fix(sdl): stop re-rasterizing text at device size — popup/dialog text no longer steps size/weight

Regression from 120abbad's glyph device-scaling: drawNativeText scaled
the font size by the layer affine (inFontSize * scaleY, int-quantized)
and re-rasterized. Under an animated layer scale — every Material
DropdownMenu/AlertDialog enter transition scale-animates (Menu.kt
scaleX/scaleY) — text re-rasterized at each integer size every frame, so
it visibly stepped through sizes and (via SDL3_ttf re-hinting) weights
as the popup grew in. User-reported, seen across dropdowns and dialogs.

Revert to rasterising at LOGICAL font size + 1:1 blit at the
affine-mapped position (the pre-120abbad behaviour): the layer scale
reaches text POSITION but not glyph size, so animated-scale popups are
stable. Keeps 120abbad's separate clip-mask AA/overlap fixes. Trade-off:
statically graphicsLayer-scaled text renders at logical size (a rare,
minor parity gap) — smoothly scaling glyphs needs blit-scaling a
logical rasterisation (noted for a future text-renderer pass).

Verified: demo dropdown opens with clean uniform text; text no longer
scales with the layer, so mid-animation == settled by construction.

CLAUDE.md

@@ -458,7 +458,57 @@ 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
+## Tooling — what's available and when to reach for it
+
+Index of the repo's tooling. Each entry says when to use it and points at its
+own README for detail; the parity/profiler/probe entries are expanded below
+since they're newer.
+
+| Tool | Reach for it when | Detail |
+|------|-------------------|--------|
+| `python scripts/build-sdl/build-all.py` | building/refreshing the static SDL3/TTF/image/FreeType libs under `libs/` (once per host, or after bumping `build-sdl.properties`) | "Building" above |
+| `scripts/compose-fork/sync.sh` | re-syncing vendored upstream `androidx.compose.*` after a fresh checkout or a `compose.properties` ref bump; `format-manifest.py` to re-align a `compose-fork.txt` | "Vendor sync workflow" above + `scripts/compose-fork/README.md` |
+| `./gradlew apiDump && python scripts/compose-coverage.py` | measuring how much upstream public API the port actually covers, per module (`--missing <module>` lists uncovered decls) | "Vendoring" above |
+| `scripts/generate-material-symbols.py` / `subset-material-symbols.py` | regenerating the Material Symbols codepoints, or hb-subsetting bundled icon fonts to used glyphs (the latter runs automatically in app Zip tasks under `-PsubsetIcons`) | — |
+| **`scripts/parity/parity.py`** | after ANY renderer/layout change: catch a screen that visually diverged native-vs-JVM (missing content, wrong shape/colour, broken clip) | `scripts/parity/README.md` + below |
+| **`scripts/probe/probe.py`** | reproducing a specific interaction bug (click/hover/hold at a point) or grabbing one screen's pixels deterministically | `scripts/probe/README.md` + below |
+| **`CDN_PROFILE=1 <app>`** | finding where a slow frame goes (per-phase main-loop timings) before optimizing | below |
+| bridge plugin (`com.bitsycore.compose-desktop-native.bridge`) | consuming the published klibs from a third-party app | `gradle-plugin/compose-desktop-native-bridge/README.md` |
+| `demo --screen=<Name>` / `--screenshot=` / `--nav3test` / `--backtest` / `--multiwintest` | driving one screen headless, or the regression probes for nav3 / predictive-back / multi-window | `demo/src/nativeMain/kotlin/MainNative.kt` |
+
+Whole-project renderer-touching change → run **parity** (broad net). Chasing
+one reported interaction → **probe** (targeted). Slow → **profiler** first,
+optimize second. See `ROADMAP.md` for the renderer work these support.
+
+### Frame profiler — `CDN_PROFILE=1`
+
+Set the env var (`CDN_PROFILE=1`, or `=<path>` for a specific file) and run any
+native app; every ~2 s of rendered frames it writes avg/max ms per main-loop
+phase (`events` / `app` pump / `pump` per-window / `render`) to a FILE
+(`cdn_profile.log` by default — works for GUI-subsystem apps like the demo that
+have no console). The line also carries render SUB-phases (`layout` / `draw` /
+`present`) and per-frame DRAW COUNTERS from `DrawStats` (`geo` =
+SDL_RenderGeometry submits, `verts`, `masks` = rounded-clip offscreen passes,
+`text`, `img` blits) — so you can see WHAT `draw` is doing, not just that it's
+slow. `CDN_FORCERENDER=1` renders every frame (bypasses idle-skip) so
+steady-state timings can be read on otherwise-idle static screens.
+
+**Caveat learned the hard way:** `present` is vsync-blocking, so every timing is
+capped by the DISPLAY refresh of the machine running the app. A 75 Hz dev
+display makes every app look "75 fps, present-bound" no matter what — so profile
+on the TARGET refresh rate before concluding anything about a frame-rate gap
+(the demo-70 / apidemo-144 report could not be reproduced on a 75 Hz box, where
+both pinned at 75 fps and the demo's `draw` was actually the lighter of the two).
+
+### Interaction probe — `scripts/probe/`
+
+Launches a native app, sends **window-client-relative** input (click / hover /
+hold, fractional coords addressed by process name so it ignores window
+position/focus) and captures the client area via `PrintWindow` (works even
+when occluded). The packaged form of the rigs that reproduced the
+square-on-click and TLS-chain bugs. Windows-only. See its README.
+
+### Parity harness — native-vs-JVM screenshot diff (`scripts/parity/`)
 
 `:demo` renders the **same commonMain screens** on two stacks: native
 (SDL/Skia, Kotlin/Native) and a `jvm()` target on upstream Compose Desktop.

NEXT-SESSION.md

@@ -0,0 +1,114 @@
+# Next session — renderer caching / retained layers
+
+## RESOLVED: the 70-vs-144 "gap" was two monitors, not a bug
+
+The user has a **75 Hz main screen + a 144 Hz second screen**. Each app's FPS is
+just the refresh of whichever monitor its window opened on (SDL locks vsync to
+the window's current display). demo happened to open on the 75 Hz screen,
+apidemo on the 144 Hz one. Move the demo to the 144 Hz screen and it hits ~144.
+
+Instrumentation confirmed there is NO performance problem: both apps have only
+~3.7 ms `draw` per frame (demo geo=40 masks=0; apidemo geo=146 masks=74 —
+apidemo does MORE and is fine), which fits both the 13.3 ms (75 Hz) and 6.7 ms
+(144 Hz) budgets with large headroom. Everything is present/vsync-bound.
+
+**Implication:** the retained-layers / dirty-region / cacheKey work is a
+SPECULATIVE optimization with NO demonstrated need right now — no app misses its
+frame budget. Don't spend the big content-redirect-primitive effort chasing a
+phantom. Revisit only if a real, animated, heavy screen is measured (with
+CDN_PROFILE, ON the target monitor) to exceed its budget.
+
+Still worth doing regardless of perf (correctness / cleanup, small):
+- Delete the dead `cacheKey` scaffolding (`GraphicsLayerModifier` /
+  `GraphicsLayerNode`, unread) OR wire it — and fix the demo's misleading
+  "(cached)" label (it uses a plain `graphicsLayer()`).
+
+--- Original investigation notes below (kept for context) ---
+
+
+Starting point after v0.1.19. Context for the perf work the profiling pointed at.
+
+## The problem (measured)
+
+The demo caps at ~70 fps on every screen while apidemo hits 144. Cause is NOT
+vsync (both use `SDL_SetRenderVSync(renderer, 1)`) — the demo is **draw-bound**:
+`CDN_PROFILE=1` showed draw ~34 ms cold / ~14 ms steady, present only ~1.3 ms.
+The renderer is **immediate-mode**: every rendered frame walks the layout tree
+and re-tessellates everything (text → glyph quads, shapes → triangles). The
+always-present sidebar (30+ text rows + icons) is re-tessellated each frame.
+
+Two open threads:
+
+### FINDING (updated) — the 70-vs-144 is NOT reproducible on a 75 Hz dev box
+
+Instrumented the draw phase (`DrawStats`: geo/verts/masks/text/img) + added
+`CDN_FORCERENDER=1`. Measured both apps forced-continuous:
+- demo:    draw ~3.7ms, geo=40,  verts=6582,  masks=0,  text=51
+- apidemo: draw ~2.9ms, geo=146, verts=18576, masks=74, text=174
+BOTH pinned at exactly 150 frames/2s = 75 fps with present ~9ms — i.e. the DEV
+DISPLAY is 75 Hz, so everything is present/vsync-bound here and the demo's draw
+is actually LIGHTER than apidemo's. The "demo is draw-bound" theory is WRONG on
+this hardware; the 70-vs-144 split is specific to the user's 144 Hz monitor and
+can't be reproduced/diagnosed from a 75 Hz box.
+
+NEXT STEP FOR THE USER (144 Hz machine): run both apps with
+`CDN_FORCERENDER=1 CDN_PROFILE=1` and share the `cdn_profile.log` lines. If demo
+shows present ~13ms (75 fps) while apidemo shows present ~7ms (144 fps) at
+similar draw times, it's a present/vsync-path difference (driver / swap
+interval / DWM), NOT draw cost — a different investigation than caching.
+
+The demo idles correctly when static (verified: 2 frames headless, even while
+hovering the sidebar), so there's NO spurious-continuous-render bug — thread A
+below is resolved as a non-issue.
+
+### A. Why does the demo render CONTINUOUSLY? (investigate FIRST — likely cheaper)
+
+Headless the demo idled (2 frames); interactively you see a constant 70-71 fps,
+so something invalidates every frame. Prime suspect: a **hover self-loop** — the
+main loop dispatches a synthetic hover every rendered frame
+(`ComposeWindow.renderFrame`: `if (hasMousePos) host.onPointerRaw(...)`), and if
+that perpetually re-invalidates a hover-reactive sidebar row, the app never
+idles while the cursor is over it. If confirmed, the fix is to stop the spurious
+invalidation so static screens idle (→ sidebar stops re-tessellating because
+nothing renders) — small and targeted, no caching needed.
+- Quick test: does the demo's FPS-title stop updating when the mouse leaves the
+  window? If yes → hover loop confirmed.
+- Look at: `renderFrame` synthetic hover; how hover state feeds `needsFrame` /
+  `hasPendingWork` / `shouldRender()`.
+
+### B. cacheKey / retained-layer texture caching (bigger — the "renderer rewrite")
+
+`Modifier.graphicsLayer(cacheKey=…)` exists as API (`GraphicsLayerModifier` /
+`GraphicsLayerNode` in `element/ModifierElements.kt`) but is **dead
+scaffolding**: `cacheKey` is stored and never read by the renderer; the node is
+a bare `Modifier.Node` with no draw behaviour (comment: "stays dormant until the
+renderer rewrite drives it"). The demo's GraphicsLayer "(cached)" section uses a
+plain `graphicsLayer()` — the label is aspirational; nothing is cached.
+
+Blocker for real texture caching: it needs to render a subtree's `drawContent()`
+into an **offscreen canvas**, and the renderer has no way to redirect
+`drawContent()` off the frame canvas. `GraphicsLayer.native.draw()` replays its
+recorded block against a canvas via `drawScope.draw(…, canvas, …)`, but
+`drawContent()` is bound to the outer `ContentDrawScope`'s canvas, which nothing
+swaps. The vector/icon offscreen path works only because it draws an explicit
+object, not opaque `drawContent()`.
+
+So the real unit of work is a **content-redirect primitive**: make the
+`ContentDrawScope` canvas swappable (or add a coordinator-level "draw this
+subtree into canvas X"). That primitive is the foundation for BOTH cacheKey
+(retained layers) AND dirty-region rendering. Infrastructure that already
+exists to build on: `OffscreenRenderer` (`createImageBitmap` + `createCanvas`),
+`SdlImageBitmap` render-target textures + `drawImage` blit-back, and the
+`NativeReleaseQueue` for freeing cached textures.
+
+Cleanup to do regardless: either implement `cacheKey` or delete the dead
+`GraphicsLayerModifier`/`GraphicsLayerNode` scaffolding and fix the demo's
+misleading "(cached)" label.
+
+## Tooling ready for this work
+
+- `CDN_PROFILE=1 <app>` → per-phase timings (layout/draw/present) to a file.
+- `scripts/parity/parity.py` → native-vs-JVM screenshot diff (regression net).
+- `scripts/probe/probe.py` → drive a native window (click/hover/hold) + capture.
+- See ROADMAP.md item 2 (dirty regions + retained layers promoted, with the
+  demo evidence) and CLAUDE.md "Tooling".

README.md

@@ -261,6 +261,24 @@ RESUMED, unfocused → STARTED, minimized → CREATED), `ViewModelStoreOwner`
 ViewModels scope per nav entry, per window, or anywhere in between — same
 semantics as Android.
 
+## Tooling
+
+Helper scripts live under `scripts/`; each has its own README with detail.
+
+| Tool | Use it to | More |
+|------|-----------|------|
+| `scripts/build-sdl/build-all.py` | build the static SDL3 / TTF / image / FreeType libs | see **Building** above |
+| `scripts/compose-fork/sync.sh` | re-sync the vendored upstream Compose sources | [README](scripts/compose-fork/README.md) |
+| `scripts/parity/parity.py` | diff every demo screen native-vs-JVM to catch render regressions | [README](scripts/parity/README.md) |
+| `scripts/probe/probe.py` | drive a native window (click/hover/hold) + screenshot it | [README](scripts/probe/README.md) |
+| `CDN_PROFILE=1 <app>` | print per-phase frame timings to find slow frames | — |
+| `python scripts/compose-coverage.py` | measure API coverage vs upstream (after `./gradlew apiDump`) | — |
+
+Building on the port from another project? Use the bridge plugin —
+[gradle-plugin/compose-desktop-native-bridge](gradle-plugin/compose-desktop-native-bridge/README.md).
+Contributor-facing detail (including how to read the parity `%differ`) is in
+[CLAUDE.md](CLAUDE.md#tooling--whats-available-and-when-to-reach-for-it).
+
 ## License
 
 [MIT](LICENSE.md).

ROADMAP.md

@@ -42,16 +42,23 @@ RSS; expect a plateau without GC sawtooth).
 - [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
+- [ ] **Dirty-region rendering** — NEXT, promoted by profiling: the app
-  textures per mask depth; allocate at the clip bbox (size-bucketed pool) to
+  re-tessellates the whole window on any invalidation. Accumulate damage from
-  cut fill-rate for the masks that remain after lazy clips.
+  invalidated layers and scissor the redraw. Biggest lever for heavy scenes.
-- [ ] **Dirty-region rendering** — the big structural item: accumulate damage
+  Evidence (CDN_PROFILE): the full demo is DRAW-bound — draw ~34ms cold /
-  from invalidated layers, scissor redraw to it. One bubble's spring should
+  ~14ms steady, present only ~1.3ms — because it re-tessellates the whole
-  not re-tessellate 84. Most likely single item to reach 75 Hz.
+  always-present sidebar (30+ text rows + icons) every rendered frame. That
+  ~14ms caps it near 70 fps; apidemo's lighter chrome fits the 6.9ms budget
+  and reaches the 144 Hz vsync ceiling. The "70 ≈ 144/2" look is coincidence,
+  NOT a vsync-half artifact — both use vsync(1). Retained layers / dirty
+  regions would stop re-tessellating the static sidebar.
+- [ ] **Retained layer textures** — cache a layer's rendered output keyed by
+  its draw content (RenderNode-style); static subtrees stop re-tessellating.
+- [ ] **Right-size clip scratch targets** (LOW — profiling showed masks aren't
+  a hot path after lazy clips): size scratch textures to the clip bbox.
+  Memory, not frame time.
 - [ ] **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
 
@@ -62,12 +69,18 @@ RSS; expect a plateau without GC sawtooth).
 
 ## 4. Correctness / parity gaps (future bug reports waiting)
 
-Ordered by likelihood of a user hitting them:
+Ordered by likelihood of a user hitting them. NOTE: this list predates several
+fixes — verify a claim against the parity heatmaps before acting (gradients and
+rotation both turned out already-working). In a parity `_diff.png`: a solid
+bright block = real gap; ghosted text = font drift.
 
-- [ ] **Gradient brushes on SDL** — `Brush.linearGradient/radialGradient`
+- [x] **Gradient brushes on SDL** — samplers were already implemented; the real
-  render solid; the per-vertex sampler infrastructure is half-present.
+  bug was UNDER-SAMPLING: fills sampled only corners, so radial/sweep on a
-- [ ] **Layer rotation** — rotationZ repositions but content doesn't rotate
+  rect/round-rect rendered FLAT (a square's corners are equidistant from centre
-  (hit-testing too).
+  → identical t). Gradient fills now grid-mesh the interior (emitRectMesh);
+  radial/sweep match upstream. (Solid fills unchanged — one quad.)
+- [x] **Layer rotation** — verified WORKING (demo GraphicsLayer "Rotation
+  (live)" + "Combined" rotate correctly); stale claim removed.
 - [ ] **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.
@@ -78,8 +91,9 @@ Ordered by likelihood of a user hitting them:
   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 +
+- [x] Promote the press/hover automation rig into `scripts/probe/`:
-  screenshot) into `scripts/`.
+  window-CLIENT-relative input (click/hover/hold, fractional coords, process-
+  addressed) + PrintWindow capture (works occluded). See scripts/probe/README.md.
 
 ## 6. Long-term
 

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

@@ -32,6 +32,7 @@ import com.compose.sdl.window.createPopupHostState
 import kotlinx.coroutines.*
 import kotlinx.coroutines.test.resetMain
 import kotlinx.coroutines.test.setMain
+import kotlinx.cinterop.toKString
 import kotlinx.cinterop.reinterpret
 import sdl3.SDL_Delay
 import sdl3.SDL_GetPerformanceCounter
@@ -150,9 +151,8 @@ 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()
+			FrameProfiler.mark()
 			Snapshot.sendApplyNotifications()
 
 			// ============
@@ -180,7 +180,7 @@ fun nativeComposeApp(content: @Composable ApplicationScope.() -> Unit) {
 			}
 
 			mainDispatcher.drainPending()
-			vProfiler?.phase(0)
+			FrameProfiler.phase("events")
 
 			// ============
 			//  App composition pump — Window()s may appear / disappear here.
@@ -200,7 +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)
+			FrameProfiler.phase("app")
 
 			// ============
 			//  Per-window pump + render.
@@ -214,16 +214,16 @@ fun nativeComposeApp(content: @Composable ApplicationScope.() -> Unit) {
 				Snapshot.sendApplyNotifications()
 				vW.frameClock.sendFrame()
 				yield()
-				vProfiler?.phase(2)
+				FrameProfiler.phase("pump")
 				if (vW.shouldRender()) {
 					vW.renderFrame()
 					vAnyRendered = true
 					if (!vW.backend.vsyncEnabled) vAllVsync = false
 				}
-				vProfiler?.phase(3)
+				FrameProfiler.phase("render")
 			}
 			runtime.reapDestroyed()
-			vProfiler?.frameDone(vAnyRendered)
+			FrameProfiler.frameDone(vAnyRendered)
 
 			// ============
 			//  Drain deferred native-resource disposals on the MAIN thread —
@@ -280,41 +280,68 @@ fun nativeComposeApp(content: @Composable ApplicationScope.() -> Unit) {
 	SDL_Quit()
 }
 
-/* CDN_PROFILE=1 — per-phase main-loop timings, printed every ~2s of rendered
+/* CDN_PROFILE=1 — per-phase timings, printed every ~2s of rendered frames.
-   frames. Phases: events (poll+dispatch), app (app-composition pump), pump
+   A named-phase SINGLETON so both the main loop (events / app / pump / render)
-   (per-window clocks/dispatch), render (layout+draw+present). Measure first,
+   AND renderFrame's sub-steps (render.layout / render.draw / render.present)
-   optimize second — see ROADMAP.md. */
+   report into one line. `mark()` resets the stopwatch; `phase(name)` charges
+   the elapsed since the last mark/phase to that name. Measure first, optimize
+   second — see ROADMAP.md. */
 @OptIn(kotlinx.cinterop.ExperimentalForeignApi::class)
-internal class FrameProfiler {
+internal object FrameProfiler {
+	// null until first checked; then true/false for the run's lifetime.
+	private var fEnabled: Boolean? = null
+	// Output file — resolved once from CDN_PROFILE. Writing to a file (not
+	// stdout) lets GUI-subsystem apps (the demo links --subsystem,windows, so
+	// it has no console) be profiled too. CDN_PROFILE=1 → "cdn_profile.log" in
+	// the cwd; CDN_PROFILE=<path> → that path.
+	private var fPath: String = "cdn_profile.log"
+	val enabled: Boolean
+		get() = fEnabled ?: run {
+			val vEnv = platform.posix.getenv("CDN_PROFILE")?.toKString()
+			if (vEnv != null && vEnv != "1" && vEnv.isNotEmpty()) fPath = vEnv
+			(vEnv != null).also { fEnabled = it }
+		}
+
 	private val fFreq = SDL_GetPerformanceFrequency().toDouble()
-	private val fSum = DoubleArray(4)
+	// Insertion-ordered so the printed line follows the call order.
-	private val fMax = DoubleArray(4)
+	private val fSum = LinkedHashMap<String, Double>()
+	private val fMax = LinkedHashMap<String, Double>()
 	private var fFrames = 0
 	private var fLastPrintMs = SDL_GetTicks()
 	private var fMark = 0uL
 
-	fun mark() { fMark = SDL_GetPerformanceCounter() }
+	fun mark() { if (enabled) fMark = SDL_GetPerformanceCounter() }
 
-	fun phase(inIndex: Int) {
+	fun phase(inName: String) {
+		if (!enabled) return
 		val vNow = SDL_GetPerformanceCounter()
 		val vMs = (vNow - fMark).toDouble() * 1000.0 / fFreq
-		fSum[inIndex] += vMs
+		fSum[inName] = (fSum[inName] ?: 0.0) + vMs
-		if (vMs > fMax[inIndex]) fMax[inIndex] = vMs
+		if (vMs > (fMax[inName] ?: 0.0)) fMax[inName] = vMs
 		fMark = vNow
 	}
 
 	fun frameDone(inRendered: Boolean) {
+		if (!enabled) return
 		if (inRendered) fFrames++
 		val vNowMs = SDL_GetTicks()
 		if (vNowMs - fLastPrintMs >= 2000u && fFrames > 0) {
-			val vNames = listOf("events", "app", "pump", "render")
+			val vParts = fSum.keys.map { vName ->
-			val vParts = vNames.mapIndexed { vI, vName ->
+				val vAvg = (fSum[vName] ?: 0.0) / fFrames
-				val vAvg = fSum[vI] / fFrames
+				"$vName=${(vAvg * 100).toInt() / 100.0}/${((fMax[vName] ?: 0.0) * 100).toInt() / 100.0}ms"
-				"$vName=${(vAvg * 100).toInt() / 100.0}/${(fMax[vI] * 100).toInt() / 100.0}ms"
 			}
-			val vTotal = fSum.sum() / fFrames
+			// Per-frame draw-work averages (see DrawStats): what's inside `draw`.
-			println("[profile] frames=$fFrames avg/max " + vParts.joinToString(" ") + " total=${(vTotal * 100).toInt() / 100.0}ms")
+			val vStats = com.compose.sdl.graphics.DrawStats
-			for (vI in fSum.indices) { fSum[vI] = 0.0; fMax[vI] = 0.0 }
+			val vDraw = "geo=${vStats.geometrySubmits / fFrames} verts=${vStats.vertices / fFrames} " +
+				"masks=${vStats.maskRealizations / fFrames} text=${vStats.textDraws / fFrames} img=${vStats.imageBlits / fFrames}"
+			val vLine = "[profile] frames=$fFrames avg/max " + vParts.joinToString(" ") + " | per-frame " + vDraw + "\n"
+			val vFile = platform.posix.fopen(fPath, "a")
+			if (vFile != null) {
+				platform.posix.fputs(vLine, vFile)
+				platform.posix.fclose(vFile)
+			}
+			vStats.reset()
+			fSum.clear(); fMax.clear()
 			fFrames = 0
 			fLastPrintMs = vNowMs
 		}
@@ -323,6 +350,8 @@ internal class FrameProfiler {
 
 /* Trigger a Kotlin/Native GC so Cleaner-managed renderer resources release
    their native memory (see the main loop's native-memory nudge). */
+private val kForceRender: Boolean = platform.posix.getenv("CDN_FORCERENDER") != null
+
 @OptIn(kotlin.native.runtime.NativeRuntimeApi::class)
 private fun collectNativeGarbage() = kotlin.native.runtime.GC.collect()
 
@@ -723,7 +752,10 @@ internal class WindowInstance(
 	//  Frame pump
 
 	fun shouldRender(): Boolean =
-		needsFrame || (recomposer?.hasPendingWork == true) || onFrame != null
+		needsFrame || (recomposer?.hasPendingWork == true) || onFrame != null || kForceRender
+
+	// TEMP measurement: CDN_FORCERENDER=1 forces every frame to render so
+	// sustained steady-state timings can be measured on otherwise-idle screens.
 
 	fun renderFrame() {
 		val vRender = renderBackend ?: return
@@ -734,6 +766,7 @@ internal class WindowInstance(
 		vRender.ensureSize(backend.pixelWidth, backend.pixelHeight)
 		host.setConstraints(backend.pixelWidth, backend.pixelHeight)
 		host.measureAndLayout()
+		FrameProfiler.phase("  layout")
 
 		// Hover refresh after layout (upstream skiko: SyntheticEventSender).
 		if (hasMousePos) {
@@ -746,7 +779,9 @@ internal class WindowInstance(
 			// Probe consumers end the app when their scenario completes.
 			facade.close()
 		}
+		FrameProfiler.phase("  draw")
 		vRender.endFrame()
+		FrameProfiler.phase("  present")
 		frameIndex++
 
 		// FPS — refreshed ~once a second, per window.

compose/ui/ui/src/commonMain/kotlin/com/compose/sdl/element/ModifierElements.kt

@@ -1,97 +0,0 @@
-package com.compose.sdl.element
-
-import androidx.compose.ui.Modifier
-import androidx.compose.ui.graphics.Shape
-import androidx.compose.ui.graphics.TransformOrigin
-import androidx.compose.ui.node.ModifierNodeElement
-
-// ==================
-// MARK: Project modifier elements
-// ==================
-// Small set of project-only modifier elements — each pairs a
-// `ModifierNodeElement<XxxNode>` with a `XxxNode : Modifier.Node`. Upstream
-// LayoutNode reads them through the chain (ModifierNodeElement IS-A
-// Modifier.Element). Everything gesture-related has been removed — pointer
-// input, text input, right/middle click, drag: all go through the standard
-// `Modifier.pointerInput` + `detectTapGestures` / `detectDragGestures` /
-// `awaitPointerEventScope` APIs upstream now supplies. What's left is the
-// two elements the renderer still owns directly:
-//   * ClipModifier — GraphicsLayer.kt lowers `clip = true` to it.
-//   * GraphicsLayerModifier — the transform / alpha / cache pipeline.
-
-class ClipModifier(val shape: Shape) : ModifierNodeElement<ClipNode>() {
-    override fun create() = ClipNode(shape)
-    override fun update(node: ClipNode) { node.shape = shape }
-    override fun hashCode(): Int = shape.hashCode()
-    override fun equals(other: Any?): Boolean = other is ClipModifier && other.shape == shape
-}
-class ClipNode(var shape: Shape) : Modifier.Node()
-
-// ==================
-// MARK: GraphicsLayerModifier
-// ==================
-
-/**
- * A "graphics layer" element: alpha + 2D transform (scale / rotation /
- * translation), with an optional cacheKey that opts the subtree into
- * render-to-texture caching across frames. See `Modifier.graphicsLayer`
- * (in `androidx.compose.ui.graphics`) for the caching semantics.
- *
- * The renderer reads this element directly via the `LayoutNode.graphicsLayer`
- * `foldIn` over the chain; the paired [GraphicsLayerNode] lifecycle stays
- * dormant until the renderer rewrite drives it.
- */
-class GraphicsLayerModifier(
-    val alpha: Float = 1f,
-    val scaleX: Float = 1f,
-    val scaleY: Float = 1f,
-    val rotationZ: Float = 0f,
-    val translationX: Float = 0f,
-    val translationY: Float = 0f,
-    val transformOrigin: TransformOrigin = TransformOrigin.Center,
-    val cacheKey: Any? = null,
-) : ModifierNodeElement<GraphicsLayerNode>() {
-
-    val needsLayer: Boolean
-        get() = alpha < 1f || cacheKey != null
-
-    val needsTransform: Boolean
-        get() = scaleX != 1f || scaleY != 1f || rotationZ != 0f ||
-            translationX != 0f || translationY != 0f
-
-    val isIdentity: Boolean
-        get() = !needsLayer && !needsTransform
-
-    override fun create() =
-        GraphicsLayerNode(alpha, scaleX, scaleY, rotationZ, translationX, translationY, transformOrigin, cacheKey)
-    override fun update(node: GraphicsLayerNode) {
-        node.alpha = alpha; node.scaleX = scaleX; node.scaleY = scaleY
-        node.rotationZ = rotationZ; node.translationX = translationX; node.translationY = translationY
-        node.transformOrigin = transformOrigin; node.cacheKey = cacheKey
-    }
-    override fun hashCode(): Int {
-        var v = alpha.hashCode()
-        v = 31 * v + scaleX.hashCode(); v = 31 * v + scaleY.hashCode()
-        v = 31 * v + rotationZ.hashCode()
-        v = 31 * v + translationX.hashCode(); v = 31 * v + translationY.hashCode()
-        v = 31 * v + transformOrigin.hashCode(); v = 31 * v + (cacheKey?.hashCode() ?: 0)
-        return v
-    }
-    override fun equals(other: Any?): Boolean =
-        other is GraphicsLayerModifier &&
-            other.alpha == alpha && other.scaleX == scaleX && other.scaleY == scaleY &&
-            other.rotationZ == rotationZ &&
-            other.translationX == translationX && other.translationY == translationY &&
-            other.transformOrigin == transformOrigin && other.cacheKey == cacheKey
-}
-
-class GraphicsLayerNode(
-    var alpha: Float,
-    var scaleX: Float,
-    var scaleY: Float,
-    var rotationZ: Float,
-    var translationX: Float,
-    var translationY: Float,
-    var transformOrigin: TransformOrigin,
-    var cacheKey: Any?,
-) : Modifier.Node()

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

@@ -0,0 +1,34 @@
+package com.compose.sdl.graphics
+
+// ==================
+// MARK: DrawStats — per-frame renderer work counters (diagnostic)
+// ==================
+
+/**
+ * Cheap per-frame counters the SDL draw path bumps and the frame profiler
+ * reads, to answer "what inside `draw` costs the time?" — geometry submissions
+ * (SDL_RenderGeometry calls), total vertices tessellated, rounded-clip mask
+ * realizations (offscreen render-target passes — the expensive ones), text
+ * blits and image blits. Reset each frame by the profiler after reading.
+ *
+ * Single-threaded (all draw is on the main thread), so plain vars are fine.
+ * Zero overhead when the profiler isn't reading — just integer increments.
+ */
+object DrawStats {
+	var geometrySubmits: Int = 0
+	var vertices: Int = 0
+	var maskRealizations: Int = 0
+	var textDraws: Int = 0
+	var imageBlits: Int = 0
+
+	fun reset() {
+		geometrySubmits = 0
+		vertices = 0
+		maskRealizations = 0
+		textDraws = 0
+		imageBlits = 0
+	}
+
+	fun summary(): String =
+		"geo=$geometrySubmits verts=$vertices masks=$maskRealizations text=$textDraws img=$imageBlits"
+}

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

@@ -259,6 +259,7 @@ internal class Sdl3Canvas(
 			fClip = vRegion
 			applyClip()
 			clearRegion(vRegion)
+			com.compose.sdl.graphics.DrawStats.maskRealizations++
 			fClipLayers.addLast(OffscreenClip(vTarget, vPrevTarget, vPending.prevClip, vRegion, vPending.bbox, vPending.deviceRound))
 			// Attribute the realized mask to the save-frame that pushed the
 			// clip: inner frames entered since must NOT composite it on their
@@ -963,6 +964,7 @@ internal class Sdl3Canvas(
 		// a 2px margin for glyph overhang / AA bleed, instead of paying a
 		// mask pass per text run.
 		admitDraw(inX - 2f, inY - 2f, inX + inBoxWidth + 2f, inY + inBoxHeight + 2f)
+		com.compose.sdl.graphics.DrawStats.textDraws++
 		fScope.flush()
 		// Paragraph-level decoration bits forwarded to every wrapped line.
 		val vUnderline = inTextDecoration?.contains(androidx.compose.ui.text.style.TextDecoration.Underline) == true
@@ -978,26 +980,25 @@ internal class Sdl3Canvas(
 		// node's real box. The per-line path below centres within a lineHeight band
 		// (1.2 em for Material Symbols), taller than the size-clamped icon node,
 		// which pushed every icon ~0.1 em below centre.
-		// Device-space scale for GLYPHS: the pen position already maps through
+		// Text is rasterised at its LOGICAL font size (stable glyph-texture cache
-		// the affine, but font size and centering boxes must scale with it too
+		// + stable hinting) and blit 1:1 at the affine-mapped pen position; the
-		// or text inside a graphicsLayer(scale) renders at full size, off
+		// layer's scale reaches position but NOT glyph size. Re-rasterising at a
-		// centre (JVM/Skia scales glyphs with the canvas). Wrap stays in LOCAL
+		// per-frame device size (an earlier attempt) made text under an animated
-		// units below so line breaks match what layout measured.
+		// layer scale — every Material popup/dropdown enter transition
-		val vTextScaleX = sqrt(fMa * fMa + fMb * fMb)
+		// (Menu.kt scaleX/scaleY) — step through integer sizes with visible
-		val vTextScaleY = sqrt(fMc * fMc + fMd * fMd)
+		// size/weight jumps, since SDL3_ttf re-hints at each size. Smoothly
-		val vDeviceFontPx =
+		// scaling glyphs would need blit-scaling a logical-size rasterisation
-			if (vTextScaleY == 1f) inFontSizePx
+		// (JVM/Skia GPU-scales the layer) — a future text-renderer enhancement;
-			else (inFontSizePx * vTextScaleY).toInt().coerceAtLeast(1)
+		// until then logical-size + no glyph scale is the stable choice.
-
 		if (inFontFamily != null && IconFont.isIconFamily(inFontFamily)) {
 			vTr.drawText(
 				inText = inText,
 				inX = mapX(inX, inY).toInt(),
 				inY = mapY(inX, inY).toInt(),
-				inBoxWidth = (inBoxWidth * vTextScaleX).toInt(),
+				inBoxWidth = inBoxWidth.toInt(),
-				inBoxHeight = (inBoxHeight * vTextScaleY).toInt(),
+				inBoxHeight = inBoxHeight.toInt(),
 				inColor = vColor,
-				inFontSize = vDeviceFontPx,
+				inFontSize = inFontSizePx,
 				inAlign = inTextAlign,
 				inFontFamily = inFontFamily,
 				inFontVariations = inFontVariations,
@@ -1040,10 +1041,10 @@ internal class Sdl3Canvas(
 				// scale with the layer. Rotation still only repositions.
 				inX = mapX(inX, vLineY).toInt(),
 				inY = mapY(inX, vLineY).toInt(),
-				inBoxWidth = (inBoxWidth * vTextScaleX).toInt(),
+				inBoxWidth = inBoxWidth.toInt(),
-				inBoxHeight = (vLineH * vTextScaleY).toInt(),
+				inBoxHeight = vLineH.toInt(),
 				inColor = vColor,
-				inFontSize = vDeviceFontPx,
+				inFontSize = inFontSizePx,
 				inAlign = inTextAlign,
 				inFontFamily = inFontFamily,
 				inFontVariations = inFontVariations,
@@ -1115,6 +1116,7 @@ internal class Sdl3Canvas(
 	) {
 		realizePendingClips()
 		val vTex = (image as? SdlImageBitmap)?.texture ?: return
+		com.compose.sdl.graphics.DrawStats.imageBlits++
 		// Commit pending frame geometry and re-assert this canvas's target + clip
 		// (an offscreen render just borrowed the render target).
 		fScope.flush()

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

@@ -141,6 +141,8 @@ internal class Sdl3DrawScope(
 			)
 		}
 		SDL_RenderGeometry(fRenderer.reinterpret(), null, fBatch, fBatchCount, null, 0)
+		com.compose.sdl.graphics.DrawStats.geometrySubmits++
+		com.compose.sdl.graphics.DrawStats.vertices += fBatchCount
 		fBatchCount = 0
 	}
 
@@ -243,7 +245,10 @@ internal class Sdl3DrawScope(
 		val vB = vT + size.height
 		when (style) {
 			Fill -> {
-				emitQuad(vL, vT, vR, vT, vR, vB, vL, vB, vSampler)
+				// Gradient fills need interior samples (see emitRectMesh); solid
+				// fills are a single flat quad.
+				if (brush is SolidColor) emitQuad(vL, vT, vR, vT, vR, vB, vL, vB, vSampler)
+				else emitRectMesh(vL, vT, vR, vB, vSampler)
 				// Axis-aligned rects are pixel-crisp and need no AA. When the affine
 				// rotates/shears them, the edges become diagonal — feather each edge
 				// outward (normals in local space; the matrix orients them on screen).
@@ -524,20 +529,30 @@ internal class Sdl3DrawScope(
 		val vY = fOriginY + topLeft.y
 		val vW = size.width
 		val vH = size.height
+		// Gradient fills sample per-vertex, so the interior must be meshed
+		// (corner-only sampling renders radial/sweep flat — see emitRectMesh);
+		// solid fills stay single quads.
+		val vGrad = brush !is SolidColor
 		if (vR <= 0f) {
-			// Trivial: just two triangles.
+			if (vGrad) emitRectMesh(vX, vY, vX + vW, vY + vH, vSampler)
-			emitQuad(vX, vY, vX + vW, vY, vX + vW, vY + vH, vX, vY + vH, vSampler)
+			else emitQuad(vX, vY, vX + vW, vY, vX + vW, vY + vH, vX, vY + vH, vSampler)
 			return
 		}
 		// Body in 3 strips: middle (full width × inner height), top edge,
 		// bottom edge — plus the 4 corner arcs.
 		if (style == Fill) {
-			// Middle strip
+			if (vGrad) {
-			emitQuad(vX, vY + vR, vX + vW, vY + vR, vX + vW, vY + vH - vR, vX, vY + vH - vR, vSampler)
+				emitRectMesh(vX, vY + vR, vX + vW, vY + vH - vR, vSampler)          // middle
-			// Top edge (between left+right corners)
+				emitRectMesh(vX + vR, vY, vX + vW - vR, vY + vR, vSampler)          // top edge
-			emitQuad(vX + vR, vY, vX + vW - vR, vY, vX + vW - vR, vY + vR, vX + vR, vY + vR, vSampler)
+				emitRectMesh(vX + vR, vY + vH - vR, vX + vW - vR, vY + vH, vSampler) // bottom edge
-			// Bottom edge
+			} else {
-			emitQuad(vX + vR, vY + vH - vR, vX + vW - vR, vY + vH - vR, vX + vW - vR, vY + vH, vX + vR, vY + vH, vSampler)
+				// Middle strip
+				emitQuad(vX, vY + vR, vX + vW, vY + vR, vX + vW, vY + vH - vR, vX, vY + vH - vR, vSampler)
+				// Top edge (between left+right corners)
+				emitQuad(vX + vR, vY, vX + vW - vR, vY, vX + vW - vR, vY + vR, vX + vR, vY + vR, vSampler)
+				// Bottom edge
+				emitQuad(vX + vR, vY + vH - vR, vX + vW - vR, vY + vH - vR, vX + vW - vR, vY + vH, vX + vR, vY + vH, vSampler)
+			}
 			// 4 corner fills — segment count adapts to the corner radius.
 			val vSeg = arcSegments(90f, vR)
 			emitFilledArc(vX + vR,          vY + vR,         vR, vR, 180f,  90f, false, vSeg, vSampler)
@@ -854,6 +869,34 @@ internal class Sdl3DrawScope(
 		emitTri(ax, ay, cx, cy, dx, dy, inSampler)
 	}
 
+	// Axis-aligned rect subdivided into a grid, so a per-vertex sampler is
+	// evaluated across the INTERIOR, not just the 4 corners. Corner-only
+	// sampling renders radial/sweep gradients flat (a square's corners are all
+	// equidistant from its centre → identical t → solid fill) and clips
+	// non-uniform multi-stop linear ramps. SDL interpolates linearly within
+	// each cell, so ~16 cells/axis tracks the gradient curve closely. Solid
+	// fills never come here (they use emitQuad — one quad).
+	private fun emitRectMesh(
+		inL: Float, inT: Float, inR: Float, inB: Float, inSampler: Sampler,
+	) {
+		// ~one cell per 8 device px per axis (1..24), so thin edge strips stay
+		// cheap while large fills get enough interior samples for the gradient.
+		val vCols = (((inR - inL) / 8f).toInt()).coerceIn(1, 24)
+		val vRows = (((inB - inT) / 8f).toInt()).coerceIn(1, 24)
+		val vW = (inR - inL) / vCols
+		val vH = (inB - inT) / vRows
+		for (vRow in 0 until vRows) {
+			val vY0 = inT + vRow * vH
+			val vY1 = if (vRow == vRows - 1) inB else vY0 + vH
+			for (vCol in 0 until vCols) {
+				val vX0 = inL + vCol * vW
+				val vX1 = if (vCol == vCols - 1) inR else vX0 + vW
+				emitTri(vX0, vY0, vX1, vY0, vX1, vY1, inSampler)
+				emitTri(vX0, vY0, vX1, vY1, vX0, vY1, inSampler)
+			}
+		}
+	}
+
 	private fun emitTri(
 		ax: Float, ay: Float, bx: Float, by: Float, cx: Float, cy: Float,
 		inSampler: Sampler,

demo/src/commonMain/kotlin/screens/GraphicsLayerScreen.kt

@@ -42,55 +42,9 @@ internal fun GraphicsLayerScreen() {
 	Column(verticalArrangement = Arrangement.spacedBy(16.dp)) {
 		ScreenTitle(
 			"Modifier.graphicsLayer",
-			"2D transform (scale / rotation / translation) + alpha + optional cacheKey opt-in. " +
+			"2D transform (scale / rotation / translation) + alpha, applied to a subtree.",
-				"When cacheKey is set, the renderer renders the subtree into an offscreen target once " +
-				"and reuses it while the key compares equal — change the key to invalidate.",
 		)
 
-		Section(
-			"Cached subtree (cacheKey)",
-			"Identical to a fresh-drawn row visually, but the subtree is rendered once and reused while " +
-				"cacheKey doesn't change. Useful for static screens — no per-frame redraw of the chrome.",
-		) {
-			Row(horizontalArrangement = Arrangement.spacedBy(16.dp)) {
-				for (vI in 0..3) {
-					Box(
-						modifier = Modifier
-							.size(64.dp)
-							.graphicsLayer(),
-					) {
-						Canvas(modifier = Modifier.size(64.dp)) {
-							drawCircle(color = vPrimary, radius = 22f, center = Offset(32f, 32f))
-							drawCircle(color = vSecondary, radius = 12f, center = Offset(32f, 32f))
-						}
-					}
-				}
-				Text("(cached)", color = vPrimary, fontSize = 12.sp)
-			}
-		}
-
-		Section("Combined (cache + rotation)", "Cached subtree blitted with live rotation around its centre.") {
-			Row(horizontalArrangement = Arrangement.spacedBy(16.dp)) {
-				for (vI in 0..4) {
-					Box(
-						modifier = Modifier
-							.size(64.dp)
-							.graphicsLayer(
-								rotationZ = vSpin + vI * 30f,
-							),
-					) {
-						Canvas(modifier = Modifier.size(64.dp)) {
-							drawRect(
-								color = if (vI % 2 == 0) vPrimary else vSecondary,
-								topLeft = Offset(10f, 10f),
-								size = Size(44f, 44f),
-							)
-						}
-					}
-				}
-			}
-		}
-
 		Section("Rotation (live)", "Each tile holds the same shape rotated at multiples of the current spin angle.") {
 			Row(horizontalArrangement = Arrangement.spacedBy(16.dp)) {
 				for (vI in 0..4) {

scripts/probe/README.md

@@ -0,0 +1,35 @@
+# Probe driver
+
+Launches a native compose-desktop-native app window, sends **window-relative**
+synthetic input, and captures the client area — for reproducing visual bugs
+deterministically (this is the packaged form of the rigs that caught the
+square-on-click and TLS-chain regressions).
+
+Why window-relative + PrintWindow: input coordinates are FRACTIONS of the
+window client rect and addressed by process name, so a probe doesn't depend on
+where the window landed or on it being foreground; capture uses `PrintWindow`,
+which grabs the window even when occluded or unfocused (a plain screen grab
+would catch whatever's on top).
+
+```bash
+# screenshot a demo screen
+python scripts/probe/probe.py demo --screen=Images --shot images.png
+
+# press-and-hold a point (fraction of the window) and capture DURING the press
+python scripts/probe/probe.py demo --screen=Buttons --hold 0.1,0.32 --shot press.png
+
+# hover then capture
+python scripts/probe/probe.py demo --screen=Shapes --hover 0.3,0.45 --shot hov.png
+
+# an app built elsewhere (e.g. the bridge example)
+python scripts/probe/probe.py shared --exe /path/to/shared.exe --click 0.5,0.5 --shot x.png
+```
+
+- `proc` = the window's process name (`demo`, `apidemo`, `shared`).
+- Actions run in order; `--hold` captures mid-press (async), `--click` /
+  `--hover` complete first. `--shot` captures after `--settle` seconds.
+- Windows-only (uses Win32 input + PrintWindow). Pillow only needed if you
+  post-process the PNG.
+
+Pairs with `scripts/parity/` (whole-screen diffing); this is for targeted
+interaction repro.

scripts/probe/_capture.ps1

@@ -0,0 +1,28 @@
+param([string]$proc, [string]$out)
+# Capture a window's CLIENT area by process name, via PrintWindow — works even
+# when the window is occluded or not focused (unlike a screen grab). Used by the
+# probe driver to screenshot native app windows deterministically.
+Add-Type -AssemblyName System.Drawing
+Add-Type @"
+using System;
+using System.Runtime.InteropServices;
+public class Cap {
+    [DllImport("user32.dll")] public static extern bool GetClientRect(IntPtr h, out RECT r);
+    [DllImport("user32.dll")] public static extern bool PrintWindow(IntPtr h, IntPtr dc, uint flags);
+    public struct RECT { public int L, T, R, B; }
+}
+"@
+$h = (Get-Process $proc -ErrorAction SilentlyContinue | Where-Object { $_.MainWindowHandle -ne 0 } | Select-Object -First 1).MainWindowHandle
+if (-not $h -or $h -eq [IntPtr]::Zero) { Write-Output "NO WINDOW"; exit 1 }
+$r = New-Object Cap+RECT
+[Cap]::GetClientRect($h, [ref]$r) | Out-Null
+$w = $r.R - $r.L; $ht = $r.B - $r.T
+if ($w -le 0 -or $ht -le 0) { Write-Output "ZERO SIZE"; exit 1 }
+$bmp = New-Object System.Drawing.Bitmap $w, $ht
+$g = [System.Drawing.Graphics]::FromImage($bmp)
+$hdc = $g.GetHdc()
+# flags=3 : PW_CLIENTONLY | PW_RENDERFULLCONTENT (captures accelerated content)
+[Cap]::PrintWindow($h, $hdc, 3) | Out-Null
+$g.ReleaseHdc($hdc)
+$bmp.Save($out, [System.Drawing.Imaging.ImageFormat]::Png)
+Write-Output "captured ${w}x${ht}"

scripts/probe/_input.ps1

@@ -0,0 +1,42 @@
+param(
+    [string]$proc,          # process name of the target window
+    [string]$action,        # click | hold | hover | move
+    [double]$fx,            # fractional X within the window client area (0..1)
+    [double]$fy,            # fractional Y
+    [int]$holdMs = 80       # press duration for 'hold'
+)
+# Window-CLIENT-relative synthetic input against a native app window, addressed
+# by process name — so probes don't depend on where the window landed or on it
+# being the foreground window (it's raised to foreground first). Coordinates are
+# FRACTIONS of the client rect, so they're resolution/DPI independent.
+Add-Type @"
+using System;
+using System.Runtime.InteropServices;
+public class In {
+    [DllImport("user32.dll")] public static extern bool GetClientRect(IntPtr h, out RECT r);
+    [DllImport("user32.dll")] public static extern bool ClientToScreen(IntPtr h, ref POINT p);
+    [DllImport("user32.dll")] public static extern bool SetForegroundWindow(IntPtr h);
+    [DllImport("user32.dll")] public static extern bool SetCursorPos(int x, int y);
+    [DllImport("user32.dll")] public static extern void mouse_event(uint f, uint x, uint y, uint d, UIntPtr e);
+    public struct RECT { public int L, T, R, B; }
+    public struct POINT { public int X, Y; }
+    public const uint DOWN = 2, UP = 4;
+}
+"@
+$h = (Get-Process $proc -ErrorAction SilentlyContinue | Where-Object { $_.MainWindowHandle -ne 0 } | Select-Object -First 1).MainWindowHandle
+if (-not $h -or $h -eq [IntPtr]::Zero) { Write-Output "NO WINDOW"; exit 1 }
+$r = New-Object In+RECT
+[In]::GetClientRect($h, [ref]$r) | Out-Null
+$p = New-Object In+POINT
+$p.X = [int](($r.R - $r.L) * $fx); $p.Y = [int](($r.B - $r.T) * $fy)
+[In]::ClientToScreen($h, [ref]$p) | Out-Null
+[In]::SetForegroundWindow($h) | Out-Null
+Start-Sleep -Milliseconds 120
+[In]::SetCursorPos($p.X, $p.Y) | Out-Null
+switch ($action) {
+    "move"  { }
+    "hover" { Start-Sleep -Milliseconds 250 }
+    "click" { Start-Sleep -Milliseconds 120; [In]::mouse_event([In]::DOWN,0,0,0,[UIntPtr]::Zero); Start-Sleep -Milliseconds 60; [In]::mouse_event([In]::UP,0,0,0,[UIntPtr]::Zero) }
+    "hold"  { Start-Sleep -Milliseconds 120; [In]::mouse_event([In]::DOWN,0,0,0,[UIntPtr]::Zero); Start-Sleep -Milliseconds $holdMs; [In]::mouse_event([In]::UP,0,0,0,[UIntPtr]::Zero) }
+}
+Write-Output "$action at $($p.X),$($p.Y)"

scripts/probe/probe.py

@@ -0,0 +1,101 @@
+#!/usr/bin/env python3
+"""
+Drive a native compose-desktop-native app window for manual/automated visual
+checks: launch it, send window-relative input (click / hover / hold / move),
+and capture the client area via PrintWindow (works even when occluded).
+
+Built from the ad-hoc rigs used to reproduce the square-on-click and TLS-chain
+bugs — window-CLIENT-relative and process-addressed, so it doesn't care where
+the window lands or whether it's focused.
+
+Examples:
+    # boot the demo's Images screen, screenshot after 4s
+    python scripts/probe/probe.py demo --screen=Images --shot=images.png
+
+    # boot bubble-wrap (example repo), press-and-hold a bubble mid-sheet, capture
+    python scripts/probe/probe.py shared --exe <path-to>/shared.exe \\
+        --hold 0.55,0.55 --shot press.png
+
+    # hover then capture
+    python scripts/probe/probe.py demo --screen=Shapes --hover 0.3,0.4 --shot hov.png
+
+`proc` is the window's process name (demo / apidemo / shared). Native apps
+built here use the mingwX64 debug exe by default; override with --exe. Actions
+run in order; --shot captures after them. Windows-only. Needs Pillow only if
+you pass --crop.
+"""
+import argparse, subprocess, sys, time, os, signal
+from pathlib import Path
+
+REPO = Path(__file__).resolve().parents[2]
+HERE = Path(__file__).resolve().parent
+DEFAULT_EXE = {
+    "demo": REPO / "demo/build/bin/mingwX64/debugExecutable/demo.exe",
+    "apidemo": REPO / "apidemo/build/bin/mingwX64/debugExecutable/apidemo.exe",
+}
+
+
+def ps(script: str, *args: str):
+    return subprocess.run(
+        ["powershell", "-ExecutionPolicy", "Bypass", "-File", str(HERE / script), *args],
+        capture_output=True, text=True,
+    )
+
+
+def frac(s: str):
+    x, y = s.split(",")
+    return float(x), float(y)
+
+
+def main():
+    ap = argparse.ArgumentParser()
+    ap.add_argument("proc", help="window process name (demo / apidemo / shared)")
+    ap.add_argument("--exe", help="path to the executable (default: mingwX64 debug for demo/apidemo)")
+    ap.add_argument("--screen", help="pass --screen=<Name> to the app (demo)")
+    ap.add_argument("--args", default="", help="extra args passed verbatim to the exe")
+    ap.add_argument("--boot", type=float, default=4.0, help="seconds to wait after launch")
+    ap.add_argument("--click", help="fractional X,Y to click (e.g. 0.5,0.5)")
+    ap.add_argument("--hold", help="fractional X,Y to press-and-hold")
+    ap.add_argument("--hover", help="fractional X,Y to hover")
+    ap.add_argument("--hold-ms", type=int, default=900, help="hold duration")
+    ap.add_argument("--settle", type=float, default=0.3, help="seconds between action and capture")
+    ap.add_argument("--shot", help="capture the window to this PNG after the actions")
+    args = ap.parse_args()
+
+    exe = Path(args.exe) if args.exe else DEFAULT_EXE.get(args.proc)
+    if not exe or not exe.exists():
+        print(f"exe not found: {exe} (build it, or pass --exe)", file=sys.stderr)
+        return 1
+
+    cmd = [str(exe)]
+    if args.screen:
+        cmd.append(f"--screen={args.screen}")
+    if args.args:
+        cmd += args.args.split()
+
+    proc = subprocess.Popen(cmd)
+    try:
+        time.sleep(args.boot)
+        if args.hover:
+            fx, fy = frac(args.hover)
+            print(ps("_input.ps1", "-proc", args.proc, "-action", "hover", "-fx", str(fx), "-fy", str(fy)).stdout.strip())
+        if args.click:
+            fx, fy = frac(args.click)
+            print(ps("_input.ps1", "-proc", args.proc, "-action", "click", "-fx", str(fx), "-fy", str(fy)).stdout.strip())
+        if args.hold:
+            fx, fy = frac(args.hold)
+            # hold runs async so we can capture DURING the press
+            p = subprocess.Popen(["powershell", "-ExecutionPolicy", "Bypass", "-File", str(HERE / "_input.ps1"),
+                                   "-proc", args.proc, "-action", "hold", "-fx", str(fx), "-fy", str(fy), "-holdMs", str(args.hold_ms)])
+        time.sleep(args.settle)
+        if args.shot:
+            out = str(Path(args.shot).resolve())
+            r = ps("_capture.ps1", "-proc", args.proc, "-out", out)
+            print(f"{r.stdout.strip()} -> {out}")
+    finally:
+        subprocess.run(["taskkill", "/F", "/IM", exe.name], capture_output=True)
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())