diff --git a/CLAUDE.md b/CLAUDE.md
index 9446fd9a..d833a1ca 100644
--- a/CLAUDE.md
+++ b/CLAUDE.md
@@ -108,6 +108,14 @@ The state module also contains protocol handler implementations (`*_handler.rs`
 
 Located in `src/screenshare/`. See [docs/developer/screenshare.md](./docs/developer/screenshare.md) for detailed architecture.
 
+### Screen Capture (wlr-screencopy)
+
+`src/state/screencopy.rs` implements `zwlr_screencopy_manager_v1` for tools like `grim` and `wl-mirror`. Reads pixels from the Skia surface after render into SHM buffers. See [docs/developer/screencopy.md](./docs/developer/screencopy.md).
+
+### Virtual Pointer
+
+`src/state/virtual_pointer.rs` implements `wlr-virtual-pointer-unstable-v1` for automation tools (`wlrctl`, `wtype`). Together with the existing `virtual-keyboard` and `wlr-foreign-toplevel` protocols, enables full compositor remote control. See [docs/developer/virtual-pointer.md](./docs/developer/virtual-pointer.md).
+
 ## Coordinate Systems & Naming Conventions
 
 Otto has two coordinate spaces — mixing them causes subtle scale-dependent bugs.
@@ -157,6 +165,6 @@ Two tiers:
 - `docs/user/` — End-user configuration and usage guides
 - `docs/developer/` — Architecture, design docs, and implementation details
 
-Key developer docs: `rendering.md`, `render_loop.md`, `wayland.md`, `screenshare.md`, `expose.md`, `dock-design.md`, `sc-layer-protocol-design.md`.
+Key developer docs: `rendering.md`, `render_loop.md`, `rendering-optimizations.md`, `wayland.md`, `screenshare.md`, `screencopy.md`, `virtual-pointer.md`, `expose.md`, `dock-design.md`, `testing.md`, `sc-layer-protocol-design.md`.
 
 Review and documentation guidelines: `.github/instructions/review.instructions.md`, `.github/instructions/documentation.instructions.md`.
diff --git a/docs/developer/rendering-optimizations.md b/docs/developer/rendering-optimizations.md
new file mode 100644
index 00000000..46da36bc
--- /dev/null
+++ b/docs/developer/rendering-optimizations.md
@@ -0,0 +1,62 @@
+# Rendering Optimizations
+
+This document covers Otto's rendering performance strategies beyond the core pipeline described in [rendering.md](./rendering.md) and [render_loop.md](./render_loop.md).
+
+## Per-window frame callback throttling
+
+**Source:** `src/state/window_throttle.rs`
+
+The single biggest lever for reducing GPU work is stopping frame callbacks to windows the user can't see. Well-behaved Wayland clients (Chromium, GTK4, Qt6) pause their internal render loops when frame callbacks stop arriving, which saves both compositor-side and client-side GPU work.
+
+### Window states
+
+Every mapped window is classified each frame into one of five states:
+
+| State | Rate | xdg activated | When |
+|-------|------|---------------|------|
+| **Focused** | Full refresh | yes | Top-of-stack on current workspace, fullscreen, or expose active |
+| **Secondary** | ~30 Hz | no | On current workspace, visible, not focused |
+| **Occluded** | ~2 Hz | no | On current workspace, fully covered by opaque content above |
+| **Minimized** | ~2 Hz | no | Explicitly minimized by the user |
+| **HiddenWorkspace** | ~2 Hz | no | Window's workspace is not active on any output |
+
+### Why 2 Hz instead of zero?
+
+Hidden states deliberately trickle callbacks at 2 Hz rather than stopping entirely. Chromium 115+ has an eviction heuristic that discards content buffers when callbacks stop for too long, causing a blank-canvas-on-restore bug. The 2 Hz rate satisfies the heuristic while saving essentially all the work.
+
+### Classification logic
+
+`classify_one()` is a pure function — no Wayland or lay-rs state, easy to unit test:
+
+```
+is_minimized?          → Minimized
+expose_active?         → Focused (all windows get smooth previews)
+is_fullscreen_window?  → Focused
+fullscreen_exists?     → Occluded (behind the fullscreen)
+is_top_of_stack?       → Focused
+in occluded_ids set?   → Occluded
+otherwise              → Secondary
+```
+
+The `occluded_ids` set is computed by the lay-rs occlusion walk (when available). Currently empty — populated as a future refinement.
+
+### Integration with the render loop
+
+`classify_windows()` runs once per frame and produces a `HashMap<ObjectId, WindowThrottleState>`. The render loop passes each window's throttle duration to Smithay's `Window::send_frame()`, which skips the callback if insufficient time has elapsed.
+
+The `is_activated` flag is sent via `xdg_toplevel.configure`, signaling toolkits to self-throttle on top of the compositor's frame-callback throttling.
+
+## Damage tracking
+
+Otto uses Smithay's `OutputDamageTracker` to render only damaged regions. The render loop skips frame submission entirely when all of these are false:
+
+- `scene_has_damage` — lay-rs scene graph reports changes
+- `dnd_needs_draw` — drag-and-drop icon is active
+- `cursor_needs_draw` — pointer is in the output
+- `has_screencopy` — a screencopy client is waiting for a frame
+
+This means an idle desktop with no animations and no cursor movement submits zero frames.
+
+## Screencopy render forcing
+
+When a screencopy client has a pending frame (`pending_screencopy_frames` is non-empty), `should_draw` is forced true regardless of damage state. This ensures capture tools always get a fresh frame. See [screencopy.md](./screencopy.md) for details.
diff --git a/docs/developer/screencopy.md b/docs/developer/screencopy.md
new file mode 100644
index 00000000..c2403546
--- /dev/null
+++ b/docs/developer/screencopy.md
@@ -0,0 +1,67 @@
+# wlr-screencopy-v1
+
+Otto implements `zwlr_screencopy_manager_v1` (version 2), allowing screen capture tools to grab output frames via shared-memory buffers.
+
+## Usage
+
+```bash
+# Full-screen screenshot (saves to file)
+grim screenshot.png
+
+# Region capture
+grim -g "100,100 800x600" region.png
+
+# Live mirroring
+wl-mirror eDP-1
+```
+
+Any client speaking wlr-screencopy-v1 works — `grim`, `wl-mirror`, `wlr-randr` (for output info), `wf-recorder`, etc.
+
+## Protocol flow
+
+```
+Client                          Otto
+  │                               │
+  ├─ capture_output(output) ─────►│
+  │                               ├─ create frame object
+  │◄── buffer(ARGB8888, w, h, s) ─┤  (advertise SHM format)
+  │◄── buffer_done ───────────────┤  (v3+)
+  │                               │
+  ├─ copy(wl_buffer) ────────────►│
+  │                               ├─ queue as PendingScreencopy
+  │                               ├─ force render (even if no damage)
+  │                               ├─ read_pixels from Skia surface
+  │◄── flags(empty) ──────────────┤
+  │◄── ready(tv_sec, tv_nsec) ────┤
+  │                               │
+  ├─ destroy ────────────────────►│
+```
+
+`capture_output_region` follows the same flow but scales the requested logical region to physical pixels using the output's fractional scale.
+
+## Architecture
+
+### Key types (`src/state/screencopy.rs`)
+
+| Type | Role |
+|------|------|
+| `ScreencopyManagerState` | Holds the Wayland global; created in `Otto::init` |
+| `ScreencopyFrameData` | Per-frame metadata: output, region, dimensions, state machine |
+| `PendingScreencopy` | Queued frame + client buffer, waiting to be filled during render |
+
+### Render integration (`src/udev/render.rs`)
+
+1. When `pending_screencopy_frames` is non-empty, `should_draw` is forced true — the compositor renders even if the scene has no damage.
+2. After `render_frame` completes and the output is scanned out, `complete_screencopy_for_output` is called.
+3. It uses `skia_surface.read_pixels()` to copy the rendered frame into each pending SHM buffer, then sends `ready` or `failed`.
+
+### Buffer format
+
+Only `ARGB8888` (4 bytes/pixel) is advertised. Stride is `width * 4`. The Skia read uses `BGRA8888` color type which matches ARGB8888 on little-endian (the Wayland convention).
+
+## Limitations
+
+- **SHM only** — no DMA-BUF export. Every frame does a GPU-to-CPU readback via `read_pixels`. Fine for screenshots; a streaming client like `wl-mirror` will consume more bandwidth than a zero-copy DMA-BUF path would.
+- **No damage reporting** — `copy_with_damage` is accepted but damage regions are not reported to the client.
+- **Cursor overlay** — the `overlay_cursor` flag is stored but the cursor is always composited into the frame (same as overlay_cursor=1).
+- **Udev backend only** — the `complete_screencopy_for_output` call site is in `src/udev/render.rs`. Winit/X11 backends don't call it yet.
diff --git a/docs/developer/testing.md b/docs/developer/testing.md
new file mode 100644
index 00000000..f342c224
--- /dev/null
+++ b/docs/developer/testing.md
@@ -0,0 +1,155 @@
+# Testing
+
+Otto has two test systems: headless integration tests for compositor behavior and WLCS for Wayland protocol conformance.
+
+## Headless integration tests
+
+Tests compositor behavior (gestures, expose, workspaces, window lifecycle, pointer interactions, scene graph) using a real compositor instance without GPU.
+
+```bash
+cargo test --features headless --test headless_basic
+```
+
+### Architecture
+
+- **Backend:** `src/headless.rs` — `HeadlessHandle` starts the compositor on a background thread
+- **Test client:** `components/otto-kit/src/testing.rs` — lightweight Wayland client with SHM buffers
+- **Tests:** `tests/headless_basic.rs`
+- **CI:** runs in the `check` job via `cargo test --features headless --test headless_basic`
+
+### HeadlessHandle API
+
+Start a compositor and interact with it:
+
+```rust
+let handle = HeadlessHandle::start(HeadlessConfig::default());
+```
+
+#### State access
+
+| Method | Returns | Description |
+|--------|---------|-------------|
+| `with_state(\|state\| { ... })` | — | Run a closure on the compositor thread (blocking) |
+| `query(\|state\| value)` | `R` | Query a value from compositor state |
+| `window_count()` | `usize` | Number of windows across all workspaces |
+| `current_workspace_index()` | `usize` | Active workspace index |
+| `is_expose_active()` | `bool` | Whether expose mode is open |
+| `is_show_desktop_active()` | `bool` | Whether show-desktop mode is active |
+
+#### Gesture simulation
+
+| Method | Description |
+|--------|-------------|
+| `swipe_begin()` | Start a 3-finger swipe gesture |
+| `swipe_update(dx, dy)` | Send swipe delta (pixels) |
+| `swipe_end()` | End swipe gesture |
+| `swipe(&[(dx, dy)])` | Complete swipe in one call |
+| `pinch_begin()` | Start a 4-finger pinch |
+| `pinch_update(scale)` | Send pinch scale |
+| `pinch_end()` | End pinch gesture |
+
+#### Pointer simulation
+
+| Method | Description |
+|--------|-------------|
+| `pointer_move(x, y)` | Move pointer to absolute logical coordinates |
+| `pointer_click()` | Left-button press + release at current position |
+
+**Note:** The first `pointer_move` into a new focus area triggers smithay's `enter` event (not `motion`). Selection updates only happen on `motion`. Establish focus with a priming move before testing hover behavior:
+
+```rust
+// Prime focus on the window selector area
+handle.pointer_move(5.0, 300.0);
+handle.settle(2);
+// Now this move triggers motion → selection update
+handle.pointer_move(target_x, target_y);
+handle.settle(10);
+```
+
+#### Expose queries
+
+| Method | Returns | Description |
+|--------|---------|-------------|
+| `expose_window_rects()` | `Vec<(title, x, y, w, h)>` | Expose layout rects in physical pixels |
+| `expose_selected_title()` | `Option<String>` | Currently hovered window title in expose |
+
+To convert expose rects (physical) to pointer coordinates (logical), divide by the output scale:
+
+```rust
+let scale = handle.query(|state| {
+    state.workspaces.outputs().next()
+        .map(|o| o.current_scale().fractional_scale())
+        .unwrap_or(1.0)
+});
+let cx = (rect_x + rect_w / 2.0) as f64 / scale;
+let cy = (rect_y + rect_h / 2.0) as f64 / scale;
+```
+
+#### Scene graph
+
+| Method | Returns | Description |
+|--------|---------|-------------|
+| `scene_snapshot()` | `SceneSnapshot` | Full scene graph snapshot |
+| `scene_json()` | `String` | Scene graph as JSON |
+| `scene_has_damage()` | `bool` | Whether scene has pending damage |
+| `is_layer_hidden(key)` | `Option<bool>` | Hidden state of a named layer |
+
+#### Animation control
+
+| Method | Description |
+|--------|-------------|
+| `settle(max_frames)` | Advance scene at 60fps until animations finish or limit reached. Returns frames with damage. |
+| `tick(dt)` | Advance one frame by `dt` seconds. Returns true if damage produced. |
+| `wait(duration)` | Sleep the test thread (lets compositor event loop run). |
+
+`settle` is deterministic — no wall-clock sleeps. Use it after gestures/pointer events to let animations complete.
+
+### TestClient API
+
+Connect to the compositor and create windows:
+
+```rust
+let mut client = TestClient::connect(&handle.socket_name)?;
+let toplevel = client.create_toplevel("my-window", 640, 480);
+handle.wait(Duration::from_millis(200));
+let _ = client.roundtrip();
+```
+
+`create_toplevel` creates a toplevel surface with a SHM buffer, sets the title, and commits. The returned `Arc<Mutex<TestToplevel>>` tracks configure events.
+
+### Writing tests
+
+Tests must use `#[serial]` (from `serial_test` crate) since they share a global compositor:
+
+```rust
+#[test]
+#[serial]
+fn my_test() {
+    let handle = start_compositor();
+    let mut client = connect_client(&handle);
+    // ... test logic ...
+}
+```
+
+## WLCS protocol conformance
+
+Tests Wayland protocol compliance (surface roles, pointer/touch routing, xdg_shell).
+
+```bash
+# One-time: build the WLCS test runner
+./compile_wlcs.sh
+
+# Build the Otto WLCS adapter
+cargo build -p wlcs_otto
+
+# Run specific test groups
+./wlcs/wlcs target/debug/libwlcs_otto.so --gtest_filter='SelfTest*:FrameSubmission*'
+./wlcs/wlcs target/debug/libwlcs_otto.so --gtest_filter='*/SurfacePointerMotionTest*'
+./wlcs/wlcs target/debug/libwlcs_otto.so --gtest_filter='XdgToplevelStableTest.*'
+
+# List all tests
+./wlcs/wlcs target/debug/libwlcs_otto.so --gtest_list_tests
+```
+
+- **Adapter:** `wlcs_otto/` — cdylib that WLCS loads, runs a headless Otto instance
+- **Key files:** `wlcs_otto/src/main_loop.rs` (event handling), `wlcs_otto/src/ffi_wrappers.rs` (C FFI)
diff --git a/docs/developer/virtual-pointer.md b/docs/developer/virtual-pointer.md
new file mode 100644
index 00000000..b197ec1f
--- /dev/null
+++ b/docs/developer/virtual-pointer.md
@@ -0,0 +1,75 @@
+# wlr-virtual-pointer-unstable-v1
+
+Otto implements `zwlr_virtual_pointer_manager_v1`, allowing external tools to synthesize pointer events that feed into the compositor's input pipeline as if they came from a real pointing device.
+
+## Usage
+
+```bash
+# Move pointer to absolute position and click
+wlrctl pointer move 500 300
+wlrctl pointer click left
+
+# Type text (uses the existing virtual-keyboard protocol)
+wlrctl keyboard type "hello"
+
+# Focus a specific window
+wlrctl toplevel focus firefox
+```
+
+Compatible clients: `wlrctl`, `ydotool` (Wayland mode), `wtype`, and any custom automation driver.
+
+### The automation trio
+
+Otto exposes three protocols that together enable full remote control:
+
+| Protocol | Purpose | Implementation |
+|----------|---------|----------------|
+| `virtual-keyboard-unstable-v1` | Synthesize key events | Smithay delegate |
+| `wlr-virtual-pointer-unstable-v1` | Synthesize pointer events | Hand-rolled (`src/state/virtual_pointer.rs`) |
+| `wlr-foreign-toplevel-management-unstable-v1` | Enumerate/focus/close windows | Smithay delegate |
+
+## Supported events
+
+All events are accumulated per-frame and flushed on `frame`:
+
+| Request | Behavior |
+|---------|----------|
+| `motion` | Relative displacement added to current pointer position |
+| `motion_absolute` | Normalized `[0, 1]` coordinates mapped to the first output's geometry |
+| `button` | Left/right/middle button press/release (standard Linux button codes) |
+| `axis` | Scroll amount on vertical/horizontal axis |
+| `axis_source` | Sets the source (wheel, finger, etc.) on the pending `AxisFrame` |
+| `axis_stop` | Stop notification for an axis (e.g. finger lifted from touchpad) |
+| `axis_discrete` | Discrete scroll step (v120 high-resolution) |
+| `frame` | Flushes all accumulated events to the pointer |
+
+## Architecture (`src/state/virtual_pointer.rs`)
+
+Smithay does not ship a virtual-pointer delegate, so the `GlobalDispatch`/`Dispatch` plumbing is hand-rolled.
+
+### Per-pointer state
+
+Each `ZwlrVirtualPointerV1` resource holds a `Mutex<VirtualPointerState>` with:
+- `pending_motion` — accumulated `(dx, dy)` from `motion` requests
+- `pending_absolute` — last `motion_absolute` position (overrides relative motion)
+- `pending_buttons` — queued `(button_code, ButtonState)` pairs
+- `axis_frame` — Smithay `AxisFrame` being built up across axis/axis_source/axis_stop/axis_discrete
+
+### Frame flush
+
+On `frame`, the accumulated state is committed in order:
+
+1. **Motion** — either absolute (mapped to output geometry) or relative (added to current location). Updates `pointer.motion()`, `layers_engine.pointer_move()`, and `surface_under()` focus.
+2. **Buttons** — each queued button dispatched via `pointer.button()`. On press, `focus_window_under_cursor` is called so clicks behave like real ones (raise + focus).
+3. **Axis** — the built-up `AxisFrame` sent via `pointer.axis()`.
+4. **Pointer frame** — `pointer.frame()` finalizes the sequence.
+
+### Click-to-focus
+
+Real libinput clicks run `focus_window_under_cursor` + `layers_engine.pointer_button_down/up` from `on_pointer_button`. The virtual pointer mirrors this in its frame flush so that `motion_absolute` + `button` correctly focuses and raises the target window.
+
+## Limitations
+
+- **No pointer constraints** — locked/confined pointer regions are not honored for virtual events.
+- **No relative-motion reporting** — the `wp_relative_pointer` extension is not notified for synthesized motion.
+- **Single seat** — events are injected into the default seat. The `seat` parameter in `create_virtual_pointer` is accepted but ignored.