(visible pixels) vertices, not O(samples), and `FrameSt
 which path each trace took. Below the threshold the raw samples are drawn instead; both
 sides go through the same instanced anti-aliased quad pipeline.
 
+## Dense traces are reduced onto the pixel grid
+
+A gathered window denser than the axis is **not** staged as a polyline. For each integer
+column of `layout.axes`, `Renderer::draw_traces` emits one exactly vertical segment from the
+lowest to the highest screen y any gathered point put in that column, flushed in framebuffer
+pixels with `u_snap` on. `FrameStats::column_traces` counts the traces that took this path.
+
+It exists because a polyline over such a window **moirés**, and three things compound to make
+it. `MinMaxPyramid::level_for()` picks the level whose buckets hold about
+`samples_per_pixel` samples, so a bucket spans **one to two pixel columns**;
+`GatherMode::Envelope` emits two points per bucket, so each bucket contributes a near-vertical
+min→max hairline *and* a near-vertical connector to the next — about 3.6 near-vertical
+hairlines per column; and each of those lands at a **fractional** x, so with ordinary
+src-alpha compositing a column's brightness is a non-linear function of the sub-pixel phases
+overlapping it. The phase pattern repeats at the beat between the bucket pitch and the pixel
+pitch, which is a visible banding, and panning sweeps the beat so it crawls. Reducing to one
+segment per column leaves the trace **one** sub-pixel phase, so there is nothing left to beat,
+and the snap then puts that one phase on the grid.
+
+Four properties of the reduction are load-bearing:
+
+- **It is gated on the gathered *density*, not on the level.** It runs when the window is
+  `GatherMode::Envelope` and carries at least two gathered points per column of the axes —
+  the raw path at 1.5–64 samples per pixel has the same beat and the same cure. The density
+  term is also what keeps it away from a *sparse* window, whose trace is drawn by the two
+  segments that enter and leave the frame (`a_sparse_window_between_samples_still_inks_both_margins`);
+  `Nearest` and `Mean` are excluded because they are one-point-per-bucket modes a band defeats.
+- **A hole stays a hole.** A column no gathered point landed in emits nothing. Bridging it
+  would draw a vertical line over x the series says nothing about.
+- **Disjoint columns are joined.** Two neighbouring columns' extremes need not overlap — a
+  fast edge rises through a whole column — and two disjoint vertical segments read as a broken
+  trace, which is what the polyline's connector prevented. The later column is extended to
+  meet the earlier one's nearest bound, so the join is made inside the column that moved.
+- **A span shorter than the stroke is drawn *as* the stroke.** The line fragment shader ramps
+  on the distance to the segment's axis and does not cap its ends, so a vertical segment
+  shorter than its own width rasterises a sliver of itself, and a flat trace's column, where
+  the two bounds are equal, nothing at all.
+
+This is the **one carve-out** in `u_snap`'s rule that chrome may be snapped and data may not
+(see `LineProgram::set_snap()`): a column *is* a pixel column and its x is already a screen
+coordinate, so there is no data position to move it off, and the y bounds the segment reports
+are untouched by the snap.
+
+`a_dense_trace_is_drawn_as_whole_pixel_columns` is the guard, and what it measures is the
+**number of distinct coverage levels** in the ink: 4 with the reduction against 126 with the
+polyline, over one sine period across 60 000 samples. A count of *partially covered pixels* is
+**not** the reading and was tried first — a 1.5 px stroke with a 1 px anti-aliasing fringe
+always spills into its neighbour column, so the two renderings measure 577 and 652, which
+separates nothing. The frame's whole geometry collapses with it: 627 segments against 60 054.
+
+## A reduced column can be graded by dwell
+
+`Style::intensity_grading` draws each of those columns **graded** rather than as one flat band:
+bright where the signal sits, faint where it only passes through, the way a digital-phosphor
+scope reads. It is off by default — a change of look, not a correction of one — and
+`FrameStats::graded_traces` counts the traces that took it, a subset of `column_traces`, since
+grading is a way of drawing the reduction and not a third path.
+
+What it puts back is the information the reduction throws away: a column's min and max are the
+**same** whether its samples fill the span evenly or pile at its two ends, so a ramp and a pulse
+are one band. The distribution cannot be recovered from the band, which is why the gather is
+asked at a level **finer than the pixel grid** (`intensity_sub_columns` sub-buckets per column)
+and the dwell read from those sub-buckets.
+
+Five things carry it, and each was arrived at by being wrong first:
+
+- **The measure is dwell per row, each segment depositing one unit spread over the rows it
+  spans** (`1 / height`). Counting +1 per row instead grades a sine's fast middle as brightly as
+  its rails, which is the pattern inverted — the middle is where the signal spends the *least*
+  time.
+- **The accumulator is one column wide.** Gathered x is monotonic, so points arrive column by
+  column: accumulate, read, clear. A full `axes.w × axes.h` density image is 33 MB at 4K; one
+  column of `axes.h` floats is L1-resident. A span is written as a `+1/height` delta at its first
+  row and the negation one past its last, so it costs two writes whatever its height, and the
+  running sum over the column is the per-row dwell.
+- **Normalisation is per column, against the column's own peak.** A frame-wide scale needs the
+  whole image to exist before anything can be drawn, *and* reads worse: a column the signal flew
+  through uniformly comes out faint at every row, so the trace fades out exactly where it moves
+  fastest. The cost of the local scale is one extra non-clearing walk of the accumulator to find
+  the peak.
+- **Alpha is quantised into `intensity_steps` tiers, one `LineBatch` and one draw call each.**
+  The line pipeline's colour is a **uniform**, not per-instance (`LineInstance` is four floats of
+  geometry; only `TextInstance` carries `a_colour`), so a per-segment alpha is inexpressible and
+  the tiers are what make the gradient drawable at all. `intensity_floor` keeps the faintest tier
+  visible — a fast riser is one sub-bucket deep everywhere and must not be dimmed out of
+  existence — and `intensity_gamma` lifts the faint end. A connector between two disjoint columns
+  is emitted at the **faintest** tier: it is inferred geometry, not a place the signal was
+  measured to dwell.
+- **Grading may not undo the reduction.** `level_for()` answers 0 — raw samples — as soon as the
+  asked-for fraction of a column falls under one bucket (`kBucket` is 64), so the finer ask is
+  **floored at level 1** whenever the flat path would have used the pyramid. Without the floor
+  every moderately dense frame became an O(visible) gather: measured in udpscope at 1.3 MS/s over
+  a 2 s window, 172 samples a pixel, the flat path sat on the pyramid and grading dropped the
+  whole figure to raw (`raw=1154 lod=0` against `raw=710 lod=453`) and cost 2.3 → 6.4 ms of draw
+  mean. The floor degrades the *look* instead — fewer sub-columns than asked for, down to what the
+  coarsest buckets give, which at 64 samples a bucket is still two or three per column. That is
+  also why raising `intensity_sub_columns` past about 8 buys nothing: the bucket size is the
+  limit, not the number.
+
+`a_dense_trace_can_be_graded_by_dwell` is the guard, and it draws the same frame both ways. It
+profiles the mean peak coverage over the band's two outer fifths against its middle fifth: the
+flat band is flat (**191 and 191**, one ink) and the graded one is **191 against 131**, a middle
+at 0.69 of the edge. That ratio is a *tier* ratio and not the dwell ratio — the default gamma
+0.45 and floor 0.3 compress it on purpose. The geometry stays bounded by the axis and not by its
+area: **8303 segments over a 572 px axis**, 14.5 runs a column, against the flat path's 627 at
+one end and `572 × 156` rows at the other. A fourth frame over a window wide enough that the
+flat path is on the pyramid is what exercises the floor (`raw=0 lod=1 … seg=634
+visible=200001`); the test's first window sits at 105 samples a pixel, where both paths are raw
+and the floor is never reached.
+
 ## One sample or a statistic of many
 
 Every number the library reports is one the samples actually hold. What differs between

include/plot/render/renderer.hpp

@@ -29,6 +29,15 @@ namespace plot {
 struct FrameStats {
   u32 raw_traces = 0;        ///< traces drawn from raw samples (level 0)
   u32 lod_traces = 0;        ///< traces drawn from a min/max pyramid level
+  /// Traces whose geometry was reduced onto the pixel grid: one vertical segment
+  /// per column, lit from that column's own extremes. Counts a trace from either
+  /// path above -- the reduction is about the *density* of the gathered points,
+  /// not about which level produced them.
+  u32 column_traces = 0;
+  /// Traces whose columns were graded by dwell rather than drawn as one flat band
+  /// (`Style::intensity_grading`). A subset of `column_traces`: grading is a way of
+  /// drawing the reduction, not a third path.
+  u32 graded_traces = 0;
   u64 visible_samples = 0;   ///< samples the gathered windows covered, summed
   u32 staged_segments = 0;   ///< line segments staged, all colour groups
   u32 staged_markers = 0;    ///< marker shapes staged over the traces
@@ -404,6 +413,13 @@ class Renderer {
   /// axis-aligned line by up to half a pixel. Set on every flush, like `dash`.
   void flush_lines(const Mapping& map, Color color, f32 width, Vec2 dash = Vec2{},
                    bool snap = false) noexcept;
+  /// The same, for a batch other than the renderer's own `lines_`. The graded trace
+  /// pass stages one batch per alpha step, and a step has to be a draw call of its
+  /// own because the line pipeline takes its colour from a uniform -- so it needs to
+  /// flush a batch it chose rather than the member one. Every other caller flushes
+  /// `lines_` and uses the overload above.
+  void flush_lines(render::LineBatch& batch, const Mapping& map, Color color, f32 width,
+                   Vec2 dash = Vec2{}, bool snap = false) noexcept;
   /// Draws every staged glyph in `batch` -- each instance carries its own tint
   /// (`a_colour`) -- and empties the batch. Text is texture-bound, so this runs
   /// once per text batch, after the line geometry it belongs over.
@@ -500,6 +516,29 @@ class Renderer {
   /// The data helpers take it too; they read the axes rect and the transforms it
   /// pins.
   void draw_traces(const Plot& plot, const PlotLayout& layout);
+  /// One gathered window reduced onto the pixel grid and staged in framebuffer
+  /// pixels: for each integer column of `layout.axes`, one exactly vertical
+  /// segment from the lowest to the highest screen y any gathered point put in
+  /// it. Returns how many columns were staged (0 stages nothing); the caller
+  /// flushes with `pixel_mapping()` and `snap = true`.
+  ///
+  /// `gather_` is the input, read with the same origin `draw_traces` asked the
+  /// series for. See `draw_traces` for when this path is taken and why.
+  u32 stage_trace_columns(const Plot& plot, const PlotLayout& layout, const Trace& trace,
+                          f64 x_origin, u32 points);
+  /// The same reduction, graded by **dwell**: a column is drawn as a stack of runs,
+  /// each at the alpha step its own share of the column's samples earns, so a flat
+  /// top is bright and the edge leading to it is faint (`Style::intensity_grading`).
+  ///
+  /// Stages into `steps_`, one `LineBatch` per alpha step, which the caller flushes
+  /// in turn -- the line pipeline's colour is a uniform, so a step is a draw call.
+  /// Returns the segments staged across all steps (0 stages nothing).
+  ///
+  /// `gather_` is the input and must have been asked for at the *finer* level
+  /// `Style::intensity_sub_columns` implies; a column's dwell is not recoverable
+  /// from its own min and max (see `Style::intensity_grading`).
+  u32 stage_trace_density(const Plot& plot, const PlotLayout& layout, const Trace& trace,
+                          f64 x_origin, u32 points);
   /// The cursors, both orientations: one stroke per cursor, placed at the coordinate its own
   /// orientation owns (`Plot::cursor_screen_coordinate()`) and spanning the other axis' whole
   /// range, in its own colour -- or in the accent at handle width while it is hovered or
@@ -611,7 +650,25 @@ class Renderer {
   render::TextBatch glyphs_under_;
   render::TextBatch glyphs_over_;
   std::vector<f32> gather_;   ///< scratch for one gathered window, re-used
-  std::vector<f32> staged_;   ///< scratch for one staged window, re-used
+  /// One (lowest y, highest y) pair per pixel column of the axes, re-used:
+  /// `stage_trace_columns()`'s working set. Two floats per column, so an 8K-wide
+  /// axis is 64 kB -- bounded by the *screen*, where `gather_` is bounded by the
+  /// window.
+  std::vector<f32> columns_;
+  /// `stage_trace_density()`'s working set, and deliberately **one column wide**: the
+  /// gathered points arrive in x order, so a column can be accumulated, read and
+  /// cleared before the next one starts. A full-axes density image would be
+  /// `axes.w * axes.h` (33 MB at 4K); this is `axes.h` (a few kB) and stays in L1.
+  ///
+  /// Each sub-bucket's span is written as a +1 at its first row and a -1 one past
+  /// its last, so accumulating a span costs two writes whatever its height; the
+  /// running sum over the column is then how many sub-buckets cover each row, which
+  /// is the dwell.
+  std::vector<f32> column_rows_;
+  /// One batch per alpha step of the grading (`Style::intensity_steps`), re-used
+  /// across frames: the line pipeline takes its colour from a uniform, so segments
+  /// at different alphas cannot share a draw call and must be staged apart.
+  std::vector<render::LineBatch> steps_;
   mutable std::vector<u8> readback_;  ///< dump_frame_ppm's row buffer
 
   /// The host's `GL_COLOR_CLEAR_VALUE`, queried on the first `begin_frame()` (GL

include/plot/render/style.hpp

@@ -177,6 +177,82 @@ struct Style {
                                    ///< sides draw through the one instanced
                                    ///< anti-aliased quad path.
 
+  // --- intensity grading: counts and ratios, never sizes -------------------
+
+  /// Draw a dense trace's pixel columns **graded by how long the signal dwells at
+  /// each height**, the way a digital-phosphor scope does, instead of as one flat
+  /// band per column: bright where samples pile up (a flat top, a peak), faint
+  /// where they only pass through (an edge, a noise excursion).
+  ///
+  /// Off by default, because it is a change of *look* and not a correction: the
+  /// flat band is the honest answer to "what did this column reach", and a host
+  /// that wants that -- or that draws few enough samples never to take the column
+  /// path at all -- should not have its traces restyled by an upgrade. A scope-like
+  /// host turns it on.
+  ///
+  /// It is not derivable from the flat band. A column's min and max are the same
+  /// whether its samples fill the span or sit piled at its two ends, which is
+  /// exactly the difference between a ramp and a pulse -- so grading needs the
+  /// gather asked at a *finer* level than the pixel grid (`intensity_sub_columns`)
+  /// and reads the distribution from those sub-buckets.
+  ///
+  /// The measure is *dwell per row*, and the weighting is the whole of it: each
+  /// sub-column segment deposits **one unit spread over the rows it spans**, so a
+  /// tall crossing leaves a thin smear and a short one at a turning point leaves a
+  /// bright mark. Counting a row at +1 per segment that touches it instead -- the
+  /// obvious thing -- grades a sine's fast middle as brightly as its rails, which
+  /// is the pattern inverted.
+  bool intensity_grading = false;
+
+  /// Sub-buckets per pixel column the gather is asked for when grading: the level
+  /// is chosen for `samples_per_pixel / intensity_sub_columns`, so a column's
+  /// density is read from this many envelopes instead of from one.
+  ///
+  /// It buys resolution in *dwell*, linearly, for the same factor in gathered
+  /// points -- and the points stay bounded by the screen, which is the property
+  /// that matters (8 of them over a 1000 px axis is 16 000 points whatever the
+  /// window holds). Below about 4 a pulse's flat levels and its edge are too few
+  /// samples apart to separate.
+  ///
+  /// It is a *want*, not a promise: the renderer floors the level it asks for at the
+  /// pyramid's coarsest bucket level, so a window too sparse to hold this many buckets
+  /// a column is graded from however many it does hold. Without that floor `level_for`
+  /// answers 0 -- raw samples -- the moment a sub-column falls under one bucket, which
+  /// turns a frame the flat path draws off the pyramid into an O(visible) gather. So
+  /// raising this past about 8 costs resolution nothing and buys it nothing either: the
+  /// bucket size is the limit, not this number.
+  u32 intensity_sub_columns = 8;
+
+  /// Quantisation of the grading, in alpha steps. Each step is one draw call (the
+  /// line pipeline's colour is a uniform, so a per-segment alpha would mean a
+  /// per-instance one), and the segments are bounded by the columns, so the cost
+  /// of a step is a draw call over a few hundred instances.
+  ///
+  /// Four is enough to read a waveform's structure and few enough that a step is a
+  /// visible tier rather than a gradient -- the point is to separate "dwelt here"
+  /// from "passed through", not to measure.
+  u32 intensity_steps = 4;
+
+  /// Alpha the faintest graded step is drawn at, as a fraction of the trace's own.
+  ///
+  /// Not 0: a column the signal merely passed through is still a column the signal
+  /// *reached*, and a riser dimmed out of existence is a square wave drawn as two
+  /// disconnected lines. This is the floor that keeps the trace continuous.
+  f32 intensity_floor = 0.3f;
+
+  /// Exponent the normalised dwell is raised to before it is quantised. Below 1 it
+  /// lifts the faint end, which is what makes a fast edge visible beside a flat top
+  /// whose dwell is three orders of magnitude greater.
+  ///
+  /// The dwell is normalised **against the column's own peak**, not against a scale
+  /// for the whole frame, so the steps spend their range on the structure *inside*
+  /// each column. One scale for the frame needs the whole `w * h` density image
+  /// (33 MB at 4K, which is why the accumulator is one column wide) and reads
+  /// worse: a column the signal flew straight through holds little dwell at any
+  /// height, so it would grade uniformly faint -- a trace that fades out exactly
+  /// where it moves fastest.
+  f32 intensity_gamma = 0.45f;
+
   /// `logical` in device pixels: the single conversion point for every logical
   /// metric above, so the arithmetic appears once per consumer instead of a `*
   /// ui_scale` scattered over the renderer and the layout.

src/render/renderer.cpp

@@ -616,9 +616,14 @@ void Renderer::upload_text_vertices(std::span<const f32> interleaved) noexcept {
 
 void Renderer::flush_lines(const Mapping& map, Color color, f32 width, Vec2 dash,
                            bool snap) noexcept {
-  const u32 segments = lines_.segment_count();
+  flush_lines(lines_, map, color, width, dash, snap);
+}
+
+void Renderer::flush_lines(render::LineBatch& batch, const Mapping& map, Color color, f32 width,
+                           Vec2 dash, bool snap) noexcept {
+  const u32 segments = batch.segment_count();
   if (segments != 0 && ready_) {
-    upload_vertices(lines_.data());
+    upload_vertices(batch.data());
     line_program_.bind();
     line_program_.set_scale(map.scale);
     line_program_.set_offset(map.offset);
@@ -643,7 +648,7 @@ void Renderer::flush_lines(const Mapping& map, Color color, f32 width, Vec2 dash
     stats_.staged_segments += segments;
     ++stats_.draw_calls;
   }
-  lines_.clear();  // keeps the capacity, so the next group re-uses the store
+  batch.clear();  // keeps the capacity, so the next group re-uses the store
 }
 
 void Renderer::flush_text(render::TextBatch& batch) noexcept {
@@ -972,6 +977,304 @@ void Renderer::draw_titles(const Plot& plot, const PlotLayout& layout) {
 
 // ------------------------------------------------------------------- data
 
+u32 Renderer::stage_trace_columns(const Plot& plot, const PlotLayout& layout, const Trace& trace,
+                                  f64 x_origin, u32 points) {
+  const Transform xt = plot.x_transform(trace.axis_x);
+  const Transform yt = plot.y_transform(trace.axis_y);
+  const f32 left = layout.axes.x;
+  // One column per whole pixel of the axes, plus the partial one at the right-hand
+  // edge: `axes.w` is fractional, and the samples that land in that last fraction
+  // are inside the frame.
+  const u32 count = static_cast<u32>(layout.axes.w) + 1u;
+  constexpr f32 kInf = std::numeric_limits<f32>::infinity();
+  columns_.resize(2u * static_cast<usize>(count));
+  for (u32 c = 0; c < count; ++c) {
+    columns_[2u * c] = kInf;         // the lowest screen y seen in this column
+    columns_[2u * c + 1u] = -kInf;   // and the highest; `lo <= hi` is "present"
+  }
+
+  for (u32 k = 0; k < points; ++k) {
+    const f64 x_abs = x_origin + static_cast<f64>(gather_[2u * k]);
+    const f64 sx = xt.to_screen_x(x_abs);
+    const f64 sy = yt.to_screen_y(static_cast<f64>(gather_[2u * k + 1u]));
+    // A NaN sample, and a non-positive one on a log axis, have no position: both
+    // arrive here as a non-finite screen coordinate.
+    if (!std::isfinite(sx) || !std::isfinite(sy)) continue;
+    const f64 column = std::floor(sx - static_cast<f64>(left));
+    // Clamped, not dropped. The gathered window carries one sample of slack on each
+    // side, and that sample's ink belongs to the edge column -- it is what the
+    // polyline drew there, under the same clip. Nothing further out can arrive: the
+    // window is the visible range's.
+    const u32 c = !(column > 0.0)                           ? 0u
+                  : column >= static_cast<f64>(count - 1u)   ? count - 1u
+                                                            : static_cast<u32>(column);
+    const f32 y = static_cast<f32>(sy);
+    if (y < columns_[2u * c]) columns_[2u * c] = y;
+    if (y > columns_[2u * c + 1u]) columns_[2u * c + 1u] = y;
+  }
+
+  const f32 width = style_.device(trace.width);
+  u32 staged = 0;
+  f32 prev_lo = 0.0f;
+  f32 prev_hi = 0.0f;
+  bool joined = false;   // the column before this one is present, so there is a join to make
+  for (u32 c = 0; c < count; ++c) {
+    f32 lo = columns_[2u * c];
+    f32 hi = columns_[2u * c + 1u];
+    if (!(lo <= hi)) {
+      // No sample landed here. A hole in the data stays a hole: bridging it would
+      // draw a vertical line over x the series says nothing about, where the
+      // polyline drew a diagonal across the gap at most one pixel wide.
+      joined = false;
+      continue;
+    }
+    // Connectivity. Two columns' extremes need not overlap -- a fast edge rises
+    // through a whole column -- and two disjoint vertical segments read as a broken
+    // trace, which is exactly what the polyline's connecting segment prevented. The
+    // later column is extended to meet the earlier one's nearest bound, so the join
+    // is made inside the column that moved and no ink appears left of it.
+    if (joined) {
+      if (hi < prev_lo) {
+        hi = prev_lo;
+      } else if (lo > prev_hi) {
+        lo = prev_hi;
+      }
+    }
+    prev_lo = lo;
+    prev_hi = hi;
+    joined = true;
+    // A span shorter than the stroke is drawn *as* the stroke: the fragment shader
+    // ramps on the distance to the segment's axis and does not cap its ends, so a
+    // vertical segment shorter than its own width rasterises a sliver of itself --
+    // and a column of a flat trace, where `lo == hi`, nothing at all.
+    if (hi - lo < width) {
+      const f32 mid = 0.5f * (lo + hi);
+      lo = mid - 0.5f * width;
+      hi = mid + 0.5f * width;
+    }
+    // The column's centre, in framebuffer pixels. Every segment therefore shares one
+    // sub-pixel phase, which is the whole point (see `draw_traces`), and the flush's
+    // snap puts that one phase on the grid.
+    const f32 x = left + static_cast<f32>(c) + 0.5f;
+    lines_.add_segment({x, lo}, {x, hi});
+    ++staged;
+  }
+  return staged;
+}
+
+u32 Renderer::stage_trace_density(const Plot& plot, const PlotLayout& layout, const Trace& trace,
+                                  f64 x_origin, u32 points) {
+  if (points < 2u) return 0u;
+  // At least two steps, or there is nothing to grade and the faintest step is also
+  // the brightest -- a one-step grading is the flat band drawn dimmer, which is not
+  // what any host asking for this wants.
+  const u32 steps = std::max(2u, style_.intensity_steps);
+  // Exactly `steps`, never merely at least: the flush loop reads `steps_.size()` as the
+  // step count and divides by it, so a batch left over from a larger setting would be
+  // flushed at the wrong alpha -- or, being uncleared, at all.
+  steps_.resize(steps);
+  for (u32 t = 0; t < steps; ++t) steps_[t].clear();
+
+  const Transform xt = plot.x_transform(trace.axis_x);
+  const Transform yt = plot.y_transform(trace.axis_y);
+  const f32 left = layout.axes.x;
+  const f32 top = layout.axes.y;
+  const u32 columns = static_cast<u32>(layout.axes.w) + 1u;
+  // One accumulator row per whole pixel of the axes, plus the partial one at the
+  // bottom edge, plus one past the end for a span's closing delta.
+  const u32 rows = static_cast<u32>(layout.axes.h) + 1u;
+  column_rows_.assign(static_cast<usize>(rows) + 1u, 0.0f);
+
+  const f32 width = style_.device(trace.width);
+  const f32 gamma = style_.intensity_gamma > 0.0f ? style_.intensity_gamma : 1.0f;
+  const f32 bottom = top + layout.axes.h;
+
+  u32 staged = 0;
+  u32 column = 0;        // the column being accumulated
+  bool open = false;     // it has at least one segment
+  u32 prev_column = 0;   // the last column that emitted anything, for the join below
+  f32 prev_lo = 0.0f;
+  f32 prev_hi = 0.0f;
+  bool joined = false;
+
+  // Reads one column's accumulator out as graded runs and clears it, so the next column
+  // starts empty. Two walks of it, which is nothing -- it is one column wide and in L1 --
+  // and the first is what makes the second's grading *relative to this column*:
+  //
+  // the brightest row of a column is always the top tier. That is deliberate and it is
+  // the difference between grading and dimming. The alternative, one scale for the whole
+  // frame, needs a pass over every column before any of them can be drawn (an
+  // `axes.w * axes.h` image, 33 MB at 4K) and makes a column the signal flew through
+  // *uniformly* faint -- a trace that fades out where it moves fastest, which is where a
+  // reader is most often looking. Per column, the dwell structure *inside* each one is
+  // what the tiers spend their range on: the rails of a sine are bright and its fast
+  // middle is faint, in every column.
+  //
+  // The runs are maximal stretches of equal tier, so this is O(rows) segments per column
+  // at worst and a handful in practice: a column of a clean waveform is a bright ridge
+  // and a faint tail.
+  const auto emit_column = [&]() {
+    const f32 x = left + static_cast<f32>(column) + 0.5f;
+    f32 peak = 0.0f;
+    {
+      f32 cover = 0.0f;
+      for (u32 r = 0; r < rows; ++r) {
+        cover += column_rows_[r];
+        if (cover > peak) peak = cover;
+      }
+    }
+    const f32 inv = peak > 0.0f ? 1.0f / peak : 0.0f;
+    // What counts as covered at all. The deposits are fractions, so this cannot be an
+    // integer test: it is a share of the column's own peak, small enough to keep a row a
+    // single thin segment grazed and large enough not to light a row on rounding alone.
+    const f32 lit_floor = 0.01f * peak;
+    f32 cover = 0.0f;
+    bool in_run = false;
+    u32 run_tier = 0;
+    u32 run_start = 0;
+    f32 band_lo = 0.0f;
+    f32 band_hi = 0.0f;
+    bool any = false;
+    // One past the last row too: it closes a run that reaches the bottom edge.
+    for (u32 r = 0; r <= rows; ++r) {
+      cover += column_rows_[r];
+      column_rows_[r] = 0.0f;
+      const bool lit = r < rows && cover > lit_floor;
+      u32 tier = 0;
+      if (lit) {
+        const f32 share = std::min(1.0f, cover * inv);
+        const u32 t = static_cast<u32>(std::pow(share, gamma) * static_cast<f32>(steps));
+        tier = std::min(steps - 1u, t);
+      }
+      if (in_run && (!lit || tier != run_tier)) {
+        f32 lo = top + static_cast<f32>(run_start);
+        f32 hi = top + static_cast<f32>(r);
+        // A run shorter than the stroke is drawn *as* the stroke, for the reason
+        // `stage_trace_columns()` gives: the fragment shader ramps on the distance to
+        // the segment's axis and does not cap its ends.
+        if (hi - lo < width) {
+          const f32 mid = 0.5f * (lo + hi);
+          lo = mid - 0.5f * width;
+          hi = mid + 0.5f * width;
+        }
+        steps_[run_tier].add_segment({x, lo}, {x, hi});
+        ++staged;
+        in_run = false;
+      }
+      if (lit && !in_run) {
+        run_start = r;
+        run_tier = tier;
+        in_run = true;
+      }
+      if (lit) {
+        if (!any) {
+          band_lo = top + static_cast<f32>(r);
+          any = true;
+        }
+        band_hi = top + static_cast<f32>(r) + 1.0f;
+      }
+    }
+    open = false;
+    if (!any) {
+      joined = false;   // nothing visible here: a hole, like an uncovered column
+      return;
+    }
+    // Connectivity, the same rule as the flat path: two columns' bands need not
+    // overlap -- a fast edge rises through a whole column -- and two disjoint
+    // verticals read as a broken trace. The gap is bridged by a segment of its own at
+    // the **faintest** tier rather than by extending a run, because a connector is
+    // inferred geometry and not a place the signal was measured to dwell: grading it
+    // bright would claim the opposite of what it is.
+    if (joined && column == prev_column + 1u) {
+      if (band_hi < prev_lo) {
+        steps_[0].add_segment({x, band_hi}, {x, prev_lo});
+        ++staged;
+      } else if (band_lo > prev_hi) {
+        steps_[0].add_segment({x, prev_hi}, {x, band_lo});
+        ++staged;
+      }
+    }
+    prev_lo = band_lo;
+    prev_hi = band_hi;
+    prev_column = column;
+    joined = true;
+  };
+
+  f64 px = 0.0;
+  f64 py = 0.0;
+  bool have_prev = false;
+  for (u32 k = 0; k < points; ++k) {
+    const f64 x_abs = x_origin + static_cast<f64>(gather_[2u * k]);
+    const f64 sx = xt.to_screen_x(x_abs);
+    const f64 sy = yt.to_screen_y(static_cast<f64>(gather_[2u * k + 1u]));
+    // A NaN sample, and a non-positive one on a log axis, have no position -- and it
+    // breaks the chain as well as skipping the point: a segment from the sample
+    // before it to the one after would cross x the series says nothing about.
+    if (!std::isfinite(sx) || !std::isfinite(sy)) {
+      have_prev = false;
+      continue;
+    }
+    if (have_prev) {
+      // The *segment*, not the point, is what carries dwell: it is a stretch of the
+      // waveform, and where it went slowly it covers few rows and where it went fast it
+      // covers many. This is where a height-based heuristic cannot reach -- a column's
+      // min and max cannot tell a ramp from a pulse, and a column's segments can (see
+      // `Style::intensity_grading`).
+      //
+      // It is charged to the column its earlier end falls in. The gather is asked at
+      // `intensity_sub_columns` buckets per column, so a segment is a fraction of a
+      // column wide and splitting one across a boundary would buy a fraction of one
+      // sub-bucket's dwell.
+      const f64 c = std::floor(px - static_cast<f64>(left));
+      // Clamped, not dropped, for the reason `stage_trace_columns()` gives: the
+      // gathered window carries one sample of slack on each side and that sample's
+      // ink belongs to the edge column.
+      const u32 ci = !(c > 0.0)                                 ? 0u
+                     : c >= static_cast<f64>(columns - 1u)      ? columns - 1u
+                                                               : static_cast<u32>(c);
+      if (open && ci != column) emit_column();   // x is increasing, so ci > column
+      if (!open) {
+        column = ci;
+        open = true;
+      }
+      const f64 lo = std::min(py, sy);
+      const f64 hi = std::max(py, sy);
+      // Wholly outside the axes is wholly clipped, and a clamp would move its ink to
+      // the edge and invent a ridge there. A segment that *crosses* the edge is
+      // clamped instead, which is the coverage the clip leaves -- so a column's shares
+      // are shares of what is visible in it, which is what is being graded.
+      if (hi >= static_cast<f64>(top) && lo <= static_cast<f64>(bottom)) {
+        const f64 ra = std::floor(lo - static_cast<f64>(top));
+        const f64 rb = std::floor(hi - static_cast<f64>(top));
+        const u32 r0 = !(ra > 0.0) ? 0u
+                       : ra >= static_cast<f64>(rows - 1u) ? rows - 1u
+                                                          : static_cast<u32>(ra);
+        const u32 r1 = !(rb > 0.0) ? 0u
+                       : rb >= static_cast<f64>(rows - 1u) ? rows - 1u
+                                                          : static_cast<u32>(rb);
+        // A segment deposits one unit of dwell, **spread over the rows it covers** --
+        // which is the whole measurement, and counting the rows instead was wrong in a
+        // way worth recording: a sub-bucket holds about as many samples as any other, so
+        // a segment crossing a hundred rows is those samples *passing through* each of
+        // them, while one covering two rows is the same samples *sitting* in them. Giving
+        // each crossed row a full unit makes a sine's fast middle as bright as its rails,
+        // which is the pattern grading exists to show.
+        //
+        // Two writes whatever the height: the running sum over the column is then the
+        // dwell at each row.
+        const f32 w = 1.0f / static_cast<f32>(r1 - r0 + 1u);
+        column_rows_[r0] += w;
+        column_rows_[r1 + 1u] -= w;
+      }
+    }
+    px = sx;
+    py = sy;
+    have_prev = true;
+  }
+  if (open) emit_column();
+  return staged;
+}
+
 void Renderer::draw_traces(const Plot& plot, const PlotLayout& layout) {
   // Both dimensions: a zero-width rect selects no level, and a zero-height one
   // would flatten every trace onto a single row -- a frame that looks like a
@@ -1013,9 +1316,30 @@ void Renderer::draw_traces(const Plot& plot, const PlotLayout& layout) {
     // This is the whole performance contract: the window, never the series.
     const u64 visible = last - first;
     const f64 samples_per_pixel = static_cast<f64>(visible) / static_cast<f64>(layout.axes.w);
-    const u32 level = samples_per_pixel >= static_cast<f64>(style_.dense_threshold_spp)
+    // Grading reads a column's *distribution*, which one envelope per column cannot
+    // carry (`Style::intensity_grading`), so the gather is asked for the level a
+    // fraction of a column implies. The *gate* is still the pixel grid's density: a
+    // window sparse enough to be drawn as a polyline is drawn as one, graded or not.
+    const bool grade = style_.intensity_grading && trace.mode == data::GatherMode::Envelope &&
+                       style_.intensity_sub_columns > 1u && style_.intensity_steps > 1u;
+    const u32 flat_level = samples_per_pixel >= static_cast<f64>(style_.dense_threshold_spp)
                                ? series.pyramid().level_for(n, samples_per_pixel)
                                : 0u;
+    // The finer level, floored at the pyramid's coarsest bucket level. Grading asks for a
+    // fraction of a column and `level_for` answers 0 -- raw samples -- as soon as that
+    // fraction falls under one bucket, so an unfloored ask turns every moderately dense
+    // frame into an O(visible) gather: measured at 1.3 MS/s over a 2 s window, 172 samples
+    // a pixel, where the flat path sits on the pyramid and grading dropped the whole figure
+    // to raw (`raw=1154 lod=0` against `raw=710 lod=453`) and cost 2.3 -> 6.4 ms a frame.
+    // The floor degrades the *look* instead: fewer sub-columns than asked for, down to
+    // whatever the coarsest buckets give, which at 64 samples a bucket is still two or three
+    // per column. Grading is a way of drawing a reduction and may not undo it.
+    const u32 level =
+        grade && flat_level > 0u
+            ? std::max(1u, series.pyramid().level_for(
+                               n, samples_per_pixel /
+                                      static_cast<f64>(style_.intensity_sub_columns)))
+            : flat_level;
 
     gather_.clear();
     // The visible range's minimum is the gather origin: the emitted f32 is then O(the span on
@@ -1111,9 +1435,73 @@ void Renderer::draw_traces(const Plot& plot, const PlotLayout& layout) {
     // trace/axis hierarchy on a high-DPI panel. `width == 0` is scatter: the
     // markers below are the whole trace.
     if (trace.width > 0.0f) {
+      // A window denser than two gathered points per pixel column is drawn as *columns*
+      // rather than as a polyline: one exactly vertical, pixel-snapped segment per column,
+      // spanning what that column's own points reach.
+      //
+      // The polyline's geometry at this density is a stack of near-vertical hairlines at
+      // *fractional* x -- an envelope bucket spans (spp/2, spp] samples, so there are one to
+      // two buckets per column, each emitting its own min->max hairline plus a connector to
+      // the next, about three and a half per column. Each is composited with its own
+      // sub-pixel coverage, so a column's brightness is a non-linear function of those
+      // phases; the phases repeat at the beat between the bucket pitch and the pixel pitch,
+      // which is a **moire** across the trace, and panning sweeps the beat so the pattern
+      // crawls. It is strongest exactly where a scope is used hardest: a high sample rate, a
+      // high signal frequency, or a noisy signal, all of which make the ribbon thin enough
+      // for its per-column coverage to be visible as texture.
+      //
+      // Reducing to one segment per column removes the cause rather than the symptom: there
+      // is one phase on the whole trace (and the snap puts it on the grid), so there is
+      // nothing left to beat. It is also O(axes.w) geometry instead of O(buckets), and it
+      // loses nothing a reader could have seen -- a column cannot show more than the extremes
+      // of what it covers, which is what the band is.
+      //
+      // Two conditions, and each refuses a case this would damage.
+      // - `Envelope` only. `Nearest` and `Mean` are one-point-per-bucket modes chosen so the
+      //   trace reads as *samples* or as a smoothed line; a min/max band would answer a
+      //   different question from the one the host asked.
+      // - Two points per column. Below that the gather is sparser than the grid, and the
+      //   polyline's segments are the shape: `Series::gather()` also falls back to raw samples
+      //   when no complete bucket fits the range, and that window's trace is drawn by the two
+      //   segments that enter and leave the frame (see
+      //   `a_sparse_window_between_samples_still_inks_both_margins`). The real envelope path
+      //   emits two to four points per column by construction, so the test holds exactly when
+      //   it ran.
+      const bool columns = trace.mode == data::GatherMode::Envelope &&
+                           static_cast<f64>(points) >= 2.0 * static_cast<f64>(layout.axes.w);
+      // `snap = true` on geometry that is not chrome, which is the one carve-out in that
+      // uniform's rule ("a line that reports a data position may not be moved off it"): a
+      // column *is* a pixel column, and its x is already a screen coordinate derived from
+      // one -- there is no data position to move it off. The y bounds, which are what the
+      // segment reports, are untouched by the snap. It holds for both staging paths.
+      const u32 staged =
+          !columns ? 0u
+          : grade  ? stage_trace_density(plot, layout, trace, x_range.min, points)
+                   : stage_trace_columns(plot, layout, trace, x_range.min, points);
+      if (staged != 0u && grade) {
+        // One draw call per alpha step, faintest first: the line pipeline's colour is a
+        // uniform, so a per-segment alpha is not expressible and a step is a batch. They
+        // are flushed in order so the bright ridge composites over the faint tail rather
+        // than under it -- the runs within a column do not overlap, but their AA fringes
+        // do, and that is the order a phosphor display puts them in.
+        for (u32 t = 0; t < steps_.size(); ++t) {
+          if (steps_[t].segment_count() == 0u) continue;
+          Color color = trace.color;
+          color.a *= style_.intensity_floor +
+                     (1.0f - style_.intensity_floor) * static_cast<f32>(t) /
+                         static_cast<f32>(steps_.size() - 1u);
+          flush_lines(steps_[t], pixel_mapping(), color, style_.device(trace.width), dash, true);
+        }
+        ++stats_.column_traces;
+        ++stats_.graded_traces;
+      } else if (staged != 0u) {
+        flush_lines(pixel_mapping(), trace.color, style_.device(trace.width), dash, true);
+        ++stats_.column_traces;
+      } else {
         stage_xy(gather_.data(), points, x_range.min, log_x, log_y);
         flush_lines(map, trace.color, style_.device(trace.width), dash);
       }
+    }
 
     // Markers, staged as line segments through this same batch in *pixel* space
     // (so their size is device pixels whatever the data scale is). The stride is

tests/test_render_smoke.cpp

@@ -38,6 +38,23 @@ constexpr plot::u32 kH = 360;
 /// 1M-wide window holds ~16 of them (the dense envelope spans the full height in
 /// every column) while a 640-wide one is a nearly flat line (the sparse path's
 /// polyline crosses every column exactly once).
+/// A sine of **one period per pixel column** for the window
+/// `a_dense_trace_can_be_graded_by_dwell` draws (60 000 x-units over a 572 px axis, so
+/// 104.9 samples a column). That is the one density at which a column *has* a dwell
+/// structure to grade: its envelope spans the full amplitude, so the band is tall, and
+/// the sub-buckets inside it are a fraction of a period each, so the ones at the sine's
+/// turning points are short and the ones crossing its middle are long. A period of many
+/// columns leaves every band a thin ribbon with nothing to distinguish, and a period of a
+/// fraction of a column fills every sub-bucket to the full amplitude -- also nothing.
+std::vector<plot::f32> fast_sine_samples() {
+  std::vector<plot::f32> y(1'000'000);
+  const double w = 2.0 * 3.14159265358979323846 / 104.9;
+  for (plot::u64 i = 0; i < y.size(); ++i) {
+    y[i] = static_cast<plot::f32>(std::sin(static_cast<double>(i) * w));
+  }
+  return y;
+}
+
 std::vector<plot::f32> sine_samples() {
   std::vector<plot::f32> y(1'000'000);
   for (plot::u64 i = 0; i < y.size(); ++i) {
@@ -868,6 +885,260 @@ PLOT_TEST(dense_and_sparse_paths_agree_on_extent) {
   ctx.destroy();
 }
 
+/// A window denser than the pixel grid inks **whole columns**, which is what kills the moire.
+///
+/// Reported as "strong moire patterns due to a large number of vertical lines" on a high rate,
+/// a high frequency or a noisy signal. The polyline at this density is a stack of near-vertical
+/// hairlines at *fractional* x -- one to two envelope buckets per column, each a min->max
+/// hairline plus a connector -- so a column's brightness is a non-linear function of those
+/// sub-pixel phases, and the phases repeat at the beat between the bucket pitch and the pixel
+/// pitch. That beat is the pattern, and panning sweeps it, so it crawls.
+///
+/// The observable is the beat's own cause, measured: how many **distinct coverage levels** the
+/// trace's ink takes. One segment per column puts every column's ink at one sub-pixel phase, so
+/// the AA ramp is evaluated at the same few distances everywhere and the ink lands on **4**
+/// levels; the polyline's hairlines sit at arbitrary phases and its ink spreads over **126** of
+/// them. That spread *is* the texture, and it is what beats against the pixel pitch.
+///
+/// A count of partially covered pixels is *not* the reading, which is worth knowing before
+/// adding one: both renderings put about the same amount of ink in the fringe (577 against 652
+/// here, of 1400-odd inked pixels), because a 1.5 px stroke with a 1 px fringe always spills
+/// into its neighbour. The fix is not that the fringe is gone, it is that it is the *same*
+/// fringe in every column.
+///
+/// The window is one period of the sine over 107 samples per pixel: dense enough for the
+/// reduction and slow enough that the band is a *thin ribbon*, which is where the beat shows.
+/// A window whose buckets each span several periods paints a solid block and hides it.
+PLOT_TEST(a_dense_trace_is_drawn_as_whole_pixel_columns) {
+  plottest::HeadlessContext ctx(kW, kH);
+  plot::Status st;
+  PLOT_CHECK(ctx.create(&st));
+  PLOT_CHECK(plot::gl::load(nullptr));
+  plot::Renderer r;
+  PLOT_CHECK(r.init(&st));
+
+  plot::data::ArraySeries s(sine_samples(), 0.0, 1.0);
+  plot::Plot p;
+  p.set_viewport({0, 0, static_cast<plot::f32>(kW), static_cast<plot::f32>(kH)});
+  const plot::u32 ax = p.add_x_axis({"t", {0.0, 60'000.0}, false, plot::AxisPlace::Bottom, false});
+  const plot::u32 ay = p.add_y_axis({"v", {-1.0, 1.0}, false, plot::AxisPlace::Left, false});
+  p.add_trace({&s, plot::Color::hex(0x40C0FF), 1.5f, ax, ay});
+
+  r.begin_frame(kW, kH, {0, 0, static_cast<plot::f32>(kW), static_cast<plot::f32>(kH)});
+  r.draw(p);
+  r.end_frame();
+  dump_if_requested(r, "columns.ppm");
+
+  const plot::Rect a = p.layout().axes;
+  // Blue over the background minus red is the trace's coverage, linearly: 8 at no coverage
+  // and 199 at full (0x40C0FF over 0x101418). Nothing else in this frame is blue-dominant --
+  // the grid and the axis line are grey -- so the two bands below are the trace's own ink,
+  // split at three quarters covered.
+  plot::u32 inked = 0;
+  plot::u32 histogram[256] = {};
+  for (plot::u32 row = 0; row < static_cast<plot::u32>(a.h); ++row) {
+    for (plot::u32 col = 0; col < static_cast<plot::u32>(a.w); ++col) {
+      plot::u8 px[4] = {};
+      const plot::u32 x = static_cast<plot::u32>(a.x) + col;
+      PLOT_CHECK(ctx.read_pixel(x, ctx.height - 1u - (static_cast<plot::u32>(a.y) + row), px));
+      const int cover = static_cast<int>(px[2]) - static_cast<int>(px[0]);
+      if (cover > 40) {   // the trace's ink, at a quarter covered or better
+        ++inked;
+        ++histogram[cover];
+      }
+    }
+  }
+  // How many *distinct* coverages the ink takes is the beat, measured. A column's ink is
+  // placed at one sub-pixel phase, so the AA ramp is evaluated at the same few distances in
+  // every column and the ink lands on a handful of levels; a polyline's hairlines sit at
+  // arbitrary phases, so its ink spreads over the whole range of them. The floor of 4 pixels
+  // per level ignores the stragglers at the ribbon's turning points.
+  plot::u32 levels = 0;
+  for (plot::u32 c = 0; c < 256u; ++c) {
+    if (histogram[c] >= 4u) ++levels;
+  }
+  std::printf("[columns] axes=[%.1f %.1f %.1f %.1f] raw=%u lod=%u columns=%u segments=%u "
+              "visible=%llu inked=%u levels=%u\n",
+              a.x, a.y, a.w, a.h, r.stats().raw_traces, r.stats().lod_traces,
+              r.stats().column_traces, r.stats().staged_segments,
+              static_cast<unsigned long long>(r.stats().visible_samples), inked, levels);
+
+  // The reduction ran, and on the *raw* path: the gate is the gathered density, not the level.
+  PLOT_CHECK(r.stats().column_traces == 1);
+  PLOT_CHECK(r.stats().visible_samples > 50'000u);
+  // One segment per column, so the frame's whole geometry -- grid and axis lines included --
+  // is bounded by the axis' width: 627 against the polyline's 60 054.
+  PLOT_CHECK(r.stats().staged_segments < static_cast<plot::u32>(a.w) + 64u);
+  // The ribbon is there, and its ink sits on a handful of coverage levels (4) rather than on a
+  // continuum of them (126).
+  PLOT_CHECK(inked > static_cast<plot::u32>(a.w));
+  PLOT_CHECK(levels <= 8u);
+
+  PLOT_CHECK(!r.last_frame_had_gl_error());
+  ctx.destroy();
+}
+
+/// `Style::intensity_grading`: a reduced column is drawn **graded by dwell** -- bright
+/// where the signal sits, faint where it only passes through -- which is the second half
+/// of the moire answer. The columns alone removed the beat; this is what puts back the
+/// information a flat band throws away, the way a phosphor display does.
+///
+/// The observable is the thing the feature claims and nothing weaker: within one column,
+/// the rows at the **edges** of the band are brighter than the rows in its **middle**.
+/// For a sine of one period per column that is exactly right -- it dwells at its turning
+/// points and crosses its middle at full speed -- and it is a property of the *rendering*
+/// rather than of the waveform, because the same frame drawn ungraded puts the band's
+/// edge and its middle at the same coverage. Both frames are measured here, so what is
+/// asserted is the difference between them.
+///
+/// It also pins what grading may *not* cost. The reduction still runs (`column_traces`),
+/// so the beat is still gone; the geometry is still bounded by the axis' width times a
+/// small factor, because a run is a stretch of equal tier and not a row; and the draw
+/// calls are bounded by the steps, since the line pipeline's colour is a uniform.
+PLOT_TEST(a_dense_trace_can_be_graded_by_dwell) {
+  plottest::HeadlessContext ctx(kW, kH);
+  plot::Status st;
+  PLOT_CHECK(ctx.create(&st));
+  PLOT_CHECK(plot::gl::load(nullptr));
+  plot::Renderer r;
+  PLOT_CHECK(r.init(&st));
+
+  plot::data::ArraySeries s(fast_sine_samples(), 0.0, 1.0);
+  plot::Plot p;
+  p.set_viewport({0, 0, static_cast<plot::f32>(kW), static_cast<plot::f32>(kH)});
+  const plot::u32 ax = p.add_x_axis({"t", {0.0, 60'000.0}, false, plot::AxisPlace::Bottom, false});
+  const plot::u32 ay = p.add_y_axis({"v", {-1.0, 1.0}, false, plot::AxisPlace::Left, false});
+  p.add_trace({&s, plot::Color::hex(0x40C0FF), 1.5f, ax, ay});
+
+  // The band's edges against its middle, averaged over columns across the axis. The
+  // shares are of the band's own height, so the reading does not depend on where the
+  // band happens to sit or on how tall it is: the outer fifth at each end against the
+  // middle fifth.
+  struct Profile {
+    plot::u32 edge = 0;     ///< mean peak coverage over the band's two outer fifths
+    plot::u32 middle = 0;   ///< mean peak coverage over its middle fifth
+    plot::u32 columns = 0;  ///< columns that held a band at all
+  };
+  auto profile = [&](const plot::Rect& a) -> Profile {
+    const plot::u32 rows = static_cast<plot::u32>(a.h);
+    std::vector<int> cover(rows, 0);
+    plot::u64 edge_sum = 0;
+    plot::u64 middle_sum = 0;
+    Profile out;
+    for (plot::u32 col = 8; col + 8 < static_cast<plot::u32>(a.w); col += 16) {
+      plot::u32 lo = rows;
+      plot::u32 hi = 0;
+      for (plot::u32 row = 0; row < rows; ++row) {
+        plot::u8 px[4] = {};
+        const plot::u32 x = static_cast<plot::u32>(a.x) + col;
+        PLOT_CHECK(ctx.read_pixel(x, ctx.height - 1u - (static_cast<plot::u32>(a.y) + row), px));
+        cover[row] = static_cast<int>(px[2]) - static_cast<int>(px[0]);
+        if (cover[row] > 20) {   // the faintest step is 0.3 of 191, so this is under it
+          if (row < lo) lo = row;
+          hi = row;
+        }
+      }
+      // A band of fewer than 15 rows has no middle fifth to speak of; at this density
+      // there is none, but the reading would be meaningless if there were.
+      if (lo >= hi || hi - lo < 15u) continue;
+      const plot::u32 h = hi - lo + 1u;
+      const plot::u32 fifth = h / 5u;
+      int edge = 0;
+      int middle = 0;
+      for (plot::u32 row = lo; row <= hi; ++row) {
+        const plot::u32 d = row - lo;
+        if (d < fifth || d >= h - fifth) {
+          if (cover[row] > edge) edge = cover[row];
+        } else if (d >= 2u * fifth && d < 3u * fifth) {
+          if (cover[row] > middle) middle = cover[row];
+        }
+      }
+      edge_sum += static_cast<plot::u64>(edge);
+      middle_sum += static_cast<plot::u64>(middle);
+      ++out.columns;
+    }
+    if (out.columns != 0) {
+      out.edge = static_cast<plot::u32>(edge_sum / out.columns);
+      out.middle = static_cast<plot::u32>(middle_sum / out.columns);
+    }
+    return out;
+  };
+
+  // Ungraded first: the reference the grading is read against.
+  r.begin_frame(kW, kH, {0, 0, static_cast<plot::f32>(kW), static_cast<plot::f32>(kH)});
+  r.draw(p);
+  r.end_frame();
+  dump_if_requested(r, "ungraded.ppm");
+  const plot::Rect a = p.layout().axes;
+  const Profile flat = profile(a);
+  const plot::u32 flat_segments = r.stats().staged_segments;
+  const plot::u32 flat_draws = r.stats().draw_calls;
+  PLOT_CHECK(r.stats().column_traces == 1 && r.stats().graded_traces == 0);
+  PLOT_CHECK(!r.last_frame_had_gl_error());
+
+  r.style().intensity_grading = true;
+  r.begin_frame(kW, kH, {0, 0, static_cast<plot::f32>(kW), static_cast<plot::f32>(kH)});
+  r.draw(p);
+  r.end_frame();
+  dump_if_requested(r, "graded.ppm");
+  const Profile lit = profile(a);
+
+  std::printf("[graded] axes=[%.1f %.1f %.1f %.1f] flat edge=%u middle=%u cols=%u seg=%u "
+              "draws=%u | graded edge=%u middle=%u cols=%u seg=%u draws=%u columns=%u "
+              "graded_traces=%u\n",
+              a.x, a.y, a.w, a.h, flat.edge, flat.middle, flat.columns, flat_segments,
+              flat_draws, lit.edge, lit.middle, lit.columns, r.stats().staged_segments,
+              r.stats().draw_calls, r.stats().column_traces, r.stats().graded_traces);
+
+  // Both frames measured a band in most of the columns they sampled, or the two
+  // profiles are not comparable and everything below is vacuous.
+  PLOT_CHECK(flat.columns > 20u && lit.columns == flat.columns);
+  // The flat band is flat: its edge and its middle are one ink (191 and 191).
+  PLOT_CHECK(flat.middle > 0u && flat.edge * 100u < flat.middle * 115u);
+  // The graded one is not, and in the direction a sine dwells in: 191 against 131, a
+  // middle at 0.69 of the edge. The figure is a *tier* ratio and not the dwell ratio --
+  // the per-row dwell at the rails is several times the crossing's, and
+  // `intensity_gamma` (0.45) and `intensity_floor` (0.3) are there to compress exactly
+  // that into a range an eye reads as shading rather than as a gap. So the threshold is
+  // what the quantised ink does, with room for a step either way.
+  PLOT_CHECK(lit.middle > 0u && lit.middle * 100u < lit.edge * 80u);
+  // The reduction still ran, so the beat the columns removed is still gone.
+  PLOT_CHECK(r.stats().column_traces == 1 && r.stats().graded_traces == 1);
+  // A run is a stretch of equal tier, not a row: the geometry stays bounded by the
+  // axis' width times a small factor, nothing like `w * h`. Measured 8303 over a 572 px
+  // axis, 14.5 runs a column -- the dwell profile's own structure (four tiers from the
+  // dim middle out to each rail) plus the quantiser's dither at each boundary. The flat
+  // path's 627 is the other end of the scale, and 572 * 156 rows is what this is not.
+  PLOT_CHECK(r.stats().staged_segments < 24u * static_cast<plot::u32>(a.w));
+  // And the draw calls by the steps: one per tier that staged anything, over the flat
+  // frame's single trace draw.
+  PLOT_CHECK(r.stats().draw_calls <= flat_draws + r.style().intensity_steps);
+
+  // And the level floor, which the window above cannot see: at 105 samples a pixel the
+  // flat path is already on raw samples, so grading asks for nothing finer. A wider
+  // window puts the flat path on the pyramid, and there a sub-column is `spp / 8` -- which
+  // falls under one bucket long before the column does, so an unfloored ask answers level
+  // 0 and gathers the whole visible window. That is a real cost and it was measured: at
+  // 1.3 MS/s over a 2 s window udpscope's own frame report went from `raw=710 lod=453` to
+  // `raw=1154 lod=0` and 2.3 -> 6.4 ms a frame. The floor is one bucket level, so the
+  // reading is simply that a graded frame stays on the pyramid.
+  p.set_x_range(ax, {0.0, 200'000.0});
+  r.begin_frame(kW, kH, {0, 0, static_cast<plot::f32>(kW), static_cast<plot::f32>(kH)});
+  r.draw(p);
+  r.end_frame();
+  std::printf("[graded] wide: raw=%u lod=%u columns=%u graded=%u seg=%u visible=%llu\n",
+              r.stats().raw_traces, r.stats().lod_traces, r.stats().column_traces,
+              r.stats().graded_traces, r.stats().staged_segments,
+              static_cast<unsigned long long>(r.stats().visible_samples));
+  PLOT_CHECK(r.stats().visible_samples > 100'000u);
+  PLOT_CHECK(r.stats().lod_traces == 1 && r.stats().raw_traces == 0);
+  PLOT_CHECK(r.stats().column_traces == 1 && r.stats().graded_traces == 1);
+  PLOT_CHECK(r.stats().staged_segments < 24u * static_cast<plot::u32>(a.w));
+
+  PLOT_CHECK(!r.last_frame_had_gl_error());
+  ctx.destroy();
+}
+
 // Resolution #1 of the brief: the renderer draws into whatever framebuffer is
 // bound and clears only its viewport, so a host that embeds the plot in a
 // sub-rect (Qt/GTK/ImGui) keeps the rest of its surface. Coordinates are given