docs(CLAUDE): document the parity harness + how to read the %differ

Explains running scripts/parity, the one-process JVM / per-screen native
mechanics, the build/parity outputs, and — the load-bearing part — what
the number means: it's a font-drift-inflated ranking signal, not a
pass/fail score. Ghosted text + dark shapes = normal; a solid bright
block or one-sided shape = regression. Baseline gradient ~2%..32%; known
SDL gaps (solid gradients, effects) inflate specific screens.

CLAUDE.md

@@ -458,6 +458,46 @@ build tell you what broke.
   `Dialog` / `DropdownMenu` / `DropdownMenuItem` / `TooltipBox` (m3 doesn't
   ship drop-in equivalents for our anchor / scrim patterns).
 
+## Parity harness — native-vs-JVM screenshot diff
+
+`:demo` renders the **same commonMain screens** on two stacks: native
+(SDL/Skia, Kotlin/Native) and a `jvm()` target on upstream Compose Desktop.
+`scripts/parity/parity.py` screenshots every screen on both and pixel-diffs
+them, so a screen that visually diverges is a **port regression** (missing
+content, wrong shape/colour, broken clip). Several past renderer regressions
+would have been caught here.
+
+```bash
+python scripts/parity/parity.py                 # all screens (builds first)
+python scripts/parity/parity.py Buttons Shapes  # a subset
+python scripts/parity/parity.py --no-build      # reuse the last renders
+```
+
+Mechanics: the JVM leg renders all screens headlessly via `ImageComposeScene`
+in ONE process (`:demo:run --args=--screenshot-all=<dir>`, wired in
+`MainJvm.kt`); the native leg launches the exe once per screen
+(`--screen=<Name> --screenshot=<x>.bmp`). Output lands in **`build/parity/`
+(gitignored)**: `<pct>_<Name>_diff.png` (amplified difference heatmap),
+`<pct>_<Name>_compare.png` (native ∣ jvm ∣ diff, side by side), and
+`report.txt` ranked worst-first. Windows-only for the native leg today; needs
+Pillow. See `scripts/parity/README.md`.
+
+**What the number means — read this before trusting it.** The `%differ` is the
+fraction of pixels whose per-channel difference exceeds a tolerance. It is NOT
+a pass/fail score and pixel-perfection is not the goal: the two stacks use
+different default fonts, so **every screen carries a steady baseline
+difference** — in the heatmap, text shows as a faint *doubled ghost* from
+slightly different line metrics/baselines. A healthy full sweep is a smooth
+gradient (~2% for a sparse screen like Counter, up to ~32% for a text-dense
+one like Tabs). Known SDL parity gaps also inflate specific screens
+predictably (Brushes ~23% — gradients render solid on SDL; Shadows / Canvas /
+GraphicsLayer — effect differences). **The signal is the RANKING and the
+delta from a screen's own history**, not the absolute value: a text-light
+screen suddenly reading 60%, or a screen jumping far above its neighbours, is
+the bug. In a `_diff.png`: ghosted/doubled text + dark shapes = normal font
+drift; a **solid bright block, or a shape present on only one side** = a real
+regression — open the `_compare.png` to see which stack is wrong.
+
 ## Conventions
 
 Kotlin standard style — plain `camelCase` for parameters, local variables,

ROADMAP.md

@@ -0,0 +1,91 @@
+# Renderer roadmap — fixes, performance, memory
+
+Grounded in the bubble-wrap findings (2026-07): lazy clips took the stress
+sheet from 23 → 39 FPS, but 75 Hz needs structural work; issue #2 showed that
+Cleaner-managed native memory starves when the Kotlin heap is quiet.
+
+## 1. Deterministic native-resource lifecycle
+
+Ownership closes resources the moment they're finished; GC stays as the
+safeguard, not the mechanism.
+
+- [x] **`NativeReleaseQueue`** — SDL calls are main-thread-only, so nothing may
+  destroy directly from a Cleaner or worker: release actions ENQUEUE, the main
+  loop DRAINS between frames. Cleaners become safe from any thread because
+  they only enqueue. Verified end-to-end (direct enqueue + a real
+  Cleaner→enqueue→drain cycle).
+- [x] **`SdlImageBitmap` textures** — previously never destroyed (manual SDL
+  memory, invisible to the GC nudge). A holder + Cleaner now enqueues
+  `SDL_DestroyTexture` / `SDL_DestroySurface` for the texture, render target
+  and never-realized decoded surface; explicit `close()` frees promptly.
+- [ ] **Wire `close()` into cache eviction** — the Cleaner is the current
+  release path for decoded/vector bitmaps (correct but GC-timed). Call
+  `SdlImageBitmap.close()` on image-cache eviction / `removeMemoryResource`
+  for prompt release (needs a cross-module hook — the caches hold the
+  `ImageBitmap` interface, not the concrete type).
+- [ ] **Cache-eviction closes (rest)** — every image/text cache closes what it
+  evicts (Sdl3ImageCache already does; sweep the rest, both renderers).
+- [ ] **Renderer `destroy()` chain** — window close synchronously frees every
+  pool (clip targets already do; text caches, image caches, shadow cache,
+  typefaces to follow).
+- [ ] **`SkiaImageBitmap.close()`** (surface + snapshot) + close-on-eviction in
+  the Skia caches. *Skia source sets build only on macOS/Linux — needs a mac
+  or CI leg to verify.*
+- [ ] **Demote the GC nudge** — once ownership does the work, stretch the 10s
+  interval / gate it on churn; it remains the backstop.
+
+Verification: `demo --leaktest`-style probe (cycle screens N times, print
+RSS; expect a plateau without GC sawtooth).
+
+## 2. SDL renderer performance (the 39 → 75 Hz path)
+
+- [x] **Frame profiler** — `CDN_PROFILE=1` env flag prints per-phase timings
+  (events / app pump / window pump / render / present) every ~2 s. Measure
+  first, optimize second.
+- [ ] **Right-size clip scratch targets** — the pool allocates window-sized
+  textures per mask depth; allocate at the clip bbox (size-bucketed pool) to
+  cut fill-rate for the masks that remain after lazy clips.
+- [ ] **Dirty-region rendering** — the big structural item: accumulate damage
+  from invalidated layers, scissor redraw to it. One bubble's spring should
+  not re-tessellate 84. Most likely single item to reach 75 Hz.
+- [ ] **Glyph atlas** — text draws currently break geometry batches (z-order
+  flush per run); an atlas texture lets glyphs ride the vertex batches.
+- [ ] **Retained layer textures** (later) — RenderNode-style caching of static
+  subtrees keyed by draw content.
+
+## 3. Skia renderer
+
+- [ ] Bounded `SkiaImageCache` (LRU + eviction close) — unbounded HashMap today.
+- [ ] Eager-close audit of the text pipeline (mirror of the SDL sweep).
+- [ ] `saveLayer` huge-bounds clamp (GPU offscreen memory spikes — see
+  CLAUDE.md pitfall).
+
+## 4. Correctness / parity gaps (future bug reports waiting)
+
+Ordered by likelihood of a user hitting them:
+
+- [ ] **Gradient brushes on SDL** — `Brush.linearGradient/radialGradient`
+  render solid; the per-vertex sampler infrastructure is half-present.
+- [ ] **Layer rotation** — rotationZ repositions but content doesn't rotate
+  (hit-testing too).
+- [ ] **Real `saveLayer` alpha on SDL** — overlapping content composites at
+  paint level; needs an offscreen (the clip-target pool can serve it).
+- [ ] **`clipPath` generic shapes** — bbox fallback clips square.
+
+## 5. Permanent tooling
+
+- [x] **Parity harness** — scripts/parity/: renders every :demo screen native +
+  JVM from the same commonMain composables, pixel-diffs, ranks by %differ,
+  emits per-screen diff heatmaps + side-by-side compares (pct in filename).
+  build/parity/ (gitignored). See scripts/parity/README.md.
+- [ ] Promote the press/hover automation rig (window-handle-relative input +
+  screenshot) into `scripts/`.
+
+## 6. Long-term
+
+- [ ] **SDL_GPU render backend** — real stencil clipping (no masks at all),
+  pipelined batching, shader gradients. Weeks of work across three
+  platforms; items 2–4 stay useful beneath it.
+
+Suggested order: profiler → lifecycle queue + leak closes → clip-target
+right-sizing → gradients → dirty regions → the rest by demand.

compose/sdl/window/src/nativeMain/kotlin/com/compose/sdl/ComposeWindow.kt

@@ -34,6 +34,8 @@ import kotlinx.coroutines.test.resetMain
 import kotlinx.coroutines.test.setMain
 import kotlinx.cinterop.reinterpret
 import sdl3.SDL_Delay
+import sdl3.SDL_GetPerformanceCounter
+import sdl3.SDL_GetPerformanceFrequency
 import sdl3.SDL_GetTicks
 import sdl3.SDL_Quit
 import sdl3.SDL_SetWindowTitle
@@ -148,7 +150,9 @@ fun nativeComposeApp(content: @Composable ApplicationScope.() -> Unit) {
 		//  Main loop
 		var vGcLastTicks = SDL_GetTicks()
 		var vRenderedSinceGc = false
+		val vProfiler = if (platform.posix.getenv("CDN_PROFILE") != null) FrameProfiler() else null
 		while (!runtime.exitRequested) {
+			vProfiler?.mark()
 			Snapshot.sendApplyNotifications()
 
 			// ============
@@ -176,6 +180,7 @@ fun nativeComposeApp(content: @Composable ApplicationScope.() -> Unit) {
 			}
 
 			mainDispatcher.drainPending()
+			vProfiler?.phase(0)
 
 			// ============
 			//  App composition pump — Window()s may appear / disappear here.
@@ -195,6 +200,7 @@ fun nativeComposeApp(content: @Composable ApplicationScope.() -> Unit) {
 			// Exit when the last window is gone (after at least one existed).
 			if (runtime.hadWindow && runtime.windows.isEmpty()) runtime.exitRequested = true
 			if (runtime.exitRequested) break
+			vProfiler?.phase(1)
 
 			// ============
 			//  Per-window pump + render.
@@ -208,13 +214,24 @@ fun nativeComposeApp(content: @Composable ApplicationScope.() -> Unit) {
 				Snapshot.sendApplyNotifications()
 				vW.frameClock.sendFrame()
 				yield()
+				vProfiler?.phase(2)
 				if (vW.shouldRender()) {
 					vW.renderFrame()
 					vAnyRendered = true
 					if (!vW.backend.vsyncEnabled) vAllVsync = false
 				}
+				vProfiler?.phase(3)
 			}
 			runtime.reapDestroyed()
+			vProfiler?.frameDone(vAnyRendered)
+
+			// ============
+			//  Drain deferred native-resource disposals on the MAIN thread —
+			//  textures/surfaces whose owner closed them or whose Cleaner fired
+			//  on a GC worker enqueue here (SDL calls aren't thread-safe). This
+			//  is the ownership path that makes the GC nudge below a mere
+			//  backstop (ROADMAP.md item 1).
+			com.compose.sdl.graphics.NativeReleaseQueue.drain()
 
 			// ============
 			//  Pace / idle-skip.
@@ -250,6 +267,7 @@ fun nativeComposeApp(content: @Composable ApplicationScope.() -> Unit) {
 		snapshotHandle.dispose()
 		for (vW in runtime.windows.toList()) runtime.scheduleDestroy(vW)
 		runtime.reapDestroyed()
+		com.compose.sdl.graphics.NativeReleaseQueue.drain()
 		appComposition.dispose()
 		appRecomposer.cancel()
 		appRecomposeJob.cancelAndJoin()
@@ -262,6 +280,47 @@ fun nativeComposeApp(content: @Composable ApplicationScope.() -> Unit) {
 	SDL_Quit()
 }
 
+/* CDN_PROFILE=1 — per-phase main-loop timings, printed every ~2s of rendered
+   frames. Phases: events (poll+dispatch), app (app-composition pump), pump
+   (per-window clocks/dispatch), render (layout+draw+present). Measure first,
+   optimize second — see ROADMAP.md. */
+@OptIn(kotlinx.cinterop.ExperimentalForeignApi::class)
+internal class FrameProfiler {
+	private val fFreq = SDL_GetPerformanceFrequency().toDouble()
+	private val fSum = DoubleArray(4)
+	private val fMax = DoubleArray(4)
+	private var fFrames = 0
+	private var fLastPrintMs = SDL_GetTicks()
+	private var fMark = 0uL
+
+	fun mark() { fMark = SDL_GetPerformanceCounter() }
+
+	fun phase(inIndex: Int) {
+		val vNow = SDL_GetPerformanceCounter()
+		val vMs = (vNow - fMark).toDouble() * 1000.0 / fFreq
+		fSum[inIndex] += vMs
+		if (vMs > fMax[inIndex]) fMax[inIndex] = vMs
+		fMark = vNow
+	}
+
+	fun frameDone(inRendered: Boolean) {
+		if (inRendered) fFrames++
+		val vNowMs = SDL_GetTicks()
+		if (vNowMs - fLastPrintMs >= 2000u && fFrames > 0) {
+			val vNames = listOf("events", "app", "pump", "render")
+			val vParts = vNames.mapIndexed { vI, vName ->
+				val vAvg = fSum[vI] / fFrames
+				"$vName=${(vAvg * 100).toInt() / 100.0}/${(fMax[vI] * 100).toInt() / 100.0}ms"
+			}
+			val vTotal = fSum.sum() / fFrames
+			println("[profile] frames=$fFrames avg/max " + vParts.joinToString(" ") + " total=${(vTotal * 100).toInt() / 100.0}ms")
+			for (vI in fSum.indices) { fSum[vI] = 0.0; fMax[vI] = 0.0 }
+			fFrames = 0
+			fLastPrintMs = vNowMs
+		}
+	}
+}
+
 /* Trigger a Kotlin/Native GC so Cleaner-managed renderer resources release
    their native memory (see the main loop's native-memory nudge). */
 @OptIn(kotlin.native.runtime.NativeRuntimeApi::class)

compose/ui/ui/src/nativeMain/kotlin/com/compose/sdl/graphics/NativeReleaseQueue.kt

@@ -0,0 +1,55 @@
+package com.compose.sdl.graphics
+
+import kotlinx.atomicfu.locks.SynchronizedObject
+import kotlinx.atomicfu.locks.synchronized
+
+// ==================
+// MARK: NativeReleaseQueue — main-thread-deferred native resource disposal
+// ==================
+
+/**
+ * A queue of native-resource release actions drained on the MAIN THREAD once
+ * per frame by the window loop (see ComposeWindow). It exists because the two
+ * ways a renderer resource dies are both off the main thread, while the calls
+ * that free them (SDL_DestroyTexture, and any renderer API) are main-thread-only:
+ *
+ *  - a Kotlin/Native `Cleaner` runs on a GC worker thread, and
+ *  - explicit `close()` can be reached from composition/effect code on other
+ *    dispatchers.
+ *
+ * Both simply ENQUEUE here; the main loop runs the actions between frames. This
+ * turns the GC (issue #2: Cleaner-managed native memory starving because the
+ * quiet Kotlin heap rarely triggers a collection) from the disposal MECHANISM
+ * into a mere safeguard — ownership closes resources promptly, the Cleaner is
+ * the backstop for whatever leaks a `close()`.
+ *
+ * Actions must be idempotent-safe at the source (guard against a resource being
+ * both explicitly closed and later GC-cleaned — see SdlImageBitmap's holder).
+ */
+object NativeReleaseQueue {
+
+	private val fLock = SynchronizedObject()
+	private var fPending = ArrayList<() -> Unit>()
+
+	/** Enqueue a release action. Safe to call from any thread. */
+	fun enqueue(action: () -> Unit) {
+		synchronized(fLock) { fPending.add(action) }
+	}
+
+	/**
+	 * Run and clear every queued action. MAIN THREAD ONLY — the actions call
+	 * renderer APIs that aren't thread-safe. Returns the number drained.
+	 */
+	fun drain(): Int {
+		val vBatch = synchronized(fLock) {
+			if (fPending.isEmpty()) return 0
+			val vTaken = fPending
+			fPending = ArrayList()
+			vTaken
+		}
+		for (vAction in vBatch) {
+			runCatching { vAction() }
+		}
+		return vBatch.size
+	}
+}

compose/ui/ui/src/sdlRendererMain/kotlin/com/compose/sdl/renderer/sdl/Sdl3Canvas.kt

@@ -228,15 +228,27 @@ internal class Sdl3Canvas(
 		}
 	}
 
-	/* Convert pending clips into real offscreen mask layers (outermost first). */
+	/* Convert pending clips into real offscreen mask layers (outermost first).
+
+	   Each clip realizes with its OWN push-time region (pending.region =
+	   enclosing-clip-at-push ∩ bbox), NOT the current fClip: a lazily-realized
+	   clip may fire arbitrarily deep inside descendants whose own clips have
+	   since narrowed fClip. Re-intersecting with that narrow clip shrank an
+	   outer clip's offscreen to the innermost descendant's rect, so its
+	   composite blitted back only that sliver and every sibling drawn into the
+	   same offscreen was discarded (the TLS-chain regression: only the first
+	   card + nothing below it, until a scroll shifted which card triggered
+	   realization). The real (already-narrowed) drawing clip is preserved
+	   across the loop and restored at the end, so subsequent draws into the
+	   innermost target stay correctly clipped. */
 	private fun realizePendingClips() {
 		if (fPendingClips.isEmpty()) return
 		val vRenderer = fRenderer.reinterpret<cnames.structs.SDL_Renderer>()
+		val vDrawClip = fClip
 		while (fPendingClips.isNotEmpty()) {
 			val vPending = fPendingClips.removeFirst()
 			fScope.flush()
-			// A clipRect after the push may have narrowed the visible area.
-			val vRegion = intersect(fClip, vPending.region)
+			val vRegion = vPending.region
 			val vTarget = fClipTargets?.target(fClipLayers.size, fSize.width.toInt(), fSize.height.toInt())
 			if (vTarget == null || vRegion[2] <= vRegion[0] || vRegion[3] <= vRegion[1]) {
 				// Degrade to the rect clip that is already active.
@@ -258,6 +270,10 @@ internal class Sdl3Canvas(
 				}
 			}
 		}
+		// Restore the real drawing clip (already narrowed by inner clipRects /
+		// scroll bounds) for draws into the now-active innermost target.
+		fClip = vDrawClip
+		applyClip()
 	}
 
 	// Flushes any pending batched geometry to SDL, then frees the scope's

compose/ui/ui/src/sdlRendererMain/kotlin/com/compose/sdl/renderer/sdl/Sdl3DrawScope.kt

@@ -893,7 +893,13 @@ internal class Sdl3DrawScope(
 	// through the current affine so scale/rotate reach the GPU; the colour is
 	// sampled at the pre-transform point so a gradient rides its shape.
 	private fun writeVertex(inX: Float, inY: Float, inColor: ComposeColor, inAlphaScale: Float = 1f) {
-		if (fBatchCount >= kBatchCapacity) flush()
+		// Auto-flush ONLY at a triangle boundary: SDL_RenderGeometry renders
+		// floor(count/3) triangles and silently drops the tail, so flushing
+		// mid-triangle tears it AND misaligns every following triangle in the
+		// new batch (seen as ~350-degree pac-man circles and glitch shapes
+		// once a frame's geometry outgrew one batch - e.g. 84 unclipped
+		// drawBehind bubbles). kBatchCapacity is a multiple of 3.
+		if (fBatchCount >= kBatchCapacity && fBatchCount % 3 == 0) flush()
 		val vBase = fBatchCount * kFloatsPerVertex
 		fVertexData[vBase + 0] = fMa * inX + fMc * inY + fMe
 		fVertexData[vBase + 1] = fMb * inX + fMd * inY + fMf
@@ -914,11 +920,13 @@ private const val kAaFeather: Float = 1.0f
 private const val kAaHalf: Float = kAaFeather * 0.5f
 
 // ============
-//  Batch capacity — 8192 vertices = ~2730 triangles per submission. At
-//  64 segments per full circle that's room for ~21 full-circle filled
-//  shapes per Canvas{} before any flush. Bigger gives fewer GPU
-//  submissions; smaller saves RAM. ~128 KB at 16 bytes per SDL_Vertex.
-private const val kBatchCapacity: Int = 8192
+//  Batch capacity — 8190 vertices = 2730 triangles per submission (kept a
+//  MULTIPLE OF 3: the auto-flush in writeVertex only fires on triangle
+//  boundaries, so the cap must land on one). At 64 segments per full circle
+//  that's room for ~21 full-circle filled shapes per Canvas{} before any
+//  flush. Bigger gives fewer GPU submissions; smaller saves RAM. ~128 KB at
+//  16 bytes per SDL_Vertex.
+private const val kBatchCapacity: Int = 8190
 
 // Floats per SDL_Vertex: position(x,y) + color(r,g,b,a) + tex_coord(x,y), tightly
 // packed. Used to stage vertices in a Kotlin FloatArray and memcpy them across.

compose/ui/ui/src/sdlRendererMain/kotlin/com/compose/sdl/renderer/sdl/Sdl3Offscreen.kt

@@ -29,6 +29,34 @@ internal var currentMainCanvas: Sdl3Canvas? = null
 // An ImageBitmap backed by an SDL render-target texture. The vector rasterises
 // into the texture (via a Canvas returned by the factory) and Sdl3Canvas.drawImageRect
 // blits it back — with the Icon tint applied through SDL_SetTextureColorMod.
+// Mutable native handles held BY REFERENCE so a Cleaner can free them without
+// capturing the (collectable) SdlImageBitmap. Both the explicit close() and the
+// GC Cleaner enqueue release of whatever is still here — the first to run nulls
+// the fields, so the other is a no-op (they can't overlap: close() only runs
+// while the bitmap is reachable, the Cleaner only after it isn't).
+@OptIn(ExperimentalForeignApi::class)
+private class SdlBitmapHandles(
+	var texture: COpaquePointer?,
+	var surface: CPointer<SDL_Surface>?,
+)
+
+/* Enqueue destruction of the handles' current contents on the main thread
+   (SDL calls aren't thread-safe; the Cleaner runs on a GC worker). Idempotent:
+   clears the fields so a second call frees nothing. Top-level so the Cleaner
+   block never captures the bitmap. */
+@OptIn(ExperimentalForeignApi::class)
+private fun enqueueRelease(inHandles: SdlBitmapHandles) {
+	val vTex = inHandles.texture
+	val vSurf = inHandles.surface
+	if (vTex == null && vSurf == null) return
+	inHandles.texture = null
+	inHandles.surface = null
+	com.compose.sdl.graphics.NativeReleaseQueue.enqueue {
+		if (vSurf != null) SDL_DestroySurface(vSurf)
+		if (vTex != null) SDL_DestroyTexture(vTex.reinterpret())
+	}
+}
+
 @OptIn(ExperimentalForeignApi::class)
 internal class SdlImageBitmap(
 	private val fRenderer: COpaquePointer,
@@ -41,19 +69,28 @@ internal class SdlImageBitmap(
 	// Sdl3EncodedImageDecoder); converted to a texture on the first draw.
 	// null → create a render TARGET for the vector-rasterisation path below
 	// (that path always constructs on the main thread).
-	private var fDecodedSurface: CPointer<SDL_Surface>? = null,
+	fDecodedSurface: CPointer<SDL_Surface>? = null,
 ) : ImageBitmap {
 
 	// RGBA render-target texture, premultiplied blend for compositing back (content
 	// is drawn over a transparent clear with ordinary BLEND, leaving premultiplied
 	// colours — see Sdl3ClipTargets for the same reasoning).
-	private var fTexture: COpaquePointer? = if (fDecodedSurface != null) null else SDL_CreateTexture(
+	private val fHandles = SdlBitmapHandles(
+		texture = if (fDecodedSurface != null) null else SDL_CreateTexture(
 			fRenderer.reinterpret(),
 			SDL_PIXELFORMAT_RGBA32,
 			SDL_TextureAccess.SDL_TEXTUREACCESS_TARGET,
 			maxOf(1, width),
 			maxOf(1, height),
-	)?.also { SDL_SetTextureBlendMode(it.reinterpret(), SDL_BLENDMODE_BLEND_PREMULTIPLIED) }
+		)?.also { SDL_SetTextureBlendMode(it.reinterpret(), SDL_BLENDMODE_BLEND_PREMULTIPLIED) },
+		surface = fDecodedSurface,
+	)
+
+	// GC backstop: if nothing calls close(), free the native handles when the
+	// bitmap is collected (deferred to the main thread via the queue). Ownership
+	// via close() is the primary path; this catches leaks.
+	@OptIn(kotlin.experimental.ExperimentalNativeApi::class)
+	private val fCleaner = kotlin.native.ref.createCleaner(fHandles) { enqueueRelease(it) }
 
 	// SDL renderer calls are NOT thread-safe, and the resources pipeline
 	// decodes on Dispatchers.Default workers — so the decode path hands over a
@@ -63,15 +100,20 @@ internal class SdlImageBitmap(
 	// render-target path above.
 	val texture: COpaquePointer?
 		get() {
-			fDecodedSurface?.let { vSurface ->
-				fTexture = SDL_CreateTextureFromSurface(fRenderer.reinterpret(), vSurface)
+			fHandles.surface?.let { vSurface ->
+				fHandles.texture = SDL_CreateTextureFromSurface(fRenderer.reinterpret(), vSurface)
 					?.also { SDL_SetTextureBlendMode(it.reinterpret(), SDL_BLENDMODE_BLEND) }
 				SDL_DestroySurface(vSurface)
-				fDecodedSurface = null
+				fHandles.surface = null
 			}
-			return fTexture
+			return fHandles.texture
 		}
 
+	/* Free the texture/surface now (deferred to the main thread). Call on
+	   cache eviction / removeMemoryResource so native memory releases promptly
+	   instead of waiting for a GC. The Cleaner remains as a backstop. */
+	fun close() = enqueueRelease(fHandles)
+
 	override fun readPixels(
 		buffer: IntArray,
 		startX: Int,

demo/src/jvmMain/kotlin/MainJvm.kt

@@ -1,16 +1,46 @@
+import androidx.compose.foundation.background
+import androidx.compose.foundation.layout.Box
+import androidx.compose.foundation.layout.fillMaxSize
+import androidx.compose.foundation.layout.padding
+import androidx.compose.foundation.rememberScrollState
+import androidx.compose.foundation.verticalScroll
 import androidx.compose.material3.MaterialTheme
 import androidx.compose.material3.darkColorScheme
+import androidx.compose.runtime.Composable
+import androidx.compose.ui.ExperimentalComposeUiApi
+import androidx.compose.ui.ImageComposeScene
+import androidx.compose.ui.Modifier
+import androidx.compose.ui.unit.Density
 import androidx.compose.ui.unit.dp
 import androidx.compose.ui.window.Window
 import androidx.compose.ui.window.application
 import androidx.compose.ui.window.rememberWindowState
+import demo.registry.allCategories
 import demo.shell.App
+import org.jetbrains.skia.EncodedImageFormat
+import java.io.File
 
-// The JVM comparison app's entry point. Runs the SAME shared App() shell + Core /
-// Material 3 screens (expressive included) as :demo, but on Compose Desktop (JVM)
-// against upstream org.jetbrains.compose. Compare side-by-side with
-// `:demo:runDebugExecutable<host>`.
-fun main() = application {
+// The JVM comparison app: the SAME shared App() + screens as :demo, on upstream
+// Compose Desktop. Interactive by default; a headless screenshot mode drives
+// the parity harness (scripts/parity — compares each screen native vs jvm).
+//
+//   --screenshot-all=<dir>   render every registered screen to <dir>/<Name>.png
+//   --width / --height       viewport size (default 1000 / 700)
+//
+// The single-screen wrapper MIRRORS MainNative's --screen path (dark theme,
+// verticalScroll + 24dp padding) so layout constraints match the native
+// screenshots pixel-for-pixel.
+fun main(args: Array<String>) {
+    val screenshotDir = args.firstOrNull { it.startsWith("--screenshot-all=") }?.substringAfter('=')
+    if (screenshotDir != null) {
+        screenshotAllScreens(
+            outDir = File(screenshotDir),
+            width = args.intArg("--width", 1000),
+            height = args.intArg("--height", 700),
+        )
+        return
+    }
+    application {
         Window(
             onCloseRequest = ::exitApplication,
             title = "ComposeDesktopNative — JVM (upstream Compose)",
@@ -21,3 +51,42 @@ fun main() = application {
             }
         }
     }
+}
+
+private fun Array<String>.intArg(name: String, default: Int): Int =
+    firstOrNull { it.startsWith("$name=") }?.substringAfter('=')?.toIntOrNull() ?: default
+
+/* Render each registered screen headlessly (density 1 to match the native
+   physical-pixel screenshots) and write a PNG per screen. */
+@OptIn(ExperimentalComposeUiApi::class)
+private fun screenshotAllScreens(outDir: File, width: Int, height: Int) {
+    outDir.mkdirs()
+    val screens = allCategories().flatMap { it.screens }.distinctBy { it.name }
+    for (screen in screens) {
+        val scene = ImageComposeScene(width, height, density = Density(1f)) {
+            ScreenHost { screen.content() }
+        }
+        try {
+            val image = scene.render()
+            val png = image.encodeToData(EncodedImageFormat.PNG) ?: continue
+            File(outDir, "${screen.name}.png").writeBytes(png.bytes)
+        } finally {
+            scene.close()
+        }
+        println("jvm screenshot: ${screen.name}")
+    }
+}
+
+/* Same wrapper as MainNative's --screen path. */
+@Composable
+private fun ScreenHost(content: @Composable () -> Unit) {
+    MaterialTheme(colorScheme = darkColorScheme()) {
+        Box(
+            modifier = Modifier
+                .fillMaxSize()
+                .background(MaterialTheme.colorScheme.background)
+                .verticalScroll(rememberScrollState())
+                .padding(24.dp),
+        ) { content() }
+    }
+}

gradle-plugin/compose-desktop-native-bridge/src/main/kotlin/com/bitsycore/compose/sdl/gradle/ResourcePackaging.kt

@@ -54,17 +54,24 @@ private val desktopTargets = mapOf(
 )
 
 private fun registerDataKresTasks(project: Project) {
-	val taskNames = project.tasks.names
-	val resPackage = project.resolveResourcePackage()
+	// The zips are PRE-REGISTERED lazily for every desktop target/variant and
+	// wired to their link task via matching{}.configureEach: executables may
+	// be declared after this afterEvaluate runs — notably by the bridge's OWN
+	// compose.desktop.native { entryPoint } DSL, which also materialises in
+	// afterEvaluate. (A name pre-scan raced that and silently packaged
+	// nothing — the "data.kres not found" crash; and Gradle forbids
+	// registering tasks from inside another task's configuration callback,
+	// so the zip cannot be created reactively either.) An unrealised
+	// registered task costs nothing: if the target has no executable, the
+	// link task never appears and the zip never runs. All content wiring
+	// happens in the zip's own configuration action, which only executes on
+	// realisation — by then the Compose plugin's prepare tasks and the final
+	// compose.resources config exist.
 	for ((target, sourceSets) in desktopTargets) {
 		for (variant in listOf("Debug", "Release")) {
 			val linkName = "link${variant}Executable$target"
-			if (linkName !in taskNames) continue
 			val zipName = "package${variant}ComposeResources$target"
-			if (zipName in taskNames) continue
-			val prepareNames = sourceSets
-				.map { "prepareComposeResourcesTaskFor" + it.replaceFirstChar { c -> c.uppercase() } }
-				.filter { it in taskNames }
+			if (zipName in project.tasks.names) continue
 			val zipTask = project.tasks.register(zipName, Zip::class.java) { task ->
 				task.description = "Bundles composeResources into data.kres next to the $target ${variant.lowercase()} executable."
 				task.archiveFileName.set("data.kres")
@@ -77,6 +84,10 @@ private fun registerDataKresTasks(project: Project) {
 				// decoders — an entry is one fseek+fread, never inflated.
 				task.entryCompression = ZipEntryCompression.STORED
 				task.duplicatesStrategy = DuplicatesStrategy.EXCLUDE
+				val resPackage = project.resolveResourcePackage()
+				val prepareNames = sourceSets
+					.map { "prepareComposeResourcesTaskFor" + it.replaceFirstChar { c -> c.uppercase() } }
+					.filter { it in project.tasks.names }
 				for (prepareName in prepareNames) {
 					val sourceSet = prepareName.removePrefix("prepareComposeResourcesTaskFor")
 						.replaceFirstChar { it.lowercase() }
@@ -87,7 +98,7 @@ private fun registerDataKresTasks(project: Project) {
 					task.dependsOn(project.tasks.named(prepareName))
 				}
 			}
-			project.tasks.named(linkName) { it.dependsOn(zipTask) }
+			project.tasks.matching { it.name == linkName }.configureEach { it.dependsOn(zipTask) }
 		}
 	}
 }

scripts/parity/README.md

@@ -0,0 +1,47 @@
+# Parity harness
+
+Renders every `:demo` screen on the **native** (SDL/Skia, Kotlin/Native) stack
+and on the **JVM** upstream-Compose stack from the *same* commonMain
+composables, then pixel-diffs them per screen. It's a regression net: a screen
+whose difference jumps far above its usual level is a port bug (missing
+content, wrong shape/colour, broken clip). Several of this project's renderer
+regressions would have surfaced here.
+
+```bash
+python scripts/parity/parity.py                 # all screens (builds first)
+python scripts/parity/parity.py Buttons Shapes  # a subset
+python scripts/parity/parity.py --no-build      # reuse the last renders
+python scripts/parity/parity.py --gpu=sdl3      # native renderer (default sdl3)
+```
+
+Output → `build/parity/` (gitignored):
+
+- `<pct>_<Name>_compare.png` — native ∣ jvm ∣ amplified-diff, side by side
+- `<pct>_<Name>_diff.png` — the amplified difference heatmap alone
+- `report.txt` — screens ranked by % differing
+
+The `<pct>` prefix is zero-padded, so a plain file listing (or the report)
+sorts worst-first.
+
+## Reading the result
+
+**Absolute % is not the metric — the ranking is.** The two stacks use
+different default fonts, so text carries a steady baseline difference (in the
+heatmap, every text line shows a faint *doubled* ghost from slightly different
+line metrics). Buttons, cards, shapes, images should align (dark). So:
+
+- **doubled/ghosted text, dark shapes** → normal font drift.
+- **a solid bright block, or a shape present on one side only** → a real
+  regression. Open the `_compare.png` to see which stack is wrong.
+
+Compare a screen's % against its neighbours and its own history: Buttons ~16%,
+Shapes ~14%, Colors ~4% are the healthy baseline (mostly text). A screen that
+reads 60% when text-light is the bug.
+
+## Requirements
+
+- Windows for the native leg today (mingwX64 exe); the JVM leg is host-neutral.
+- Pillow (`pip install pillow`).
+- The JVM leg renders all screens headlessly via `ImageComposeScene` in one
+  process (`:demo:run --args=--screenshot-all=…`); native takes one exe launch
+  per screen.

scripts/parity/parity.py

@@ -0,0 +1,156 @@
+#!/usr/bin/env python3
+"""
+Parity harness — render every demo screen on the NATIVE (SDL/Skia) stack and on
+the JVM upstream-Compose stack, then pixel-diff them per screen.
+
+The two stacks share the exact same commonMain screen composables, so a screen
+that suddenly diverges from its usual difference level is a PORT REGRESSION
+(missing content, wrong shape/colour, broken clip). Absolute pixel-perfection
+is NOT the goal — fonts differ between stacks, so text-heavy screens carry a
+steady baseline difference. The signal is the RANKING: a screen that jumps from
+~8% to ~60% different is the bug. Four of the renderer regressions this project
+hit would have surfaced here.
+
+Usage (from repo root):
+    python scripts/parity/parity.py                 # all screens
+    python scripts/parity/parity.py Buttons Shapes  # a subset
+    python scripts/parity/parity.py --no-build      # reuse existing screenshots
+
+Outputs to build/parity/ (gitignored):
+    <pct>_<Name>_compare.png   native | jvm | amplified-diff, side by side
+    <pct>_<Name>_diff.png      the amplified pixel-difference heatmap alone
+    report.txt                 ranked table
+The <pct> prefix is zero-padded so a plain file listing sorts worst-first.
+
+Windows-only for the native leg today (mingwX64 exe). Needs Pillow.
+Native uses the SDL renderer; pass --gpu to change.
+"""
+import subprocess, sys, os, shutil
+from pathlib import Path
+from PIL import Image, ImageChops, ImageDraw
+
+REPO = Path(__file__).resolve().parents[2]
+OUT = REPO / "build" / "parity"
+NATIVE_EXE = REPO / "demo" / "build" / "bin" / "mingwX64" / "debugExecutable" / "demo.exe"
+WIDTH, HEIGHT = 1000, 700
+# Per-channel tolerance: below this a pixel counts as "same" (JPEG-ish noise,
+# sub-pixel AA, font hinting). Tuned so unrelated screens sit well under it.
+TOL = 32
+
+GRADLEW = str(REPO / ("gradlew.bat" if os.name == "nt" else "gradlew"))
+
+
+def run(cmd, **kw):
+    print("+ " + " ".join(str(c) for c in cmd))
+    return subprocess.run(cmd, cwd=REPO, **kw)
+
+
+def build():
+    run([GRADLEW, ":demo:linkDebugExecutableMingwX64", "--console=plain"], check=True)
+
+
+def jvm_shots(dst: Path):
+    run([GRADLEW, ":demo:run", f"--args=--screenshot-all={dst}", "--console=plain"], check=True)
+
+
+def native_shot(name: str, dst: Path, gpu: str):
+    bmp = dst / f"{name}.bmp"
+    run([str(NATIVE_EXE), f"--screen={name}", f"--screenshot={bmp}",
+         f"--gpu={gpu}", f"--width={WIDTH}", f"--height={HEIGHT}"],
+        check=False, timeout=60)
+    if bmp.exists():
+        Image.open(bmp).convert("RGB").save(dst / f"{name}.png")
+        bmp.unlink()
+
+
+def diff_pair(native: Image.Image, jvm: Image.Image):
+    """Return (percent_differing, amplified_diff_image)."""
+    w = min(native.width, jvm.width)
+    h = min(native.height, jvm.height)
+    a = native.crop((0, 0, w, h)).convert("RGB")
+    b = jvm.crop((0, 0, w, h)).convert("RGB")
+    diff = ImageChops.difference(a, b)
+    # A pixel "differs" if any channel exceeds TOL.
+    gray = diff.convert("L")
+    mask = gray.point(lambda p: 255 if p > TOL else 0)
+    differing = sum(mask.point(lambda p: 1 if p else 0).getdata())
+    pct = 100.0 * differing / (w * h)
+    amplified = diff.point(lambda p: min(255, p * 4))
+    return pct, amplified
+
+
+def side_by_side(native, jvm, diff, path):
+    w = min(native.width, jvm.width)
+    h = min(native.height, jvm.height)
+    canvas = Image.new("RGB", (w * 3 + 20, h + 20), (16, 16, 16))
+    canvas.paste(native.crop((0, 0, w, h)), (0, 10))
+    canvas.paste(jvm.crop((0, 0, w, h)), (w + 10, 10))
+    canvas.paste(diff, (w * 2 + 20, 10))
+    canvas.save(path)
+
+
+def main():
+    argv = [a for a in sys.argv[1:] if not a.startswith("--")]
+    flags = [a for a in sys.argv[1:] if a.startswith("--")]
+    no_build = "--no-build" in flags
+    gpu = next((f.split("=", 1)[1] for f in flags if f.startswith("--gpu=")), "sdl3")
+
+    OUT.mkdir(parents=True, exist_ok=True)
+    jvm_dir = OUT / "_jvm"
+    # Clear last run's pct-prefixed visualizations so stale percentages don't
+    # linger alongside the fresh ones.
+    for old in list(OUT.glob("*_compare.png")) + list(OUT.glob("*_diff.png")):
+        old.unlink()
+
+    if not no_build:
+        build()
+        if jvm_dir.exists():
+            shutil.rmtree(jvm_dir)
+        jvm_dir.mkdir(parents=True)
+        jvm_shots(jvm_dir)
+
+    # Screen set: the JVM run enumerates the full registry → use its PNGs as the
+    # source of truth for names (native takes one exe launch each).
+    names = sorted(p.stem for p in jvm_dir.glob("*.png"))
+    if argv:
+        wanted = {n.lower() for n in argv}
+        names = [n for n in names if n.lower() in wanted]
+    if not names:
+        print("No screens matched.", file=sys.stderr)
+        return 1
+
+    results = []
+    for name in names:
+        jvm_png = jvm_dir / f"{name}.png"
+        if not no_build or not (OUT / f"{name}.native.png").exists():
+            native_shot(name, OUT, gpu)
+            src = OUT / f"{name}.png"
+            if src.exists():
+                src.replace(OUT / f"{name}.native.png")
+        native_png = OUT / f"{name}.native.png"
+        if not native_png.exists():
+            results.append((name, None))
+            continue
+        native = Image.open(native_png).convert("RGB")
+        jvm = Image.open(jvm_png).convert("RGB")
+        pct, amp = diff_pair(native, jvm)
+        # Zero-padded pct prefix → worst-first in any file listing.
+        prefix = f"{pct:06.2f}"
+        amp.save(OUT / f"{prefix}_{name}_diff.png")
+        side_by_side(native, jvm, amp, OUT / f"{prefix}_{name}_compare.png")
+        results.append((name, pct))
+        print(f"  {name:28s} {pct:6.2f}% differing")
+
+    results.sort(key=lambda r: (-1 if r[1] is None else r[1]), reverse=True)
+    lines = ["screen                        %differ", "-" * 40]
+    for name, pct in results:
+        lines.append(f"{name:28s}  {'NATIVE FAILED' if pct is None else f'{pct:6.2f}%'}")
+    report = "\n".join(lines)
+    (OUT / "report.txt").write_text(report + "\n", encoding="utf-8")
+    print("\n" + report)
+    print(f"\nDiff images + report in {OUT}")
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())