perf(app): lighter stress demo + profileable release build

- StressTestScreen: use MinMaxLTTB LOD (caps output <= pixelWidth,
  ~half the vertices vs MinMax's ~2x) and move the sample generator
  loop to Dispatchers.Default so it stops contending with the UI thread.
- SignalGenerator: run the shared producer loop off the main thread.
- Add a release buildType (R8 + resource shrink, debuggable=false,
  debug-signed for on-device profiling); proguard-rules.pro needs no
  keeps (no kotlinx.serialization; Compose/coroutines ship consumer rules).
- minSdk 26 -> 24.

Off-main producer un-starves the feed (280 -> ~1300 push/s); release
build (AOT+R8) vs debug at matched load: CPU -12%, fps +48%.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

app/build.gradle.kts

@@ -10,7 +10,7 @@ android {
 
     defaultConfig {
         applicationId = "dev.dtrentin.chart.demo"
-        minSdk = 26
+        minSdk = 24
         targetSdk = 35
         versionCode = 1
         versionName = "0.1.0"
@@ -23,6 +23,20 @@ android {
 
     buildFeatures { compose = true }
 
+    buildTypes {
+        release {
+            isMinifyEnabled = true          // R8 — the perf lever
+            isShrinkResources = true        // strip unused resources
+            isDebuggable = false            // let ART fully optimize (perf test)
+            // debug key so the release apk installs on a dev device (profiling build, not store build)
+            signingConfig = signingConfigs.getByName("debug")
+            proguardFiles(
+                getDefaultProguardFile("proguard-android-optimize.txt"),
+                "proguard-rules.pro"
+            )
+        }
+    }
+
     sourceSets["main"].kotlin.srcDirs("src/main/kotlin")
 }
 

app/proguard-rules.pro

@@ -0,0 +1,12 @@
+# Release R8 rules for :app (profiling build).
+#
+# Intentionally minimal. Consumer rules ship with the libraries:
+#   - Jetpack Compose (androidx.compose.*) — bundled consumer R8 rules
+#   - kotlinx-coroutines — bundled consumer R8 rules (keeps ServiceLoader/volatile)
+#   - AGP auto-keeps the manifest entry point (.MainActivity)
+#
+# No kotlinx.serialization in this app (chart-realtime does not apply the
+# serialization plugin), so no @Serializer keep rules are required.
+#
+# App-specific keeps are added below ONLY when a build/run failure proves them
+# necessary. Do NOT blanket-keep.

app/src/main/kotlin/dev/dtrentin/chart/demo/SignalGenerator.kt

@@ -1,9 +1,11 @@
 package dev.dtrentin.chart.demo
 
 import dev.dtrentin.chart.RealtimeChartState
+import kotlinx.coroutines.CoroutineScope
+import kotlinx.coroutines.Dispatchers
 import kotlinx.coroutines.delay
 import kotlinx.coroutines.isActive
-import kotlinx.coroutines.CoroutineScope
+import kotlinx.coroutines.withContext
 
 /**
  * Synthetic multi-signal producer. Single coroutine pushes one sample per active signal
@@ -34,9 +36,13 @@ suspend fun CoroutineScope.runSignalGenerator(
 
     val tickIntervalMs = (1_000L / sampleRateHz).coerceAtLeast(1L)
     val dt = 1f / sampleRateHz.toFloat()
+
+    // Push work off Main: sample generation + state.push run on Dispatchers.Default so they
+    // don't contend with Compose draw on the Main dispatcher. withContext inherits the caller's
+    // Job → cancelling the caller (e.g. leaving the composable) cancels this loop (no leak).
+    withContext(Dispatchers.Default) {
         val startMs = System.currentTimeMillis()
         var t = 0f
-
         while (isActive) {
             val ts = startMs + (t * 1000f).toLong()
             for (i in signalNames.indices) {
@@ -48,4 +54,5 @@ suspend fun CoroutineScope.runSignalGenerator(
             t += dt
             delay(tickIntervalMs)
         }
+    }
 }

app/src/main/kotlin/dev/dtrentin/chart/demo/StressTestScreen.kt

@@ -26,7 +26,7 @@ import dev.dtrentin.chart.RealtimeChart
 import dev.dtrentin.chart.RealtimeChartState
 import dev.dtrentin.chart.interaction.InteractionConfig
 import dev.dtrentin.chart.interaction.rememberChartInteractionState
-import dev.dtrentin.chart.lod.MinMaxLodStrategy
+import dev.dtrentin.chart.lod.MinMaxLttbLodStrategy
 import dev.dtrentin.chart.model.AxisConfig
 import dev.dtrentin.chart.model.AxisLabelMode
 import dev.dtrentin.chart.model.ChartConfig
@@ -39,8 +39,10 @@ import dev.dtrentin.chart.render.DecimalAxisFormatter
 import dev.dtrentin.chart.render.TimeAxisFormatter
 import kotlin.math.PI
 import kotlin.math.sin
+import kotlinx.coroutines.Dispatchers
 import kotlinx.coroutines.delay
 import kotlinx.coroutines.isActive
+import kotlinx.coroutines.withContext
 
 private const val STRESS_SIGNAL_COUNT = 8
 private const val STRESS_SAMPLE_HZ = 200
@@ -58,8 +60,9 @@ private val STRESS_COLORS = listOf(
 
 /**
  * Stress test: 8 simultaneous signals @ 200 Hz each (1600 pushes/sec total). Uses
- * MinMax LoD (preserves peaks under heavy throughput). 10 s window — 16k samples
+ * MinMaxLTTB LoD (preserves min/max envelope, output capped at pixelWidth). 10 s window —
- * visible at once.
+ * 16k samples visible at once. Producer runs on [Dispatchers.Default] to keep sample
+ * generation off the Main dispatcher (away from Compose draw).
  */
 @Composable
 fun StressTestScreen() {
@@ -81,7 +84,7 @@ fun StressTestScreen() {
                 ),
                 render = RenderConfig(
                     theme = chartTheme,
-                    lodStrategy = MinMaxLodStrategy(),
+                    lodStrategy = MinMaxLttbLodStrategy(),
                 ),
             ),
         )
@@ -96,6 +99,10 @@ fun StressTestScreen() {
                 SignalConfig(color = STRESS_COLORS[i], strokeWidth = 1.5f),
             )
         }
+        // Push work off Main: generation + state.push run on Dispatchers.Default so they
+        // don't contend with Compose draw on the Main dispatcher. withContext inherits the
+        // effect's Job → leaving the composable cancels this loop (no leak).
+        withContext(Dispatchers.Default) {
             val tickIntervalMs = (1_000L / STRESS_SAMPLE_HZ).coerceAtLeast(1L)
             val startMs = System.currentTimeMillis()
             var n = 0L
@@ -115,6 +122,7 @@ fun StressTestScreen() {
                 delay(tickIntervalMs)
             }
         }
+    }
 
     var pushPerSec by remember { mutableIntStateOf(0) }
     var totalPushes by remember { mutableLongStateOf(0L) }

chart-realtime/build.gradle.kts

@@ -64,7 +64,7 @@ android {
     compileSdk = 35
 
     defaultConfig {
-        minSdk = 26
+        minSdk = 24
     }
 
     compileOptions {

chart-realtime/src/commonMain/kotlin/dev/dtrentin/chart/RealtimeChart.kt

@@ -5,7 +5,9 @@ import androidx.compose.foundation.background
 import androidx.compose.foundation.gestures.detectDragGestures
 import androidx.compose.foundation.gestures.detectTapGestures
 import androidx.compose.foundation.gestures.detectTransformGestures
+import androidx.compose.foundation.layout.Box
 import androidx.compose.runtime.Composable
+import androidx.compose.runtime.mutableFloatStateOf
 import androidx.compose.runtime.remember
 import androidx.compose.ui.Modifier
 import androidx.compose.ui.geometry.Offset
@@ -27,11 +29,16 @@ import dev.dtrentin.chart.model.ChartTheme
 import dev.dtrentin.chart.render.AxisRenderer.drawXAxis
 import dev.dtrentin.chart.render.AxisRenderer.drawYAxis
 import dev.dtrentin.chart.render.AxisRenderer.resolveYRange
+import dev.dtrentin.chart.render.AxisRenderer.stabilizeYRange
 
 /**
- * Renders all signals held by [state] on a Canvas. Recomposition is driven by Compose
+ * Renders all signals held by [state] into a [Box] of two stacked Canvas layers: a COLD
- * snapshot observation of `state.dataVersion`, so the Canvas redraws only when new data
+ * layer (Y-axis line / grid / labels, redrawn only when the stabilized Y range or size
- * arrives (batched per frame by the Compose snapshot system).
+ * changes) and a HOT layer above it (signal polylines, X-axis / grid / time-labels,
+ * crosshair). The hot layer reads `state.dataVersion` INSIDE its draw lambda → draw-phase
+ * invalidation (no recomposition), coalesced to one draw per frame by the Compose snapshot
+ * system; it publishes a tick-stabilized Y range (see [stabilizeYRange]) to the cold layer
+ * so per-frame Y-label layout stops running every frame.
  *
  * v0.5.0 wiring:
  * - Decimation strategy from `state.config.render.lodStrategy` (default MinMaxLTTB).
@@ -42,7 +49,7 @@ import dev.dtrentin.chart.render.AxisRenderer.resolveYRange
  *   chart behaves identically to v0.4.0 (read-only).
  *
  * @param state holds all signal data and config.
- * @param modifier applied to Canvas.
+ * @param modifier applied to the container [Box].
  * @param xWindowSeconds visible X window in seconds; overrides `state.config.data.xWindowSeconds` at call site.
  * @param theme visual theme; overrides `state.config.render.theme` at call site.
  * @param interaction optional state holder enabling user gestures. Create via
@@ -75,15 +82,32 @@ public fun RealtimeChart(
     // T9: zero-alloc Y-range out param. Layout: [0] = yMin, [1] = yMax.
     val yRangeOut = remember { FloatArray(2) }
 
+    // ── Cold-layer Y-range state (T-split) ────────────────────────────────────
+    // Stabilized (tick-quantized) Y range published to the COLD canvas. Written ONLY by
+    // the hot draw lambda, ONLY when the quantized bounds actually change. The cold canvas
+    // reads these two states → its Y-label TextMeasurer layout + Y-axis/grid draw run only
+    // on that change (or a resize), NOT every frame. Init NaN → cold skips drawing until
+    // the first hot frame publishes a real range.
+    val stableYMin = remember { mutableFloatStateOf(Float.NaN) }
+    val stableYMax = remember { mutableFloatStateOf(Float.NaN) }
+    // Plain (non-snapshot) mirror of the last-published stabilized range. Lets the hot draw
+    // detect a change WITHOUT reading the Compose state — so writing it invalidates the COLD
+    // draw only, never the hot draw itself (no self-invalidation). Layout: [0]=yMin, [1]=yMax.
+    val lastStable = remember { floatArrayOf(Float.NaN, Float.NaN) }
+    // Zero-alloc scratch for stabilizeYRange output. [0] = yMin, [1] = yMax.
+    val stableRangeOut = remember { FloatArray(2) }
+
     // Cross-frame caches read by pointer-input lambdas. Plain LongArray slots (NOT Compose
     // state) — writes inside draw must NOT invalidate composition. Pointer-input lambdas
     // read the most recently-rendered values (1-frame lag is acceptable for gestures).
     // Layout: [0] = latestMs, [1] = windowStartMs, [2] = windowMs.
     val interactionCache = remember { longArrayOf(Long.MIN_VALUE, 0L, 0L) }
 
-    val baseModifier = modifier.background(theme.backgroundColor)
+    // Gestures + background live on the container Box (was the single Canvas). The two
+    // stacked child canvases (cold below, hot above) fill it via matchParentSize().
+    val containerModifier = modifier.background(theme.backgroundColor)
     val gestureModifier = if (interaction != null) {
-        baseModifier
+        containerModifier
             .pointerInput(interaction) {
                 detectTransformGestures { _, _, zoom, _ ->
                     if (zoom != 1f) interaction.applyZoom(zoom, fallbackXWindowSeconds = xWindowSeconds)
@@ -135,9 +159,32 @@ public fun RealtimeChart(
                     )
                 }
             }
-    } else baseModifier
+    } else containerModifier
 
-    Canvas(modifier = gestureModifier) {
+    Box(modifier = gestureModifier) {
+        // ── COLD layer (drawn first → below) ───────────────────────────────────
+        // Y-axis line + Y grid + Y labels. Reads ONLY the stabilized Y range + size +
+        // insets — never dataVersion. Re-executes (re-measuring Y labels via TextMeasurer)
+        // only when the stabilized Y range or the canvas size changes. Eliminates the
+        // per-frame Y-label layout that previously ran inside the single hot draw pass.
+        Canvas(modifier = Modifier.matchParentSize()) {
+            val yMin = stableYMin.floatValue
+            val yMax = stableYMax.floatValue
+            if (yMin.isNaN() || yMax.isNaN() || yMax <= yMin) return@Canvas
+            val chartBottom = size.height - chartBottomInsetPx
+            drawYAxis(
+                yMin, yMax, theme, textMeasurer,
+                config.axis.yLabelMode, config.axis.yLabelDecimals,
+                chartLeftPx, chartBottom, showGrid = config.axis.showGrid,
+            )
+        }
+
+        // ── HOT layer (drawn second → above) ───────────────────────────────────
+        // Signal polyline + X-axis/grid/time-labels + crosshair. Reads state.dataVersion
+        // → draw-phase invalidation (composition NOT invalidated), coalesced per frame.
+        // X labels stay hot on purpose: in Following mode windowStartMs advances every
+        // frame so X ticks/labels scroll — caching them would freeze the scroll.
+        Canvas(modifier = Modifier.matchParentSize()) {
             val currentVersion = state.dataVersion
             // Recompose may run when interaction state (crosshair / mode) changes even if
             // dataVersion did not — so still draw when interaction is non-null and crosshair
@@ -185,7 +232,9 @@ public fun RealtimeChart(
             for (i in signalsArr.indices) {
                 val entry = signalsArr[i]
                 if (!entry.config.visible) { entry.scratchCount = 0; continue }
-            val n = entry.buffer.snapshot(windowStartMs, windowMs, entry.scratchTs, entry.scratchV)
+                // O(log n) bisect + in-window walk (T2). Does NOT copy the full tier ring per
+                // frame — see TieredBuffer.snapshotWindow / CircularBuffer.copyWindow.
+                val n = entry.buffer.snapshotWindow(windowStartMs, windowMs, entry.scratchTs, entry.scratchV)
                 entry.scratchCount = n
                 for (j in 0 until n) {
                     val v = entry.scratchV[j]
@@ -198,11 +247,26 @@ public fun RealtimeChart(
             }
             if (!hasData) { dataMin = -1f; dataMax = 1f }
             resolveYRange(config, dataMin, dataMax, yRangeOut)
-        val yMin = yRangeOut[0]
+            // Stabilize (quantize to the axis-tick grid). SHARED by the signal projection
-        val yMax = yRangeOut[1]
+            // below AND the cold Y grid/labels, so gridlines and signal stay pixel-aligned.
+            // The hot signal uses THIS frame's freshly-computed stable range (so data never
+            // clips); the cold canvas reads the published Compose state and therefore trails
+            // by at most one frame on the rare transition where the range crosses a tick —
+            // imperceptible (both layers then re-converge on the identical numeric range).
+            stabilizeYRange(yRangeOut[0], yRangeOut[1], stableRangeOut)
+            val yMin = stableRangeOut[0]
+            val yMax = stableRangeOut[1]
+            // Publish to the cold layer ONLY when the quantized bounds change. Compared
+            // against a plain (non-snapshot) mirror so this hot lambda never READS the
+            // Compose state → the write invalidates the COLD draw only, never itself.
+            if (stableRangeOut[0] != lastStable[0] || stableRangeOut[1] != lastStable[1]) {
+                lastStable[0] = stableRangeOut[0]
+                lastStable[1] = stableRangeOut[1]
+                stableYMin.floatValue = stableRangeOut[0]
+                stableYMax.floatValue = stableRangeOut[1]
+            }
 
             drawXAxis(windowStartMs, windowMs, theme, textMeasurer, config.axis.xLabelMode, chartLeftPx, chartBottom, t0, showGrid = config.axis.showGrid)
-        drawYAxis(yMin, yMax, theme, textMeasurer, config.axis.yLabelMode, config.axis.yLabelDecimals, chartLeftPx, chartBottom, showGrid = config.axis.showGrid)
 
             for (i in signalsArr.indices) {
                 val entry = signalsArr[i]
@@ -236,7 +300,8 @@ public fun RealtimeChart(
                 }
             }
 
-        // Crosshair overlay (drawn last → above signals + axes).
+            // Crosshair overlay (drawn last → above signals + axes). Uses the same
+            // stabilized range as the signal so dot markers land on the polyline.
             val crosshair = interaction?.crosshair
             if (crosshair != null) {
                 drawCrosshair(
@@ -254,6 +319,7 @@ public fun RealtimeChart(
 
             lastRenderedVersion[0] = currentVersion
         }
+    }
 }
 
 /**

chart-realtime/src/commonMain/kotlin/dev/dtrentin/chart/buffer/CircularBuffer.kt

@@ -43,6 +43,76 @@ internal class CircularBuffer(val capacity: Int) {
         return currentSize
     }
 
+    /**
+     * Copy only the samples whose timestamp lies in `[fromMs, toMs)` into the caller's arrays,
+     * appended starting at [outOffset], in chronological (ascending) order. Returns the count
+     * written (0 when empty / window outside data).
+     *
+     * Locates the first in-window sample via O(log n) binary search over the ring's logical
+     * order, then walks ONLY the in-window subrange — NO full-buffer copy (unlike
+     * [snapshot] + linear filter). Respects ring wrap-around via the same physical-index
+     * mapping as [snapshot].
+     *
+     * Bounds match the linear per-tier filter in `TieredBuffer.snapshot`:
+     *  - lower bound inclusive (`ts >= fromMs`)
+     *  - upper bound exclusive (`ts < toMs`)
+     *
+     * Writing stops early if the caller's arrays fill (`outIdx >= outTimestamps.size`),
+     * mirroring the bounds guard in the linear path.
+     *
+     * Thread safety: single-writer/single-reader. `writeIndex`/`size` are read once (volatile)
+     * for a consistent view — identical contract to [snapshot].
+     */
+    fun copyWindow(
+        fromMs: Long,
+        toMs: Long,
+        outTimestamps: LongArray,
+        outValues: FloatArray,
+        outOffset: Int,
+    ): Int {
+        val currentWrite = writeIndex
+        val currentSize = size.coerceAtMost(capacity)
+        if (currentSize == 0) return 0
+        val startIdx: Long = if (currentSize < capacity) 0L else currentWrite - capacity
+
+        // First logical index with ts >= fromMs. When not wrapped, logical == physical and
+        // timestamps[0, currentSize) is already chronological → reuse the contiguous
+        // bisectStart primitive. When wrapped, bisect over the ring's logical order.
+        val startLogical: Int =
+            if (currentSize < capacity) bisectStart(timestamps, currentSize, fromMs)
+            else bisectStartRing(startIdx, currentSize, fromMs)
+
+        var outIdx = outOffset
+        var i = startLogical
+        while (i < currentSize) {
+            val src = (((startIdx + i) % capacity + capacity) % capacity).toInt()
+            val ts = timestamps[src]
+            if (ts >= toMs) break
+            if (outIdx >= outTimestamps.size) break
+            outTimestamps[outIdx] = ts
+            outValues[outIdx] = values[src]
+            outIdx++
+            i++
+        }
+        return outIdx - outOffset
+    }
+
+    /**
+     * Ring-aware lower-bound bisect over logical indices `[0, count)`: first `i` where
+     * `timestamps[phys(i)] >= targetMs`, else `count`. `phys(i) = (startIdx + i) mod capacity`.
+     * Assumes the ring's logical order is non-decreasing (chronological push). O(log count).
+     */
+    private fun bisectStartRing(startIdx: Long, count: Int, targetMs: Long): Int {
+        var lo = 0
+        var hi = count
+        while (lo < hi) {
+            val mid = (lo + hi) ushr 1
+            val src = (((startIdx + mid) % capacity + capacity) % capacity).toInt()
+            if (timestamps[src] < targetMs) lo = mid + 1 else hi = mid
+        }
+        return lo
+    }
+
     fun latestTimestampMs(): Long {
         if (size == 0) return -1L
         return timestamps[((writeIndex - 1L + capacity) % capacity).toInt()]
@@ -66,8 +136,8 @@ internal class CircularBuffer(val capacity: Int) {
  *  - `count` must be `<= ts.size`
  *  - `count == 0` → returns 0
  *
- * O(log count). Zero-alloc. Used by `TieredBuffer.snapshotWindow` to locate the
+ * O(log count). Zero-alloc. Used by `CircularBuffer.copyWindow` for the non-wrapped ring
- * window-start index in each tier's snapshot instead of linear-scanning all n samples.
+ * (logical == physical order) to locate the window-start index without a linear pre-scan.
  */
 internal fun bisectStart(ts: LongArray, count: Int, targetMs: Long): Int {
     if (count <= 0) return 0

chart-realtime/src/commonMain/kotlin/dev/dtrentin/chart/buffer/TieredBuffer.kt

@@ -200,13 +200,16 @@ internal class TieredBuffer {
     }
 
     /**
-     * Bisect-based variant of [snapshot]. Returns identical content + ordering for the
+     * Window read used by the draw hot path. Returns content + ordering IDENTICAL to
-     * same (windowStartMs, windowMs) args, but locates the window-start index in each
+     * [snapshot] for the same (windowStartMs, windowMs) args, but reads each tier via
-     * tier's chronologically-sorted snapshot via O(log n) bisect instead of an
+     * [CircularBuffer.copyWindow] — an O(log n) bisect to the window start plus a walk over
-     * O(n) linear pre-scan. Linear walk runs only over the in-window subrange.
+     * ONLY the in-window subrange. Unlike [snapshot] it does NOT copy the full ring into a
+     * scratch array before filtering (which grows to 60k samples/tier0 and dominated the
+     * per-frame cost).
      *
      * Output ordering: tier2 oldest first, then tier1, then tier0 newest — same as [snapshot].
-     * Tier boundary clamps (tier0BoundaryMs / tier1BoundaryMs) preserved verbatim.
+     * Tier boundary clamps (tier0BoundaryMs / tier1BoundaryMs) and guards preserved verbatim.
+     * Edge inclusion matches [snapshot]: lower bound inclusive, upper bound exclusive.
      *
      * Returns 0 on empty buffer or window entirely outside data.
      */
@@ -226,48 +229,21 @@ internal class TieredBuffer {
         var outIdx = 0
 
         if (windowStartMs < tier1BoundaryMs) {
-            val n2 = tier2.snapshot(t2Ts, t2Vs)
             // Tier2 records satisfy ts < tier1BoundaryMs (older than tier1 horizon),
             // so upper clamp is min(windowEndMs, tier1BoundaryMs).
             val tier2UpperExclusive = if (windowEndMs < tier1BoundaryMs) windowEndMs else tier1BoundaryMs
-            val start = bisectStart(t2Ts, n2, windowStartMs)
+            outIdx += tier2.copyWindow(windowStartMs, tier2UpperExclusive, outTimestamps, outValues, outIdx)
-            var i = start
-            while (i < n2) {
-                val ts = t2Ts[i]
-                if (ts >= tier2UpperExclusive) break
-                if (outIdx >= outTimestamps.size) return outIdx
-                outTimestamps[outIdx] = ts; outValues[outIdx] = t2Vs[i]; outIdx++
-                i++
-            }
         }
 
         if (windowStartMs < tier0BoundaryMs && windowEndMs > tier1BoundaryMs) {
-            val n1 = tier1.snapshot(t1Ts, t1Vs)
             // Tier1 records satisfy tier1BoundaryMs <= ts < tier0BoundaryMs.
             val tier1Lower = if (windowStartMs > tier1BoundaryMs) windowStartMs else tier1BoundaryMs
             val tier1UpperExclusive = if (windowEndMs < tier0BoundaryMs) windowEndMs else tier0BoundaryMs
-            val start = bisectStart(t1Ts, n1, tier1Lower)
+            outIdx += tier1.copyWindow(tier1Lower, tier1UpperExclusive, outTimestamps, outValues, outIdx)
-            var i = start
-            while (i < n1) {
-                val ts = t1Ts[i]
-                if (ts >= tier1UpperExclusive) break
-                if (outIdx >= outTimestamps.size) return outIdx
-                outTimestamps[outIdx] = ts; outValues[outIdx] = t1Vs[i]; outIdx++
-                i++
-            }
         }
 
         val tier0Start = if (windowStartMs > tier0BoundaryMs) windowStartMs else tier0BoundaryMs
-        val n0 = tier0.snapshot(t0Ts, t0Vs)
+        outIdx += tier0.copyWindow(tier0Start, windowEndMs, outTimestamps, outValues, outIdx)
-        val start0 = bisectStart(t0Ts, n0, tier0Start)
-        var i = start0
-        while (i < n0) {
-            val ts = t0Ts[i]
-            if (ts >= windowEndMs) break
-            if (outIdx >= outTimestamps.size) return outIdx
-            outTimestamps[outIdx] = ts; outValues[outIdx] = t0Vs[i]; outIdx++
-            i++
-        }
 
         return outIdx
     }

chart-realtime/src/commonMain/kotlin/dev/dtrentin/chart/interaction/InverseProjection.kt

@@ -33,8 +33,8 @@ internal object InverseProjection {
 
     /**
      * Returns the value of the sample closest to [targetTsMs] in [entry]'s already-populated
-     * scratch arrays. Caller must have invoked `entry.buffer.snapshot(...)` for the current
+     * scratch arrays. Caller must have invoked `entry.buffer.snapshotWindow(...)` for the
-     * frame (i.e. `entry.scratchCount`, `entry.scratchTs`, `entry.scratchV` are populated).
+     * current frame (i.e. `entry.scratchCount`, `entry.scratchTs`, `entry.scratchV` are populated).
      *
      * Uses binary search on the chronologically-sorted `scratchTs` (snapshot output ordering
      * is documented in `TieredBuffer.snapshot`).

chart-realtime/src/commonMain/kotlin/dev/dtrentin/chart/render/AxisRenderer.kt

@@ -211,4 +211,33 @@ internal object AxisRenderer {
         }
     }
 
+    /**
+     * Quantizes an exact Y range to the axis-tick grid so downstream (cold-layer) Y-label
+     * layout runs only when the tick-aligned bounds actually change — not every frame.
+     *
+     * Expands OUTWARD to a multiple of the range's own nice step (via
+     * [NumberFormat.niceInterval]; [drawYAxis] re-derives the same — equal in-band, never
+     * finer), so:
+     *  - the stabilized range always CONTAINS the exact range → a signal projected with it
+     *    never clips, and
+     *  - signal and Y gridlines/labels share the identical `[yMin, yMax]` → never mutually
+     *    misalign (the cold layer draws off the same stabilized bounds).
+     *
+     * Under small frame-to-frame data wobble the result is IDENTICAL (bounds only move when
+     * the data crosses a tick line or the range crosses a niceInterval band), which is what
+     * keeps the cold layer cold. Non-finite / non-positive-range input passes through
+     * unchanged. Writes `out[0] = yMin`, `out[1] = yMax` (caller owns the 2-element array).
+     */
+    internal fun stabilizeYRange(exactMin: Float, exactMax: Float, out: FloatArray) {
+        val range = (exactMax - exactMin).toDouble()
+        if (!exactMin.isFinite() || !exactMax.isFinite() || range <= 0.0) {
+            out[0] = exactMin
+            out[1] = exactMax
+            return
+        }
+        val step = NumberFormat.niceInterval(range, targetTickCount = 5)
+        out[0] = (floor(exactMin / step) * step).toFloat()
+        out[1] = (ceil(exactMax / step) * step).toFloat()
+    }
+
 }

chart-realtime/src/commonMain/kotlin/dev/dtrentin/chart/render/LineSignalRenderer.kt

@@ -1,10 +1,12 @@
 package dev.dtrentin.chart.render
 
 import androidx.compose.ui.graphics.Color
+import androidx.compose.ui.graphics.Paint
+import androidx.compose.ui.graphics.PaintingStyle
 import androidx.compose.ui.graphics.Path
 import androidx.compose.ui.graphics.drawscope.DrawScope
-import androidx.compose.ui.graphics.drawscope.Stroke
 import androidx.compose.ui.graphics.drawscope.clipRect
+import androidx.compose.ui.graphics.drawscope.drawIntoCanvas
 
 /**
  * Default [SignalRenderer] impl: stroked polyline through pre-decimated points.
@@ -15,32 +17,42 @@ import androidx.compose.ui.graphics.drawscope.clipRect
  * Implementation is stateless from the caller's perspective — all scratch (Path, FloatArrays)
  * is caller-owned and passed in per call.
  *
- * v0.5.0 T8: a process-wide [Stroke] cache keyed by `strokeWidth` eliminates the per-frame
+ * Anti-aliasing is disabled on the signal stroke. `DrawScope.drawPath` forces AA on with no
- * `Stroke(width = ...)` allocation. Cache assumes single-threaded Compose UI access (the
+ * opt-out, which pins Skia to CPU coverage-mask rasterization (aaa_fill_path / blitAntiH). We
- * renderer is only ever invoked from the UI thread during Canvas draw). Capped at
+ * therefore draw through a common [Paint] (`isAntiAlias = false`) via [drawIntoCanvas], flipping
- * [STROKE_CACHE_MAX] entries (sane upper bound: typical apps use < 8 distinct widths).
+ * Skia to GPU tessellation and removing that CPU cost. Configured stroke width is preserved.
- * When the cap is hit the cache is cleared rather than running an LRU eviction (simpler;
+ *
- * a chart that uses > 16 widths is already pathological).
+ * A process-wide [Paint] cache keyed by `strokeWidth` eliminates the per-frame `Paint(...)`
+ * allocation. Cache assumes single-threaded Compose UI access (the renderer is only ever
+ * invoked from the UI thread during Canvas draw). Capped at [PAINT_CACHE_MAX] entries (sane
+ * upper bound: typical apps use < 8 distinct widths). When the cap is hit the cache is cleared
+ * rather than running an LRU eviction (simpler; a chart that uses > 16 widths is already
+ * pathological). Color varies per signal and is mutated on the cached instance per call.
  */
 public object LineSignalRenderer : SignalRenderer {
 
-    // T8: process-wide stroke cache. Keyed by Float (strokeWidth in px units, as supplied by
+    // Process-wide paint cache. Keyed by Float (strokeWidth in px units, as supplied by
     // SignalConfig.strokeWidth). Single-threaded UI access assumption — no synchronization.
-    private val strokeCache: HashMap<Float, Stroke> = HashMap(8)
+    private val paintCache: HashMap<Float, Paint> = HashMap(8)
-    private const val STROKE_CACHE_MAX: Int = 16
+    private const val PAINT_CACHE_MAX: Int = 16
 
     /**
-     * Internal test hook. Returns the same [Stroke] instance across calls with equal
+     * Internal test hook. Returns the same [Paint] instance across calls with equal
-     * [strokeWidth]. Mutates the cache (creates an entry on miss).
+     * [strokeWidth]. Mutates the cache (creates an entry on miss). The returned paint has
+     * `isAntiAlias = false`, `style = Stroke`, and `strokeWidth` set to [strokeWidth].
      */
-    internal fun internalStrokeForWidth(strokeWidth: Float): Stroke = strokeForWidth(strokeWidth)
+    internal fun internalPaintForWidth(strokeWidth: Float): Paint = paintForWidth(strokeWidth)
 
-    private fun strokeForWidth(strokeWidth: Float): Stroke {
+    private fun paintForWidth(strokeWidth: Float): Paint {
-        val cached = strokeCache[strokeWidth]
+        val cached = paintCache[strokeWidth]
         if (cached != null) return cached
-        if (strokeCache.size >= STROKE_CACHE_MAX) strokeCache.clear()
+        if (paintCache.size >= PAINT_CACHE_MAX) paintCache.clear()
-        val fresh = Stroke(width = strokeWidth)
+        val fresh = Paint().apply {
-        strokeCache[strokeWidth] = fresh
+            isAntiAlias = false
+            style = PaintingStyle.Stroke
+            this.strokeWidth = strokeWidth
+        }
+        paintCache[strokeWidth] = fresh
         return fresh
     }
 
@@ -69,9 +81,10 @@ public object LineSignalRenderer : SignalRenderer {
         for (i in 1 until count) {
             path.lineTo(chartLeft + lodX[i], chartBottom - ((lodY[i] - yMin) * invY) * chartBottom)
         }
-        val stroke = strokeForWidth(strokeWidth)
+        val paint = paintForWidth(strokeWidth)
+        paint.color = color
         clipRect(left = chartLeft, top = 0f, right = chartRight, bottom = chartBottom) {
-            drawPath(path = path, color = color, style = stroke)
+            drawIntoCanvas { it.drawPath(path, paint) }
         }
     }
 }

chart-realtime/src/commonTest/kotlin/dev/dtrentin/chart/buffer/CircularBufferTest.kt

@@ -1,5 +1,6 @@
 package dev.dtrentin.chart.buffer
 
+import kotlin.random.Random
 import kotlin.test.*
 
 class CircularBufferTest {
@@ -199,4 +200,115 @@ class CircularBufferTest {
         assertEquals(1002L, ts[cap - 3]); assertEquals(2f, v[cap - 3])
         assertEquals(1001L, ts[cap - 4]); assertEquals(1f, v[cap - 4])
     }
+
+    // ---------- T2: copyWindow bisect vs linear-filter equivalence ----------
+
+    /**
+     * Equivalence oracle: full snapshot + chronological in-window filter [fromMs, toMs)
+     * — mirrors the old linear per-tier scan that copyWindow replaces.
+     */
+    private fun linearWindow(buf: CircularBuffer, fromMs: Long, toMs: Long): Pair<LongArray, FloatArray> {
+        val cap = buf.capacity
+        val ts = LongArray(cap); val v = FloatArray(cap)
+        val n = buf.snapshot(ts, v)
+        val outTs = ArrayList<Long>(); val outV = ArrayList<Float>()
+        for (i in 0 until n) {
+            if (ts[i] >= fromMs && ts[i] < toMs) { outTs.add(ts[i]); outV.add(v[i]) }
+        }
+        return outTs.toLongArray() to outV.toFloatArray()
+    }
+
+    private fun assertCopyWindowMatchesLinear(buf: CircularBuffer, fromMs: Long, toMs: Long, msg: String) {
+        val (refTs, refV) = linearWindow(buf, fromMs, toMs)
+        val outTs = LongArray(buf.capacity); val outV = FloatArray(buf.capacity)
+        val n = buf.copyWindow(fromMs, toMs, outTs, outV, 0)
+        assertEquals(refTs.size, n, "$msg: count mismatch")
+        for (i in 0 until n) {
+            assertEquals(refTs[i], outTs[i], "$msg: ts mismatch at i=$i")
+            assertEquals(refV[i], outV[i], "$msg: v mismatch at i=$i")
+        }
+    }
+
+    @Test fun copyWindow_emptyBuffer_returnsZero() {
+        val buf = CircularBuffer(8)
+        assertEquals(0, buf.copyWindow(0L, 1000L, LongArray(8), FloatArray(8), 0))
+    }
+
+    @Test fun copyWindow_notWrapped_matchesLinear() {
+        val buf = CircularBuffer(16)
+        for (i in 0 until 10) buf.push(i * 10L, i.toFloat()) // ts 0..90
+        assertCopyWindowMatchesLinear(buf, 20L, 70L, "notWrapped[20,70)")
+        assertCopyWindowMatchesLinear(buf, 0L, 100L, "notWrapped full")
+        assertCopyWindowMatchesLinear(buf, 25L, 66L, "notWrapped off-grid")
+    }
+
+    @Test fun copyWindow_wrapped_matchesLinear() {
+        val cap = 5; val buf = CircularBuffer(cap)
+        for (i in 0..7) buf.push(i * 10L, i.toFloat()) // retained ts {30,40,50,60,70}
+        assertCopyWindowMatchesLinear(buf, 0L, 1000L, "wrapped full")
+        assertCopyWindowMatchesLinear(buf, 40L, 60L, "wrapped [40,60)")
+        assertCopyWindowMatchesLinear(buf, 45L, 65L, "wrapped off-grid")
+        assertCopyWindowMatchesLinear(buf, 0L, 30L, "wrapped before retained")
+    }
+
+    @Test fun copyWindow_windowBeforeAllData_returnsZero() {
+        val buf = CircularBuffer(8)
+        buf.push(1000L, 1f); buf.push(1010L, 2f)
+        assertEquals(0, buf.copyWindow(0L, 500L, LongArray(8), FloatArray(8), 0))
+    }
+
+    @Test fun copyWindow_windowAfterAllData_returnsZero() {
+        val buf = CircularBuffer(8)
+        buf.push(100L, 1f); buf.push(110L, 2f)
+        assertEquals(0, buf.copyWindow(1000L, 2000L, LongArray(8), FloatArray(8), 0))
+    }
+
+    @Test fun copyWindow_lowerInclusiveUpperExclusive() {
+        val buf = CircularBuffer(8)
+        for (i in 1..5) buf.push(i * 10L, i.toFloat()) // 10,20,30,40,50
+        val ts = LongArray(8); val v = FloatArray(8)
+        // [20,40): includes 20 (lower inclusive), excludes 40 (upper exclusive).
+        val n = buf.copyWindow(20L, 40L, ts, v, 0)
+        assertEquals(2, n)
+        assertEquals(20L, ts[0]); assertEquals(30L, ts[1])
+    }
+
+    @Test fun copyWindow_singlePointWindow() {
+        val buf = CircularBuffer(8)
+        for (i in 1..5) buf.push(i * 10L, i.toFloat())
+        val ts = LongArray(8); val v = FloatArray(8)
+        // [30,31) → exactly the point at 30.
+        assertEquals(1, buf.copyWindow(30L, 31L, ts, v, 0))
+        assertEquals(30L, ts[0]); assertEquals(3f, v[0])
+        // [30,30) → empty (upper exclusive == lower).
+        assertEquals(0, buf.copyWindow(30L, 30L, ts, v, 0))
+    }
+
+    @Test fun copyWindow_writesAtOffset_preservesPrefixAndReturnsCount() {
+        val buf = CircularBuffer(8)
+        for (i in 1..5) buf.push(i * 10L, i.toFloat()) // 10..50
+        val ts = LongArray(8) { -1L }; val v = FloatArray(8) { -1f }
+        val offset = 3
+        val n = buf.copyWindow(20L, 50L, ts, v, offset) // {20,30,40}
+        assertEquals(3, n)
+        // Prefix untouched.
+        for (i in 0 until offset) { assertEquals(-1L, ts[i]); assertEquals(-1f, v[i]) }
+        // Window appended at offset.
+        assertEquals(20L, ts[offset]); assertEquals(30L, ts[offset + 1]); assertEquals(40L, ts[offset + 2])
+        assertEquals(2f, v[offset]); assertEquals(3f, v[offset + 1]); assertEquals(4f, v[offset + 2])
+    }
+
+    @Test fun copyWindow_randomWindows_wrapped_matchesLinear() {
+        val cap = 64; val buf = CircularBuffer(cap)
+        // Push far more than capacity → heavy wrap. Monotonic ts.
+        val total = 500
+        for (i in 0 until total) buf.push(i * 7L, i.toFloat())
+        val maxTs = (total - 1) * 7L
+        val rnd = Random(0xC0FFEE)
+        repeat(300) {
+            val a = rnd.nextLong(-50L, maxTs + 50L)
+            val span = rnd.nextLong(0L, maxTs + 100L)
+            assertCopyWindowMatchesLinear(buf, a, a + span, "randWrapped[from=$a,span=$span]")
+        }
+    }
 }

chart-realtime/src/commonTest/kotlin/dev/dtrentin/chart/buffer/TieredBufferTest.kt

@@ -1,5 +1,6 @@
 package dev.dtrentin.chart.buffer
 
+import kotlin.random.Random
 import kotlin.test.Test
 import kotlin.test.assertEquals
 import kotlin.test.assertTrue
@@ -601,4 +602,76 @@ class TieredBufferTest {
             assertTrue(ts[i] >= ts[i - 1], "monotonic-non-decreasing ts violated at i=$i")
         }
     }
+
+    // ---------- T2: snapshotWindow (ring bisect) equivalence over a WRAPPED tier0 ----------
+
+    /**
+     * Fill tier0 well past its 60k capacity while keeping every sample inside the 5-min
+     * tier0 horizon (step 4ms → 70k samples span 280s < 300s). Forces ring wrap-around
+     * (writeIndex >> capacity) so the bisect must resolve physical wrap. All windows here
+     * land in the tier0-only region, isolating the ring-wrap path.
+     */
+    private fun filledWrappedTier0(): TieredBuffer {
+        val buf = TieredBuffer()
+        val n = TieredBuffer.TIER0_CAPACITY + 10_000 // 70_000 > capacity → wraps
+        for (i in 0 until n) buf.push(i * 4L, i.toFloat())
+        return buf
+    }
+
+    @Test fun snapshotWindow_randomWindows_overWrappedTier0_equalsSnapshot() {
+        val buf = filledWrappedTier0()
+        val maxTs = (TieredBuffer.TIER0_CAPACITY + 10_000 - 1) * 4L
+        // Oldest retained tier0 ts after wrap.
+        val minRetained = 10_000L * 4L
+        val tsA = outTs(); val vsA = outVs()
+        val tsB = outTs(); val vsB = outVs()
+        val rnd = Random(0xBADC0DE)
+        repeat(200) {
+            val start = rnd.nextLong(minRetained - 500L, maxTs + 500L)
+            val span = rnd.nextLong(1L, 50_000L)
+            val nA = buf.snapshot(start, span, tsA, vsA)
+            val nB = buf.snapshotWindow(start, span, tsB, vsB)
+            assertSnapshotEquivalent(nA, tsA, vsA, nB, tsB, vsB, "randWrapped[start=$start,span=$span]")
+        }
+    }
+
+    @Test fun snapshotWindow_windowStraddlingRingWrap_equalsSnapshot() {
+        val buf = filledWrappedTier0()
+        val maxTs = (TieredBuffer.TIER0_CAPACITY + 10_000 - 1) * 4L
+        val minRetained = 10_000L * 4L
+        // Window centred in the retained range — physical wrap point falls inside it.
+        val mid = (minRetained + maxTs) / 2L
+        val start = mid - 30_000L
+        val span = 60_000L
+        val tsA = outTs(); val vsA = outVs()
+        val tsB = outTs(); val vsB = outVs()
+        val nA = buf.snapshot(start, span, tsA, vsA)
+        val nB = buf.snapshotWindow(start, span, tsB, vsB)
+        assertSnapshotEquivalent(nA, tsA, vsA, nB, tsB, vsB, "straddleWrap")
+        assertTrue(nB > 0, "expected samples inside straddle window, got $nB")
+    }
+
+    @Test fun snapshotWindow_singlePointWindow_equalsSnapshot() {
+        val buf = TieredBuffer()
+        for (i in 1..5) buf.push(i * 10L, i.toFloat()) // ts 10..50 in tier0
+        val tsA = outTs(); val vsA = outVs()
+        val tsB = outTs(); val vsB = outVs()
+        // [30,31) → single point at ts=30.
+        val nA = buf.snapshot(30L, 1L, tsA, vsA)
+        val nB = buf.snapshotWindow(30L, 1L, tsB, vsB)
+        assertSnapshotEquivalent(nA, tsA, vsA, nB, tsB, vsB, "singlePoint")
+        assertEquals(1, nB); assertEquals(30L, tsB[0]); assertEquals(3f, vsB[0])
+    }
+
+    @Test fun snapshotWindow_pointsExactlyOnWindowBounds_equalsSnapshot() {
+        val buf = TieredBuffer()
+        for (i in 1..5) buf.push(i * 10L, i.toFloat()) // 10,20,30,40,50
+        val tsA = outTs(); val vsA = outVs()
+        val tsB = outTs(); val vsB = outVs()
+        // Bounds land exactly on samples: start=20 (inclusive), end=40 (exclusive).
+        val nA = buf.snapshot(20L, 20L, tsA, vsA) // window [20,40)
+        val nB = buf.snapshotWindow(20L, 20L, tsB, vsB)
+        assertSnapshotEquivalent(nA, tsA, vsA, nB, tsB, vsB, "onBounds")
+        assertEquals(2, nB) // {20,30}; 40 excluded, 20 included
+    }
 }

chart-realtime/src/commonTest/kotlin/dev/dtrentin/chart/render/AxisRendererTest.kt

@@ -40,4 +40,42 @@ class AxisRendererTest {
         assertTrue(out[1].isFinite())
         assertTrue(out[1] >= out[0])
     }
+
+    // T-split: stabilizeYRange snaps outward to the tick grid and CONTAINS the exact range.
+    @Test fun stabilizeYRange_snapsToTicks_andContainsExactRange() {
+        val out = FloatArray(2)
+        // range = 9.1 → niceInterval(9.1, 5) = 2 → floor(0.3/2)*2 = 0, ceil(9.4/2)*2 = 10.
+        AxisRenderer.stabilizeYRange(0.3f, 9.4f, out)
+        assertEquals(0f, out[0])
+        assertEquals(10f, out[1])
+        assertTrue(out[0] <= 0.3f && out[1] >= 9.4f)
+    }
+
+    // T-split: small frame-to-frame wobble inside a tick band → IDENTICAL bounds (cold stays cold).
+    @Test fun stabilizeYRange_stableUnderSmallWobble() {
+        val a = FloatArray(2)
+        val b = FloatArray(2)
+        AxisRenderer.stabilizeYRange(0.31f, 9.38f, a)
+        AxisRenderer.stabilizeYRange(0.34f, 9.42f, b)
+        assertEquals(a[0], b[0])
+        assertEquals(a[1], b[1])
+    }
+
+    // T-split: empty / non-positive range passes through unchanged (no NaN/Inf).
+    @Test fun stabilizeYRange_passthroughForEmptyRange() {
+        val out = FloatArray(2)
+        AxisRenderer.stabilizeYRange(5f, 5f, out)
+        assertEquals(5f, out[0])
+        assertEquals(5f, out[1])
+    }
+
+    // T-split: negative range spanning zero snaps symmetrically outward.
+    @Test fun stabilizeYRange_negativeRange_snapsOutward() {
+        val out = FloatArray(2)
+        // range = 2.4 → niceInterval(2.4, 5) = 0.5 → floor(-1.2/0.5)*0.5 = -1.5, ceil(1.2/0.5)*0.5 = 1.5.
+        AxisRenderer.stabilizeYRange(-1.2f, 1.2f, out)
+        assertEquals(-1.5f, out[0])
+        assertEquals(1.5f, out[1])
+        assertTrue(out[0] <= -1.2f && out[1] >= 1.2f)
+    }
 }

chart-realtime/src/commonTest/kotlin/dev/dtrentin/chart/render/LineSignalRendererTest.kt

@@ -1,22 +0,0 @@
-package dev.dtrentin.chart.render
-
-import kotlin.test.Test
-import kotlin.test.assertNotSame
-import kotlin.test.assertSame
-
-class LineSignalRendererTest {
-
-    // T8: Stroke cache reuses same instance across calls with equal strokeWidth.
-    @Test fun strokeCache_reusesAcrossCallsForSameWidth() {
-        val a = LineSignalRenderer.internalStrokeForWidth(2f)
-        val b = LineSignalRenderer.internalStrokeForWidth(2f)
-        assertSame(a, b)
-    }
-
-    // T8: Different widths get distinct Stroke instances.
-    @Test fun strokeCache_distinctInstancesForDifferentWidths() {
-        val a = LineSignalRenderer.internalStrokeForWidth(1.5f)
-        val b = LineSignalRenderer.internalStrokeForWidth(3f)
-        assertNotSame(a, b)
-    }
-}

chart-realtime/src/iosTest/kotlin/dev/dtrentin/chart/render/LineSignalRendererTest.kt

@@ -0,0 +1,45 @@
+package dev.dtrentin.chart.render
+
+import androidx.compose.ui.graphics.PaintingStyle
+import kotlin.test.Test
+import kotlin.test.assertEquals
+import kotlin.test.assertFalse
+import kotlin.test.assertNotSame
+import kotlin.test.assertSame
+
+/**
+ * Lives in `iosTest` (Skiko-backed) rather than `commonTest`: the renderer now builds an
+ * `androidx.compose.ui.graphics.Paint`, which on the Android JVM unit-test host delegates to
+ * the non-mockable `android.graphics.Paint` stub (`Method setAntiAlias ... not mocked`). Skiko
+ * provides a real `Paint` backend, so these assertions run for real here. The renderer itself is
+ * `commonMain` code identical on both platforms.
+ */
+class LineSignalRendererTest {
+
+    // Paint cache reuses the same instance across calls with equal strokeWidth (no per-frame alloc).
+    @Test fun paintCache_reusesAcrossCallsForSameWidth() {
+        val a = LineSignalRenderer.internalPaintForWidth(2f)
+        val b = LineSignalRenderer.internalPaintForWidth(2f)
+        assertSame(a, b)
+    }
+
+    // Different widths get distinct Paint instances.
+    @Test fun paintCache_distinctInstancesForDifferentWidths() {
+        val a = LineSignalRenderer.internalPaintForWidth(1.5f)
+        val b = LineSignalRenderer.internalPaintForWidth(3f)
+        assertNotSame(a, b)
+    }
+
+    // Anti-aliasing is off on the cached signal paint (the whole point: skip CPU coverage-mask raster).
+    @Test fun paint_antiAliasDisabled() {
+        val paint = LineSignalRenderer.internalPaintForWidth(2f)
+        assertFalse(paint.isAntiAlias)
+    }
+
+    // Configured stroke width is honored and style is Stroke.
+    @Test fun paint_honorsStrokeWidthAndStyle() {
+        val paint = LineSignalRenderer.internalPaintForWidth(4.5f)
+        assertEquals(4.5f, paint.strokeWidth)
+        assertEquals(PaintingStyle.Stroke, paint.style)
+    }
+}