diff --git a/Cargo.lock b/Cargo.lock
index 43346228..6b7e8eb9 100644
--- a/Cargo.lock
+++ b/Cargo.lock
@@ -3418,6 +3418,7 @@ dependencies = [
  "photocraft-doc",
  "photocraft-geom",
  "photocraft-raster",
+ "photocraft-vector",
  "pollster 1.0.1",
  "rayon",
  "wgpu",
diff --git a/crates/compose/src/effects.rs b/crates/compose/src/effects.rs
index 8e16d106..c691ddaf 100644
--- a/crates/compose/src/effects.rs
+++ b/crates/compose/src/effects.rs
@@ -8,9 +8,14 @@
 //! interior effects blend with the layer's mode. Layer opacity applies to
 //! the whole stack; fill opacity only to the layer's own pixels.
 //!
-//! Shapes come from the layer's alpha (after its mask). Distances use an
-//! exact Euclidean distance transform; soft falloffs use three box-blur
-//! passes (≈ Gaussian) whose total support equals the effect size.
+//! Interior effects are painted relative to the layer's shape and then take the shape's alpha
+//! (an overlay recolours a half-transparent edge without adding coverage). The outside parts
+//! of strokes blend onto the backdrop with their own modes, an upper stroke covering the
+//! lower ones. Drop shadows are knocked out only through see-through fill.
+//!
+//! Shapes come from the layer's alpha (after its mask). Strokes measure distances with a
+//! 5 × 5 chamfer metric seeded at sub-pixel edge offsets (as Photoshop does); other effects
+//! use an exact Euclidean distance transform. Soft falloffs are Gaussian blurs.
 
 use photocraft_color::blend::BlendMode;
 use photocraft_doc::{
@@ -177,6 +182,7 @@ fn edt_nearest(inside: &[bool], w: usize, h: usize) -> (Vec<f32>, Vec<usize>) {
 /// plus the index of the nearest `inside` pixel (0 on inside pixels). Photoshop's effect
 /// distances follow this metric: its stroke corners measure 3.236 = √5 + 1 at offset (1, 3)
 /// and 3.650 = √5 + √2 at (2, 3), where the exact distances are 3.162 and 3.606.
+#[cfg(test)]
 fn chamfer_nearest(inside: &[bool], w: usize, h: usize) -> (Vec<f32>, Vec<usize>) {
     chamfer_from(inside.iter().map(|&b| if b { 0.0 } else { CHAMFER_INF }).collect(), w, h)
 }
@@ -193,7 +199,7 @@ fn chamfer_from(mut d: Vec<f32>, w: usize, h: usize) -> (Vec<f32>, Vec<usize>) {
     const FWD: [(i64, i64, f32); 8] = [(-1, 0, 1.0), (-1, -1, S2), (0, -1, 1.0), (1, -1, S2), (-1, -2, S5), (1, -2, S5), (-2, -1, S5), (2, -1, S5)];
     let mut near: Vec<usize> = (0..w * h).collect();
     let (wi, hi) = (w as i64, h as i64);
-    let mut relax = |x: i64, y: i64, sign: i64, d: &mut Vec<f32>, near: &mut Vec<usize>| {
+    let relax = |x: i64, y: i64, sign: i64, d: &mut Vec<f32>, near: &mut Vec<usize>| {
         let i = (y * wi + x) as usize;
         let mut best = d[i];
         let mut bn = near[i];
@@ -242,13 +248,6 @@ enum Metric {
     Chamfer,
 }
 
-fn nearest(inside: &[bool], w: usize, h: usize, m: Metric) -> (Vec<f32>, Vec<usize>) {
-    match m {
-        Metric::Euclidean => edt_nearest(inside, w, h),
-        Metric::Chamfer => chamfer_nearest(inside, w, h),
-    }
-}
-
 fn dist_outside_by(s: &Map, m: Metric) -> Vec<f32> {
     if m == Metric::Chamfer {
         // Seeds start at their sub-pixel edge offset (1 − coverage).
@@ -257,7 +256,7 @@ fn dist_outside_by(s: &Map, m: Metric) -> Vec<f32> {
         return d.iter().zip(&s.v).map(|(d, &a)| if a > INSIDE_EPS { -0.5 } else { d - 0.5 }).collect();
     }
     let inside: Vec<bool> = s.v.iter().map(|&a| a > INSIDE_EPS).collect();
-    let (d, near) = nearest(&inside, s.w, s.h, m);
+    let (d, near) = edt_nearest(&inside, s.w, s.h);
     (0..d.len()).map(|i| if inside[i] { -0.5 } else { d[i] - 0.5 + (1.0 - s.v[near[i]].min(1.0)) }).collect()
 }
 
@@ -290,7 +289,7 @@ fn dist_inside_by(s: &Map, m: Metric) -> Vec<f32> {
         return d.iter().zip(&s.v).map(|(d, &a)| if a <= INSIDE_EPS { -0.5 } else { d - 0.5 }).collect();
     }
     let outside: Vec<bool> = s.v.iter().map(|&a| a <= INSIDE_EPS).collect();
-    let (d, _) = nearest(&outside, s.w, s.h, m);
+    let (d, _) = edt_nearest(&outside, s.w, s.h);
     d.iter().zip(&outside).map(|(d, o)| if *o { -0.5 } else { d - 0.5 }).collect()
 }
 
@@ -752,8 +751,12 @@ pub fn composite_with_effects(layer: &Layer, content: &Buffer, backdrop: &mut Bu
         if let Effect::DropShadow(s) = e {
             let mut m = fx(i, 0);
             if s.knocks_out {
+                // The layer hides the shadow beneath it only where its fill is see-through: at
+                // 100 % fill the layer covers it anyway (and anti-aliased edges are not
+                // attenuated twice), at 0 % the shape shows the bare backdrop.
+                let see_through = 1.0 - layer.fill_opacity.clamp(0.0, 1.0);
                 for (v, a) in m.v.iter_mut().zip(&shape.v) {
-                    *v *= 1.0 - a;
+                    *v *= 1.0 - a * see_through;
                 }
             }
             paint_color(&mut work, &m, rgb(&s.color), s.common.blend, s.common.opacity);
@@ -849,28 +852,55 @@ pub fn composite_with_effects(layer: &Layer, content: &Buffer, backdrop: &mut Bu
     }
     let vdout = maps.vdout.as_ref().filter(|_| vector_shape && !strokes.is_empty()).map(|v| maps.crop_vec(v, big, FAR));
     // Outside parts lie beneath the layer and blend onto the backdrop with their own mode. A
-    // higher stroke knocks out the ones below it (Photoshop: a Multiply stroke listed above a
-    // wider Normal stroke multiplies the backdrop, not the lower stroke).
-    let mut cover = vec![0.0f32; w * h];
-    for st in strokes.iter().copied() {
-        let (_, out_w) = widths(st);
-        if out_w <= 0.0 {
-            continue;
+    // higher stroke covers the ones below it (Photoshop: a Multiply stroke listed above a wider
+    // Normal stroke multiplies the backdrop, not the lower stroke). Each stroke's share is its
+    // coverage × opacity not yet taken by strokes above; for Normal strokes this equals stacking
+    // them top over bottom.
+    if strokes.iter().any(|st| widths(st).1 > 0.0) {
+        let base = work.clone();
+        let mut acc = vec![[0.0f32; 4]; w * h]; // premultiplied Σ share·blended
+        let mut cover = vec![0.0f32; w * h];
+        for st in strokes.iter().copied() {
+            let (_, out_w) = widths(st);
+            if out_w <= 0.0 {
+                continue;
+            }
+            // Shape layers: Photoshop strokes the vector outline, estimated from
+            // local coverage; the stroke never shows through the shape's pixels.
+            let d = vdout.as_ref().unwrap_or(&dout);
+            let mut share = vec![0.0f32; w * h];
+            for (i, sh) in share.iter_mut().enumerate() {
+                let k = if inside(shape.v[i]) {
+                    if vector_shape { 0.0 } else { 1.0 }
+                } else {
+                    (out_w + 0.5 - d[i]).clamp(0.0, 1.0)
+                };
+                *sh = k * st.common.opacity * (1.0 - cover[i]);
+                cover[i] += *sh;
+            }
+            // The stroke blended over the backdrop at full coverage.
+            let mut blended = base.clone();
+            let ones = Map { w, h, v: share.iter().map(|&s| if s > 0.0 { 1.0 } else { 0.0 }).collect() };
+            paint_fx(&mut blended, &ones, &st.paint, sb, anchor, big, st.common.blend, 1.0, patterns);
+            for ((a, b), s) in acc.iter_mut().zip(&blended.px).zip(&share) {
+                if *s > 0.0 {
+                    for c in 0..3 {
+                        a[c] += s * b[c] * b[3];
+                    }
+                    a[3] += s * b[3];
+                }
+            }
         }
-        // Shape layers: Photoshop strokes the vector outline, estimated from
-        // local coverage; the stroke never shows through the shape's pixels.
-        let d = vdout.as_ref().unwrap_or(&dout);
-        let mut m = Map::new(w, h, 0.0);
-        for (((mv, a), dv), c) in m.v.iter_mut().zip(&shape.v).zip(d).zip(cover.iter_mut()) {
-            let k = if inside(*a) {
-                if vector_shape { 0.0 } else { 1.0 }
-            } else {
-                (out_w + 0.5 - dv).clamp(0.0, 1.0)
-            };
-            *mv = k * (1.0 - *c);
-            *c += *mv;
+        for ((p, a), c) in work.px.iter_mut().zip(&acc).zip(&cover) {
+            if *c <= 0.0 {
+                continue;
+            }
+            let k = 1.0 - c.min(1.0);
+            let alpha = p[3] * k + a[3];
+            if alpha > 0.0 {
+                *p = [(p[0] * p[3] * k + a[0]) / alpha, (p[1] * p[3] * k + a[1]) / alpha, (p[2] * p[3] * k + a[2]) / alpha, alpha.min(1.0)];
+            }
         }
-        paint_fx(&mut work, &m, &st.paint, sb, anchor, big, st.common.blend, st.common.opacity, patterns);
     }
     for (i, e) in rev() {
         if let Effect::BevelEmboss(b) = e
@@ -1006,6 +1036,21 @@ mod tests {
         assert!(rev > 0.99);
     }
 
+    #[test]
+    fn chamfer_uses_euclidean_step_lengths() {
+        let (w, h) = (9, 9);
+        let mut inside = vec![false; w * h];
+        inside[0] = true;
+        let (d, near) = chamfer_nearest(&inside, w, h);
+        let at = |x: usize, y: usize| d[y * w + x];
+        assert!((at(3, 0) - 3.0).abs() < 1e-5);
+        assert!((at(2, 2) - 2.0 * std::f32::consts::SQRT_2).abs() < 1e-5);
+        assert!((at(1, 2) - 5f32.sqrt()).abs() < 1e-5);
+        assert!((at(1, 3) - (5f32.sqrt() + 1.0)).abs() < 1e-5, "knight + straight, not √10");
+        assert!((at(2, 3) - (5f32.sqrt() + std::f32::consts::SQRT_2)).abs() < 1e-5);
+        assert!(near.iter().all(|&n| n == 0));
+    }
+
     #[test]
     fn offsets_follow_light() {
         let (dx, dy) = offset(120.0, 10.0);
diff --git a/crates/compose/src/tests.rs b/crates/compose/src/tests.rs
index ae276649..4b9869b6 100644
--- a/crates/compose/src/tests.rs
+++ b/crates/compose/src/tests.rs
@@ -660,3 +660,99 @@ fn effect_maps_are_cached_and_invalidated_by_pixel_changes() {
     }
     assert_ne!(first.px, again.px);
 }
+
+// ---------- PSD-fidelity effect semantics (fitted on Photoshop composites) ----------
+
+#[test]
+fn stroke_corners_follow_the_5x5_chamfer_metric() {
+    // Square 10..30, outside stroke 3: pixel (9, 7) is offset (1, 3) from the corner pixel.
+    // Photoshop measures √5 + 1 = 3.236 there (exact distance 3.162): coverage 3 + 1 − 3.236.
+    let d = fx_doc(vec![stroke(3.0, StrokePosition::Outside)]);
+    let a = 1.0 - px(&d, 9, 7)[0]; // blue stroke over white: red channel drops by coverage
+    assert!((a - 0.764).abs() < 0.01, "{a}");
+    // (2, 2) stays Euclidean (two diagonal steps).
+    let a = 1.0 - px(&d, 8, 8)[0];
+    assert!((a - 1.0).abs() < 0.01, "{a}");
+}
+
+#[test]
+fn inside_stroke_respects_partial_edge_coverage() {
+    // A 25 %-alpha edge column puts the edge 0.75 px further out: a 3 px inside stroke then
+    // reaches 75 % into the fourth column, and the edge column takes the stroke colour at 25 %.
+    let mut d = doc_white(40, 40);
+    let mut l = solid_layer("sq", Rect::new(10, 10, 30, 30), [1.0, 0.0, 0.0, 1.0]);
+    l.surface_mut().unwrap().fill_rect(Rect::new(10, 10, 11, 30), &[1.0, 0.0, 0.0, 0.25]);
+    l.effects.items = vec![stroke(3.0, StrokePosition::Inside)];
+    d.layers.push(l);
+    let edge = px(&d, 10, 20);
+    assert!(close4(edge, [0.75, 0.75, 1.0, 1.0]), "edge = blue stroke at 25 % over white: {edge:?}");
+    let p = px(&d, 13, 20);
+    assert!((p[2] - 0.75).abs() < 0.02 && (p[0] - 0.25).abs() < 0.02, "{p:?}");
+}
+
+#[test]
+fn interior_effects_keep_the_layer_alpha() {
+    // A colour overlay replaces a half-transparent pixel's colour without adding coverage.
+    let mut d = Document::new("t", Size::new(40, 40), ColorMode::Rgb, SampleType::U8);
+    let mut l = solid_layer("sq", Rect::new(10, 10, 30, 30), [1.0, 0.0, 0.0, 0.5]);
+    l.effects.items = vec![Effect::ColorOverlay { common: FxCommon::new(photocraft_color::BlendMode::Normal, 1.0), color: Color::rgb(0.0, 1.0, 0.0) }];
+    d.layers.push(l);
+    assert!(close4(px(&d, 20, 20), [0.0, 1.0, 0.0, 0.5]), "{:?}", px(&d, 20, 20));
+}
+
+#[test]
+fn upper_stroke_blends_with_the_backdrop_not_the_lower_stroke() {
+    // Multiply yellow (3 px) listed above a Normal blue (8 px): the inner ring multiplies the
+    // white backdrop (stays yellow), the outer ring is blue.
+    let yellow = Effect::Stroke(StrokeFx {
+        common: FxCommon::new(photocraft_color::BlendMode::Multiply, 1.0),
+        size: 3.0,
+        position: StrokePosition::Outside,
+        paint: FxPaint::Color(Color::rgb(1.0, 1.0, 0.0)),
+    });
+    let d = fx_doc(vec![yellow, stroke(8.0, StrokePosition::Outside)]);
+    assert!(close4(px(&d, 8, 20), [1.0, 1.0, 0.0, 1.0]), "{:?}", px(&d, 8, 20));
+    assert!(close4(px(&d, 4, 20), [0.0, 0.0, 1.0, 1.0]), "{:?}", px(&d, 4, 20));
+}
+
+#[test]
+fn shadow_knockout_needs_see_through_fill() {
+    // At 100 % fill an anti-aliased edge is not attenuated twice: the result equals no knockout.
+    let mut base = Document::new("t", Size::new(40, 40), ColorMode::Rgb, SampleType::U8);
+    base.layers.push(solid_layer("bg", Rect::new(0, 0, 40, 40), [1.0; 4]));
+    let mut l = solid_layer("sq", Rect::new(10, 10, 30, 30), [1.0, 0.0, 0.0, 1.0]);
+    l.surface_mut().unwrap().fill_rect(Rect::new(10, 10, 11, 30), &[1.0, 0.0, 0.0, 0.5]);
+    l.effects.items = vec![Effect::DropShadow(Shadow { distance: 0.0, spread: 1.0, size: 4.0, ..shadow(0.0, 0.0) })];
+    base.layers.push(l);
+    let on = px(&base, 10, 20);
+    if let Effect::DropShadow(s) = &mut base.layers[1].effects.items[0] {
+        s.knocks_out = false;
+    }
+    assert!(close4(on, px(&base, 10, 20)), "{on:?}");
+}
+
+#[test]
+fn linked_pattern_overlay_anchors_at_the_effects_reference_point() {
+    use photocraft_doc::pattern::Pattern;
+    // 2 × 1 tile: red, blue. Anchored at x = 11 (reference point), x = 11 is red, 12 blue.
+    let mut tile = photocraft_raster::Surface::new(PixelFormat::RGBA8);
+    tile.fill_rect(Rect::new(0, 0, 1, 1), &[1.0, 0.0, 0.0, 1.0]);
+    tile.fill_rect(Rect::new(1, 0, 2, 1), &[0.0, 0.0, 1.0, 1.0]);
+    let pat = Pattern::new("rb", tile, 2, 1);
+    let overlay = Effect::PatternOverlay {
+        common: FxCommon::new(photocraft_color::BlendMode::Normal, 1.0),
+        name: "rb".into(),
+        id: pat.id.clone(),
+        scale: 1.0,
+        angle: 0.0,
+        link: true,
+        phase: (0.0, 0.0),
+    };
+    let mut d = fx_doc(vec![overlay]);
+    d.patterns.push(pat);
+    // Without a reference point the layer's top-left (10) anchors the tiling.
+    assert!(close4(px(&d, 10, 20), [1.0, 0.0, 0.0, 1.0]), "{:?}", px(&d, 10, 20));
+    d.layers[1].effects.reference = Some((11.0, 0.0));
+    assert!(close4(px(&d, 11, 20), [1.0, 0.0, 0.0, 1.0]), "{:?}", px(&d, 11, 20));
+    assert!(close4(px(&d, 12, 20), [0.0, 0.0, 1.0, 1.0]), "{:?}", px(&d, 12, 20));
+}
diff --git a/crates/engine/src/tests.rs b/crates/engine/src/tests.rs
index 2850dd54..4411679f 100644
--- a/crates/engine/src/tests.rs
+++ b/crates/engine/src/tests.rs
@@ -420,3 +420,18 @@ fn layer_locks_are_set_and_enforced() {
     let l = st.doc.layer(st.active_layer.unwrap()).unwrap();
     assert!(l.locks.position && !l.locks.pixels);
 }
+
+#[test]
+fn moving_a_layer_moves_its_effects_reference_point() {
+    let mut s = session_with_doc();
+    s.execute("layer.new.layer", json!({})).unwrap();
+    s.edit("ref", |doc, active| {
+        doc.layer_mut(active.unwrap()).unwrap().effects.reference = Some((-201.0, 9.0));
+        Ok(())
+    })
+    .unwrap();
+    s.execute("layer.translate", json!({"dx": 10, "dy": 5})).unwrap();
+    let d = s.active().unwrap();
+    let l = d.doc.layer(d.active_layer.unwrap()).unwrap();
+    assert_eq!(l.effects.reference, Some((-191.0, 14.0)));
+}
diff --git a/crates/gpu/Cargo.toml b/crates/gpu/Cargo.toml
index 9cd2c144..dc08bcd3 100644
--- a/crates/gpu/Cargo.toml
+++ b/crates/gpu/Cargo.toml
@@ -24,3 +24,5 @@ rayon = { workspace = true }
 [dev-dependencies]
 wgpu = { version = "30" }
 pollster = "1"
+# Shape layers in the effect parity tests.
+photocraft-vector = { workspace = true }
diff --git a/crates/gpu/src/compose.wgsl b/crates/gpu/src/compose.wgsl
index fa789be8..5650dce0 100644
--- a/crates/gpu/src/compose.wgsl
+++ b/crates/gpu/src/compose.wgsl
@@ -38,12 +38,12 @@ const F_MASK_TEX: u32 = 2u;      // mask pixels live in `mask_tex`
 const F_TEX: u32 = 4u;           // layer pixels live in `layer_tex`
 const F_GRADIENT: u32 = 8u;      // gradient fill (stops in `lut_tex`)
 const F_KNOCKOUT: u32 = 16u;     // effect paint: the layer knocks out the coverage (drop shadow)
-const F_UNDER: u32 = 32u;        // effect paint: slide beneath A (outside stroke)
 const F_VECTOR: u32 = 64u;       // effect paint: shape layer (outside stroke never inside)
 const F_ATOP: u32 = 128u;        // effect merge: clipped layer over an opaque base
 const F_GATE: u32 = 256u;        // effect paint: coverage only inside the layer's shape
 const F_REL: u32 = 512u;         // effect paint: coverage relative to the layer's alpha
 const F_STROKE_OUT: u32 = 1024u; // effect paint: outside stroke band
+const F_FIRST: u32 = 2048u;      // outside strokes: nothing accumulated yet
 
 @group(0) @binding(0) var<uniform> chunk: Chunk;
 @group(0) @binding(1) var<uniform> op: Op;
@@ -675,9 +675,10 @@ fn fs_fxinit(in: VOut) -> @location(0) vec4<f32> {
 }
 
 // effects::paint: composite `colour × coverage × opacity` into A. B is the layer (its alpha `a`
-// is the effect shape). Coverage `kind`: 0 the effect map, 1 full; then F_KNOCKOUT m × (1 − a),
+// is the effect shape). Coverage `kind`: 0 the effect map, 1 full; then F_KNOCKOUT m × (1 − a k)
+// (k = `p4.w`),
 // F_GATE inside ? m : 0, F_REL inside ? min(m / a, 1) : 0, F_STROKE_OUT inside ? (vector ? 0 : 1)
-// : m. F_UNDER paints beneath A (outside strokes).
+// : m.
 @fragment
 fn fs_fxpaint(in: VOut) -> @location(0) vec4<f32> {
     let p = local(in.pos);
@@ -687,7 +688,7 @@ fn fs_fxpaint(in: VOut) -> @location(0) vec4<f32> {
     let inside_shape = a > INSIDE_EPS;
     var m = 1.0;
     if (op.kind == 0) { m = map_value(d, op.p2.w); }
-    if ((op.flags & F_KNOCKOUT) != 0u) { m = m * (1.0 - a); }
+    if ((op.flags & F_KNOCKOUT) != 0u) { m = m * (1.0 - a * op.p4.w); }
     if ((op.flags & F_GATE) != 0u && !inside_shape) { m = 0.0; }
     if ((op.flags & F_REL) != 0u) {
         if (inside_shape) { m = min(m / a, 1.0); } else { m = 0.0; }
@@ -696,20 +697,55 @@ fn fs_fxpaint(in: VOut) -> @location(0) vec4<f32> {
         m = select(1.0, 0.0, (op.flags & F_VECTOR) != 0u);
     }
     let k = m * op.opacity;
-    if ((op.flags & F_UNDER) != 0u) {
-        // Painted into a transparent buffer, then A goes over it.
-        var under = vec4(0.0);
-        if (k > 0.0) {
-            let c = fx_color(d);
-            under = composite(M_NORMAL, vec4(0.0), vec4(c.rgb, c.a * k), 1.0);
-        }
-        return composite(M_NORMAL, under, dst, 1.0);
-    }
     if (k <= 0.0) { return dst; }
     let c = fx_color(d);
     return composite(op.mode, dst, vec4(c.rgb, c.a * k), 1.0);
 }
 
+// Outside strokes (A = the exterior result before them, B = the layer, C = coverage so far in
+// `.r`, D (`layer_tex`) = accumulated premultiplied colour). Each stroke takes its band × opacity
+// not yet covered above it; `kind` 0 accumulates that share of the stroke blended over A at full
+// coverage, `kind` 1 the coverage.
+@fragment
+fn fs_fxstroke(in: VOut) -> @location(0) vec4<f32> {
+    let p = local(in.pos);
+    let d = doc_px(p);
+    let a = textureLoad(tex_b, p, 0).a;
+    var k = map_value(d, 0.0);
+    if (a > INSIDE_EPS) { k = select(1.0, 0.0, (op.flags & F_VECTOR) != 0u); }
+    let first = (op.flags & F_FIRST) != 0u;
+    var cover = 0.0;
+    if (!first) { cover = textureLoad(tex_c, p, 0).r; }
+    let share = k * op.opacity * (1.0 - cover);
+    if (op.kind == 1) { return vec4(cover + share, 0.0, 0.0, 0.0); }
+    var acc = vec4(0.0);
+    if (!first) { acc = textureLoad(layer_tex, p, 0); }
+    if (share > 0.0) {
+        let base = textureLoad(tex_a, p, 0);
+        var bl = base;
+        if (i32(op.p2.z) != 3) {
+            let c = fx_color(d);
+            bl = composite(op.mode, base, vec4(c.rgb, c.a), 1.0);
+        }
+        acc += share * vec4(bl.rgb * bl.a, bl.a);
+    }
+    return acc;
+}
+
+// Resolve outside strokes: A = exterior result, B = accumulated colour, C = coverage.
+@fragment
+fn fs_fxstrokeend(in: VOut) -> @location(0) vec4<f32> {
+    let p = local(in.pos);
+    let w = textureLoad(tex_a, p, 0);
+    let cover = textureLoad(tex_c, p, 0).r;
+    if (cover <= 0.0) { return w; }
+    let acc = textureLoad(tex_b, p, 0);
+    let k = 1.0 - min(cover, 1.0);
+    let alpha = w.a * k + acc.a;
+    if (alpha <= 0.0) { return w; }
+    return vec4((w.rgb * w.a * k + acc.rgb) / alpha, min(alpha, 1.0));
+}
+
 fn mix_premul(a: vec4<f32>, b: vec4<f32>, k: f32) -> vec4<f32> {
     let alpha = a.a + (b.a - a.a) * k;
     if (alpha <= 0.0) { return vec4(0.0); }
diff --git a/crates/gpu/src/fx.rs b/crates/gpu/src/fx.rs
index 73de9ae3..c1b61a9f 100644
--- a/crates/gpu/src/fx.rs
+++ b/crates/gpu/src/fx.rs
@@ -132,24 +132,24 @@ impl MapProgram {
         let mut count = 0;
         let mut out = vec![None; n];
         let mut held: Vec<(usize, usize)> = Vec::new();
-        for i in 0..n {
+        for (i, (stage, slot)) in self.stages.iter().zip(out.iter_mut()).enumerate() {
             // Release temporaries whose last reader was before this stage.
-            held.retain(|&(stage, t)| {
-                if last_use[stage] < i {
+            held.retain(|&(st, t)| {
+                if last_use[st] < i {
                     free.push(t);
                     false
                 } else {
                     true
                 }
             });
-            if self.stages[i].out.is_some() {
+            if stage.out.is_some() {
                 continue;
             }
             let t = free.pop().unwrap_or_else(|| {
                 count += 1;
                 count - 1
             });
-            out[i] = Some(t);
+            *slot = Some(t);
             held.push((i, t));
         }
         (out, count)
@@ -213,7 +213,7 @@ impl B {
 /// Distance up to which a field must be exact for a band / dilation of width `w`
 /// (`clamp(w + 0.5 - d)`, with `d` up to 1.5 px beyond the pixel distance).
 fn reach_for(w: f32) -> i32 {
-    (w.max(0.0).min(photocraft_compose::effects::MAX_REACH)).ceil() as i32 + 2
+    w.clamp(0.0, photocraft_compose::effects::MAX_REACH).ceil() as i32 + 2
 }
 
 /// `compose::effects::blur`'s kernel: (radius, normalised weights), or None below sigma 0.2.
@@ -478,11 +478,10 @@ pub(crate) fn shape_sources(layer: &Layer) -> (Option<&Surface>, Option<&Surface
     (content, mask)
 }
 
-/// Everything a non-group layer's shape depends on except tile contents.
-pub(crate) fn shape_key(layer: &Layer, canvas: Rect, region: Rect) -> u64 {
+/// Everything a non-group layer's shape depends on except tile contents and position.
+pub(crate) fn shape_key(layer: &Layer, canvas: Rect) -> u64 {
     let mut h = std::collections::hash_map::DefaultHasher::new();
     layer.id.0.hash(&mut h);
-    (region.x0, region.y0, region.x1, region.y1).hash(&mut h);
     let (content, mask) = shape_sources(layer);
     match &layer.content {
         LayerContent::Fill(f) => {
@@ -503,7 +502,7 @@ pub(crate) fn shape_key(layer: &Layer, canvas: Rect, region: Rect) -> u64 {
 }
 
 /// Deep identity of a group (children, their pixels by tile pointer, settings).
-pub(crate) fn group_key(layer: &Layer, region: Rect, light: &GlobalLight) -> u64 {
+pub(crate) fn group_key(layer: &Layer, light: &GlobalLight) -> u64 {
     fn surface_fp(s: &Surface, h: &mut std::collections::hash_map::DefaultHasher) {
         format!("{:?}{:?}", s.format(), s.default_pixel()).hash(h);
         for (c, t) in s.tiles() {
@@ -540,7 +539,7 @@ pub(crate) fn group_key(layer: &Layer, region: Rect, light: &GlobalLight) -> u64
         h.write_u8(0xfe);
     }
     let mut h = std::collections::hash_map::DefaultHasher::new();
-    (region.x0, region.y0, region.x1, region.y1, light.angle.to_bits(), light.altitude.to_bits()).hash(&mut h);
+    (light.angle.to_bits(), light.altitude.to_bits()).hash(&mut h);
     match &layer.content {
         LayerContent::Group(g) => {
             for c in &g.children {
diff --git a/crates/gpu/src/lib.rs b/crates/gpu/src/lib.rs
index 933aa800..0108191b 100644
--- a/crates/gpu/src/lib.rs
+++ b/crates/gpu/src/lib.rs
@@ -331,13 +331,13 @@ fn uniform_entry(binding: u32, size: u64) -> wgpu::BindGroupLayoutEntry {
 /// The compositor's WGSL source (exposed for validation in tests).
 pub const SHADER: &str = include_str!("compose.wgsl");
 
-/// Pass-level resources resolved for one plan.
-struct Bound<'p> {
+/// Pass-level resources resolved for one plan: residents (texture, mask), effect maps (view and
+/// region) and pattern textures per pass.
+struct Bound {
     views: Vec<(Option<ResidentRef>, Option<ResidentRef>)>,
     keys: Vec<(LayerId, Role)>,
     maps: Vec<Option<(wgpu::TextureView, Rect)>>,
     patterns: Vec<Option<wgpu::TextureView>>,
-    _plan: std::marker::PhantomData<&'p ()>,
 }
 
 impl Compositor {
@@ -491,7 +491,7 @@ impl Compositor {
     }
 
     /// Bring every surface `plan` samples up to date and resolve its effect maps and patterns.
-    fn bind_plan<'p>(&mut self, device: &wgpu::Device, queue: &wgpu::Queue, doc: DocId, canvas: Rect, plan: &Plan<'p>, stats: &mut Stats) -> Bound<'p> {
+    fn bind_plan(&mut self, device: &wgpu::Device, queue: &wgpu::Queue, doc: DocId, canvas: Rect, plan: &Plan<'_>, stats: &mut Stats) -> Bound {
         let mut views: Vec<(Option<ResidentRef>, Option<ResidentRef>)> = Vec::with_capacity(plan.passes.len());
         let mut keys: Vec<(LayerId, Role)> = Vec::new();
         let mut maps = Vec::with_capacity(plan.passes.len());
@@ -513,7 +513,7 @@ impl Compositor {
             }));
             patterns.push(p.pattern.map(|pat| self.pattern_view(device, queue, pat)));
         }
-        Bound { views, keys, maps, patterns, _plan: std::marker::PhantomData }
+        Bound { views, keys, maps, patterns }
     }
 
     /// Run `plan` over `chunks` (document rects, at most CHUNK square), handing each finished
@@ -568,7 +568,7 @@ impl Compositor {
             }
             let slot = |s: Option<u32>| s.map_or(dummy, |s| &self.pool[s as usize].1);
             let (tex, mask) = &bound.views[i];
-            let tv = tex.map_or(dummy, |(k, _)| resident_views[k]);
+            let tv = tex.map_or(slot(p.d), |(k, _)| resident_views[k]);
             let mv = mask.map_or(dummy, |(k, _)| resident_views[k]);
             let lv = luts[i].as_ref().map_or(dummy, |(_, v)| v);
             let map = bound.maps[i].as_ref().map_or(dummy, |(v, _)| v);
@@ -724,7 +724,7 @@ impl Compositor {
         let frame = self.frame;
         for (t, used) in &mut self.temps {
             let s = t.texture.size();
-            if s.width == w && s.height == h && !taken.iter().any(|v| *v == t.view) {
+            if s.width == w && s.height == h && !taken.contains(&t.view) {
                 *used = frame;
                 return t.view.clone();
             }
@@ -741,14 +741,18 @@ impl Compositor {
         let region = f.region;
         let canvas = doc.bounds();
         let is_group = matches!(layer.content, LayerContent::Group(_));
-        let shape_key = if is_group { fx::group_key(layer, region, &doc.global_light) } else { fx::shape_key(layer, canvas, region) };
+        let shape_key = if is_group { fx::group_key(layer, &doc.global_light) } else { fx::shape_key(layer, canvas) };
         let (content_src, mask_src) = if is_group { (None, None) } else { fx::shape_sources(layer) };
         let frame = self.frame;
+        let (rw, rh) = (region.width(), region.height());
 
-        // What changed: everything (new region / settings), some tiles, or nothing.
+        // What changed: everything (new settings / size), the position only, some tiles, or
+        // nothing.
         let mut damage = Rect::EMPTY;
-        let reuse = self.fx.get(&layer.id).is_some_and(|e| e.region == region && e.shape_key == shape_key);
-        if reuse {
+        let prev = self.fx.get(&layer.id).filter(|e| e.shape_key == shape_key).map(|e| e.region);
+        let mut rebuild = true;
+        if prev == Some(region) {
+            rebuild = false;
             let e = &self.fx[&layer.id];
             for (old, cur) in e.tiles.iter().zip([content_src, mask_src]) {
                 let d = match (old, cur) {
@@ -760,9 +764,22 @@ impl Compositor {
                 damage = if damage.is_empty() { d } else { damage.union(&d) };
             }
             damage = damage.intersect(&region);
-        } else {
+        } else if let Some(old) = prev
+            && (old.width(), old.height()) == (rw, rh)
+            && !region.is_empty()
+        {
+            // Moved by whole pixels: maps are computed relative to the region, so if the shape
+            // moved unchanged with it, every map is still exact.
+            let v = fx::shape(doc, layer, region);
+            let e = self.fx.get_mut(&layer.id).expect("entry");
+            if v == e.shape_cpu {
+                e.region = region;
+                rebuild = false;
+            }
+        }
+        if rebuild {
             let shape = Tex::map(device, "pc_fx_shape", region, MAP32);
-            let n = region.width() as usize * region.height() as usize;
+            let n = rw as usize * rh as usize;
             self.fx.insert(layer.id, FxEntry { doc: doc.id, region, shape_key, tiles: [None, None], _pin: is_group.then(|| layer.clone()), shape_cpu: vec![0.0; n], shape, fields: HashMap::new(), progs: Vec::new(), last_used: frame });
             damage = region;
         }
@@ -774,15 +791,11 @@ impl Compositor {
             return;
         }
 
-        let t_trace = std::time::Instant::now();
+        let t_trace = web_time_now();
         // Shape over the damage (compose's own alpha).
         if !damage.is_empty() {
             stats.fx_shapes += 1;
-            let ts = std::time::Instant::now();
             let v = fx::shape(doc, layer, damage);
-            if std::env::var_os("PHOTOCRAFT_FX_TRACE").is_some() {
-                eprintln!("  shape {}x{}: {:.2} ms", damage.width(), damage.height(), ts.elapsed().as_secs_f64() * 1000.0);
-            }
             fx::paste(&mut e.shape_cpu, region, damage, &v);
             e.shape.write_r32(queue, region, damage, &v);
         }
@@ -808,11 +821,7 @@ impl Compositor {
                 _ => region,
             };
             let reach = have.filter(|r| *r >= reach).unwrap_or(reach);
-            let tf = std::time::Instant::now();
             let v = fx::field(kind, reach, &e.shape_cpu, region, out);
-            if std::env::var_os("PHOTOCRAFT_FX_TRACE").is_some() {
-                eprintln!("  field {kind:?} reach {reach} over {}x{}: {:.2} ms", out.width(), out.height(), tf.elapsed().as_secs_f64() * 1000.0);
-            }
             if have.is_none_or(|r| r < reach) {
                 e.fields.insert(kind, (reach, Tex::map(device, "pc_fx_field", region, MAP32)));
             }
@@ -820,8 +829,8 @@ impl Compositor {
             stats.fx_pixels += out.width() as u64 * out.height() as u64;
         }
 
-        if std::env::var_os("PHOTOCRAFT_FX_TRACE").is_some() && !damage.is_empty() {
-            eprintln!("fx `{}`: damage {:?} shape+fields {:.2} ms", layer.name, damage, t_trace.elapsed().as_secs_f64() * 1000.0);
+        if let Some(t) = t_trace.filter(|_| !damage.is_empty()) {
+            eprintln!("effect maps of `{}`: shape + distance fields over {damage:?} in {:.2} ms (CPU)", layer.name, t.elapsed().as_secs_f64() * 1000.0);
         }
         e.progs.truncate(progs.len());
         while e.progs.len() < progs.len() {
@@ -899,6 +908,18 @@ impl Compositor {
     }
 }
 
+/// A start time when `PHOTOCRAFT_FX_TRACE=1` (it prints the CPU time of effect shapes and
+/// distance fields); never on wasm.
+fn web_time_now() -> Option<std::time::Instant> {
+    #[cfg(not(target_arch = "wasm32"))]
+    {
+        static T: std::sync::OnceLock<bool> = std::sync::OnceLock::new();
+        T.get_or_init(|| std::env::var_os("PHOTOCRAFT_FX_TRACE").is_some()).then(std::time::Instant::now)
+    }
+    #[cfg(target_arch = "wasm32")]
+    None
+}
+
 /// The uniform record of a pass.
 fn op_words(p: &plan::Pass<'_>, tex: Option<[i32; 4]>, mask: Option<[i32; 4]>, map: Option<Rect>) -> Vec<u8> {
     let mut flags = p.flags;
diff --git a/crates/gpu/src/plan.rs b/crates/gpu/src/plan.rs
index 50a978f2..a959a05f 100644
--- a/crates/gpu/src/plan.rs
+++ b/crates/gpu/src/plan.rs
@@ -42,6 +42,11 @@ pub enum Kernel {
     FxPaint,
     /// Layer + interior effects over the exterior result, then layer opacity against the backdrop.
     FxMerge,
+    /// Outside strokes: accumulate one stroke's share (`kind` 0, premultiplied colour) or its
+    /// coverage (`kind` 1).
+    FxStroke,
+    /// Outside strokes: resolve the accumulated strokes over the exterior result.
+    FxStrokeEnd,
     /// Copy slot A into `dst` (an existing slot) over the pass clip (no draw).
     CopyRect,
     /// Copy all of slot A into the fresh slot `dst` (no draw).
@@ -72,6 +77,8 @@ impl Kernel {
             Kernel::FxInit => "fs_fxinit",
             Kernel::FxPaint => "fs_fxpaint",
             Kernel::FxMerge => "fs_fxmerge",
+            Kernel::FxStroke => "fs_fxstroke",
+            Kernel::FxStrokeEnd => "fs_fxstrokeend",
             Kernel::MShift => "fs_mshift",
             Kernel::MDilate => "fs_mdilate",
             Kernel::MBlur => "fs_mblur",
@@ -99,7 +106,7 @@ impl Kernel {
     }
 
     /// Every kernel with a pipeline.
-    pub const DRAWN: [Kernel; 18] = [
+    pub const DRAWN: [Kernel; 20] = [
         Kernel::Content,
         Kernel::Mask,
         Kernel::Blend,
@@ -110,6 +117,8 @@ impl Kernel {
         Kernel::FxInit,
         Kernel::FxPaint,
         Kernel::FxMerge,
+        Kernel::FxStroke,
+        Kernel::FxStrokeEnd,
         Kernel::MShift,
         Kernel::MDilate,
         Kernel::MBlur,
@@ -151,6 +160,8 @@ pub struct Pass<'a> {
     pub a: Option<Slot>,
     pub b: Option<Slot>,
     pub c: Option<Slot>,
+    /// A fourth slot, bound in place of `tex` (passes without layer pixels).
+    pub d: Option<Slot>,
     pub mode: BlendMode,
     pub opacity: f32,
     /// Layer pixels (raster / text / shape / smart cache / fill cache).
@@ -194,6 +205,7 @@ impl<'a> Pass<'a> {
             a: None,
             b: None,
             c: None,
+            d: None,
             mode: BlendMode::Normal,
             opacity: 1.0,
             tex: None,
@@ -285,11 +297,9 @@ enum Cov {
 
 /// Shader flags for effect passes (keep in sync with compose.wgsl). Coverage `m` of a paint, with
 /// `a` the layer's alpha and `inside` = `a > 0.5/255`:
-/// knockout `m × (1 − a)`; gate `inside ? m : 0`; rel `inside ? min(m / a, 1) : 0`;
+/// knockout `m × (1 − a × k)` (k in `p4.w`); gate `inside ? m : 0`; rel `inside ? min(m / a, 1) : 0`;
 /// stroke-out `inside ? (vector ? 0 : 1) : m`.
 pub const F_KNOCKOUT: u32 = 16;
-/// Paint beneath A instead of over it (outside strokes).
-pub const F_UNDER: u32 = 32;
 /// Shape layer (outside strokes never show inside it).
 pub const F_VECTOR: u32 = 64;
 /// Merge onto an opaque clipping base, keeping its alpha.
@@ -297,6 +307,8 @@ pub const F_ATOP: u32 = 128;
 pub const F_GATE: u32 = 256;
 pub const F_REL: u32 = 512;
 pub const F_STROKE_OUT: u32 = 1024;
+/// First outside stroke (no accumulated share yet).
+pub const F_FIRST: u32 = 2048;
 
 /// `FxInit` kinds: A at `opacity` × alpha; an opaque copy of A; A's colour with alpha `opacity`
 /// inside B's shape; A with alpha × B's alpha.
@@ -337,9 +349,9 @@ impl<'a> Planner<'a> {
     fn emit(&mut self, mut pass: Pass<'a>) -> Slot {
         let dst = self.alloc();
         pass.dst = dst;
-        let (a, b, c) = (pass.a, pass.b, pass.c);
+        let (a, b, c, d) = (pass.a, pass.b, pass.c, pass.d);
         self.passes.push(pass);
-        for s in [a, b, c].into_iter().flatten() {
+        for s in [a, b, c, d].into_iter().flatten() {
             self.release(s);
         }
         dst
@@ -618,8 +630,10 @@ impl<'a> Planner<'a> {
         };
         for &(i, e) in &rev {
             if let Effect::DropShadow(s) = e {
+                // The layer hides the shadow beneath it where its fill is see-through.
                 let flags = if s.knocks_out { F_KNOCKOUT } else { 0 };
-                w = self.paint(w, content, Cov::Map(map(i, 0), 0.0), &Paint::Color(s.color.to_rgb()), s.common.blend, s.common.opacity, flags, clip, sb);
+                let see_through = 1.0 - layer.fill_opacity.clamp(0.0, 1.0);
+                w = self.paint_k(w, content, Cov::Map(map(i, 0), 0.0), &Paint::Color(s.color.to_rgb()), s.common.blend, s.common.opacity, flags, clip, sb, see_through);
             }
         }
         for &(i, e) in &rev {
@@ -688,15 +702,39 @@ impl<'a> Planner<'a> {
         // The shape's own alpha.
         let c = self.retain(content);
         l = init(self, l, Some(c), INIT_ALPHA, 1.0);
-        // Outside stroke parts slide beneath the layer (top instance first).
-        for (i, e) in items.iter().copied().enumerate() {
-            if let Effect::Stroke(st) = e
-                && stroke_widths(st).1 > 0.0
-            {
+        // Outside stroke parts lie beneath the layer and blend onto the exterior result with their
+        // own mode; a higher stroke covers the ones below it (each takes its coverage × opacity
+        // not yet taken above it, blended over the result as it was before the strokes).
+        let outs: Vec<(usize, &photocraft_doc::StrokeFx)> = items.iter().copied().enumerate().filter_map(|(i, e)| if let Effect::Stroke(st) = e { Some((i, st)) } else { None }).filter(|(_, st)| stroke_widths(st).1 > 0.0).collect();
+        if !outs.is_empty() {
+            let (mut acc, mut cover): (Option<Slot>, Option<Slot>) = (None, None);
+            for (n, &(i, st)) in outs.iter().enumerate() {
                 let paint = self.fx_paint(&st.paint, anchor);
-                let flags = F_UNDER | F_STROKE_OUT | if vector_shape { F_VECTOR } else { 0 };
-                l = self.paint(l, content, Cov::Map(map(i, 0), 0.0), &paint, BlendMode::Normal, st.common.opacity, flags, clip, sb);
+                let flags = F_STROKE_OUT | if vector_shape { F_VECTOR } else { 0 } | if n == 0 { F_FIRST } else { 0 };
+                // Share × blended colour, then coverage (both read the coverage so far).
+                let mut p = self.fx_pass_kernel(Kernel::FxStroke, w, content, &paint, st.common.blend, st.common.opacity, flags, clip, sb);
+                if matches!(paint, Paint::None) {
+                    // A missing pattern paints nothing: the stroke's share keeps the result as is.
+                    p.params[2][2] = 3.0;
+                }
+                p.adjust_kind = 0;
+                p.map = Some(map(i, 0));
+                p.c = cover.map(|c| self.retain(c));
+                p.d = acc;
+                let a = self.emit(p);
+                let mut p = self.fx_pass_kernel(Kernel::FxStroke, w, content, &Paint::Color([0.0; 3]), BlendMode::Normal, st.common.opacity, flags, clip, sb);
+                p.adjust_kind = 1;
+                p.map = Some(map(i, 0));
+                p.c = cover;
+                cover = Some(self.emit(p));
+                acc = Some(a);
             }
+            let mut p = Pass::new(Kernel::FxStrokeEnd, 0);
+            p.a = Some(w);
+            p.b = acc;
+            p.c = cover;
+            p.clip = Some(clip);
+            w = self.emit(p);
         }
         for &(i, e) in &rev {
             if let Effect::BevelEmboss(b) = e
@@ -748,28 +786,22 @@ impl<'a> Planner<'a> {
         }
     }
 
-    /// One effect paint into `dst` (consumed); reads the layer's alpha from `content` (kept).
+    /// An effect pass reading `dst` (A, retained) and the layer `content` (B, retained) with its
+    /// paint set up.
     #[allow(clippy::too_many_arguments)]
-    fn paint(&mut self, dst: Slot, content: Slot, cov: Cov, paint: &Paint<'a>, blend: BlendMode, opacity: f32, flags: u32, clip: Rect, sb: Rect) -> Slot {
-        if matches!(paint, Paint::None) {
-            // Missing pattern: compose paints nothing.
-            return dst;
-        }
-        let mut p = Pass::new(Kernel::FxPaint, 0);
-        p.a = Some(dst);
+    fn fx_pass(&mut self, dst: Slot, content: Slot, paint: &Paint<'a>, blend: BlendMode, opacity: f32, flags: u32, clip: Rect, sb: Rect) -> Pass<'a> {
+        self.fx_pass_kernel(Kernel::FxPaint, dst, content, paint, blend, opacity, flags, clip, sb)
+    }
+
+    #[allow(clippy::too_many_arguments)]
+    fn fx_pass_kernel(&mut self, kernel: Kernel, dst: Slot, content: Slot, paint: &Paint<'a>, blend: BlendMode, opacity: f32, flags: u32, clip: Rect, sb: Rect) -> Pass<'a> {
+        let mut p = Pass::new(kernel, 0);
+        p.a = Some(self.retain(dst));
         p.b = Some(self.retain(content));
         p.mode = blend;
         p.opacity = opacity;
         p.flags = flags;
         p.clip = Some(clip);
-        match cov {
-            Cov::Map(m, outside) => {
-                p.adjust_kind = 0;
-                p.map = Some(m);
-                p.params[2][3] = outside;
-            }
-            Cov::One => p.adjust_kind = 1,
-        }
         match paint {
             Paint::None => {}
             Paint::Color(c) => p.color = [c[0], c[1], c[2], 1.0],
@@ -795,6 +827,33 @@ impl<'a> Planner<'a> {
                 p.pattern = Some(pat);
             }
         }
+        p
+    }
+
+    /// One effect paint into `dst` (consumed); reads the layer's alpha from `content` (kept).
+    #[allow(clippy::too_many_arguments)]
+    fn paint(&mut self, dst: Slot, content: Slot, cov: Cov, paint: &Paint<'a>, blend: BlendMode, opacity: f32, flags: u32, clip: Rect, sb: Rect) -> Slot {
+        self.paint_k(dst, content, cov, paint, blend, opacity, flags, clip, sb, 1.0)
+    }
+
+    /// [`Self::paint`] with the knockout strength (`F_KNOCKOUT`: coverage × (1 − alpha × k)).
+    #[allow(clippy::too_many_arguments)]
+    fn paint_k(&mut self, dst: Slot, content: Slot, cov: Cov, paint: &Paint<'a>, blend: BlendMode, opacity: f32, flags: u32, clip: Rect, sb: Rect, knockout: f32) -> Slot {
+        if matches!(paint, Paint::None) {
+            // Missing pattern: compose paints nothing.
+            return dst;
+        }
+        let mut p = self.fx_pass(dst, content, paint, blend, opacity, flags, clip, sb);
+        self.release(dst);
+        match cov {
+            Cov::Map(m, outside) => {
+                p.adjust_kind = 0;
+                p.map = Some(m);
+                p.params[2][3] = outside;
+            }
+            Cov::One => p.adjust_kind = 1,
+        }
+        p.extra[3] = knockout;
         self.emit(p)
     }
 }
diff --git a/crates/gpu/tests/parity.rs b/crates/gpu/tests/parity.rs
index 0e953425..35ae977a 100644
--- a/crates/gpu/tests/parity.rs
+++ b/crates/gpu/tests/parity.rs
@@ -451,6 +451,14 @@ fn effect_cases() -> Vec<(&'static str, Vec<Effect>)> {
         ("gradient overlay", vec![Effect::GradientOverlay { common: FxCommon::new(BlendMode::Normal, 0.8), gradient: Gradient { style: GradientStyle::Radial, ..gradient() }, dither: false }]),
         ("pattern overlay", vec![Effect::PatternOverlay { common: FxCommon::new(BlendMode::Normal, 0.9), name: pat.name.clone(), id: pat.id.clone(), scale: 1.0, angle: 0.0, link: true, phase: (2.0, 1.0) }]),
         ("pattern overlay scaled rotated", vec![Effect::PatternOverlay { common: FxCommon::new(BlendMode::Screen, 0.8), name: pat.name.clone(), id: pat.id.clone(), scale: 1.7, angle: 30.0, link: false, phase: (0.0, 0.0) }]),
+        (
+            "missing pattern",
+            vec![
+                Effect::PatternOverlay { common: FxCommon::new(BlendMode::Normal, 1.0), name: "nope".into(), id: "nope".into(), scale: 1.0, angle: 0.0, link: true, phase: (0.0, 0.0) },
+                Effect::Stroke(stroke(3.0, StrokePosition::Outside, FxPaint::Pattern { name: "nope".into(), id: "nope".into(), scale: 1.0 })),
+                Effect::Stroke(stroke(6.0, StrokePosition::Outside, FxPaint::Color(Color::rgb(0.0, 0.0, 0.0)))),
+            ],
+        ),
         (
             "multiple instances",
             vec![
@@ -637,7 +645,7 @@ fn layer_effects_update_incrementally() {
     }
     let s = fx_check(&mut g, &d, "drop shadow spread");
     assert_eq!(s.fx_programs, 1, "{s:?}");
-    // Moving the layer (new region) rebuilds everything.
+    // Moving the layer by whole pixels moves its maps with it: nothing is rebuilt.
     let moved = {
         let src = d.layers[1].surface().unwrap();
         let b = src.content_bounds();
@@ -646,5 +654,32 @@ fn layer_effects_update_incrementally() {
         dst
     };
     *d.layers[1].surface_mut().unwrap() = moved;
-    fx_check(&mut g, &d, "moved");
+    let s = fx_check(&mut g, &d, "moved");
+    assert_eq!(s.fx_programs, 0, "{s:?}");
+    // A change of shape with the move rebuilds.
+    d.layers[1].surface_mut().unwrap().fill_rect(Rect::new(300, 300, 310, 310), &[1.0, 1.0, 1.0, 0.5]);
+    fx_check(&mut g, &d, "moved and painted");
+}
+
+#[test]
+fn layer_effects_on_shape_layers() {
+    let Some(mut g) = gpu() else { return };
+    use photocraft_doc::vector::{Path, ShapeLayer, Subpath};
+    for alpha in [1.0, 0.55] {
+        let mut d = fx_doc(90, 70, SampleType::U8);
+        let path = Path::new(vec![Subpath::polygon(&[(14.3, 12.6), (70.2, 18.1), (60.7, 58.4), (24.9, 50.2)])]);
+        let mut fill = Color::rgb(0.3, 0.6, 0.9);
+        fill.alpha = alpha;
+        let mut sh = ShapeLayer { path, fill: Some(Fill::Solid(fill)), stroke: None, live: None, cache: None, psd_raw: None };
+        sh.cache = Some(photocraft_vector::render_shape(&sh, d.pixel_format(), d.bounds()));
+        let mut l = Layer::new("shape", LayerContent::Shape(sh));
+        l.effects.items = vec![
+            Effect::Stroke(stroke(3.0, StrokePosition::Outside, FxPaint::Color(Color::rgb(1.0, 1.0, 1.0)))),
+            Effect::Stroke(stroke(4.0, StrokePosition::Center, FxPaint::Gradient(gradient()))),
+            Effect::DropShadow(shadow(BlendMode::Multiply, 0.7, 120.0, 4.0, 5.0, 0.0)),
+            Effect::BevelEmboss(bevel(BevelStyle::InnerBevel, true, 4.0, 1.0)),
+        ];
+        d.layers.push(l);
+        fx_check(&mut g, &d, &format!("shape layer alpha {alpha}"));
+    }
 }
diff --git a/crates/io/examples/oracle_diff.rs b/crates/io/examples/oracle_diff.rs
index 1fe76b93..0bd5c3fc 100644
--- a/crates/io/examples/oracle_diff.rs
+++ b/crates/io/examples/oracle_diff.rs
@@ -5,6 +5,9 @@
 //! cargo run --release -p photocraft-io --example oracle_diff -- file.psd 0 png out.png # ours | Photoshop | diff heatmap
 //! cargo run --release -p photocraft-io --example oracle_diff -- file.psd 0 col [x]     # column samples
 //! cargo run --release -p photocraft-io --example oracle_diff -- file.psd 0 row y x0 x1 # row samples
+//! cargo run --release -p photocraft-io --example oracle_diff -- file.psd 0 worst [n]   # n worst pixels
+//! cargo run --release -p photocraft-io --example oracle_diff -- file.psd 0 grid x0 y0 x1 y1 [ch] # value grids
+//! cargo run --release -p photocraft-io --example oracle_diff -- file.psd 0 layerpx x y  # each layer's pixel
 //! DUMP_FX=1 …                                                                          # raw effects descriptors
 //! ```
 fn main() {
diff --git a/crates/io/examples/pattern_fit.rs b/crates/io/examples/pattern_fit.rs
deleted file mode 100644
index 46552918..00000000
--- a/crates/io/examples/pattern_fit.rs
+++ /dev/null
@@ -1,31 +0,0 @@
-//! TEMPORARY: transfer-curve fit for an adjustment layer vs Photoshop's merged image.
-//! args: file.psd "layer name" x0 y0 x1 y1
-fn main() {
-    let a: Vec<String> = std::env::args().collect();
-    let bytes = std::fs::read(&a[1]).unwrap();
-    let file = photocraft_psd::PsdFile::from_bytes(&bytes).unwrap();
-    let imp = photocraft_io::import(&a[1], &bytes).unwrap();
-    let mut doc = imp.document.clone();
-    let merged = photocraft_io::merged_composite(&file).unwrap();
-    let w = doc.size.width as usize;
-    let r: Vec<usize> = a[3..7].iter().map(|s| s.parse().unwrap()).collect();
-    let id = doc.walk().into_iter().find(|t| t.2.name == a[2]).map(|t| t.2.id).expect("layer");
-    doc.layer_mut(id).unwrap().visible = false;
-    let below = photocraft_compose::flatten(&doc).px;
-    let mut acc = vec![(0.0f64, 0usize); 256];
-    for y in r[1]..r[3] {
-        for x in r[0]..r[2] {
-            let i = y * w + x;
-            let v = (below[i][0] * 255.0).round().clamp(0.0, 255.0) as usize;
-            acc[v].0 += f64::from(merged[i][0]) * 255.0;
-            acc[v].1 += 1;
-        }
-    }
-    let mut line = Vec::new();
-    for (v, (s, n)) in acc.iter().enumerate() {
-        if *n > 0 {
-            line.push(format!("{v}:{:.1}", s / *n as f64));
-        }
-    }
-    println!("{}", line.join(" "));
-}
diff --git a/crates/io/src/adjust_map.rs b/crates/io/src/adjust_map.rs
index 7e42a320..24adb467 100644
--- a/crates/io/src/adjust_map.rs
+++ b/crates/io/src/adjust_map.rs
@@ -240,9 +240,10 @@ fn write_gradient_map(stops: &[(f32, [f32; 3])], reverse: bool) -> Vec<u8> {
         v.extend_from_slice(&((t.clamp(0.0, 1.0) * 4096.0).round() as u32).to_be_bytes());
         v.extend_from_slice(&50u32.to_be_bytes());
         put16(&mut v, 0);
-        for k in 0..4 {
-            put16(&mut v, if k < 3 { (c[k].clamp(0.0, 1.0) * 65535.0).round() as u16 } else { 0 });
+        for x in c {
+            put16(&mut v, (x.clamp(0.0, 1.0) * 65535.0).round() as u16);
         }
+        put16(&mut v, 0); // fourth component
         put16(&mut v, 0);
     }
     put16(&mut v, 2);
diff --git a/crates/io/tests/roundtrip.rs b/crates/io/tests/roundtrip.rs
index 63cb8cf8..e7c6bc74 100644
--- a/crates/io/tests/roundtrip.rs
+++ b/crates/io/tests/roundtrip.rs
@@ -352,3 +352,51 @@ fn clearing_effects_drops_stale_blocks() {
     let rec = f.layers().iter().find(|r| r.name() == back.layers[1].name).unwrap();
     assert!(rec.block(b"lfx2").is_none());
 }
+
+#[test]
+fn link_groups_and_effects_reference_round_trip() {
+    use photocraft_doc::{Document, Layer, LayerContent, Size};
+    let mut d = Document::new("l", Size::new(8, 8), ColorMode::Rgb, SampleType::U8);
+    let fmt = d.pixel_format();
+    let mut a = Layer::raster("a", fmt);
+    a.surface_mut().unwrap().fill_rect(photocraft_geom::Rect::new(0, 0, 4, 4), &[1.0, 0.0, 0.0, 1.0]);
+    a.link_group = Some(7);
+    a.effects.reference = Some((-201.0, 47.5));
+    let mut b = Layer::raster("b", fmt);
+    b.link_group = Some(7);
+    let mut inner = Layer::raster("inner", fmt);
+    inner.link_group = Some(3);
+    let g = Layer::group("g", vec![inner]);
+    d.layers = vec![a, b, g];
+    let back = roundtrip(&d);
+    assert_eq!(back.layers[0].link_group, Some(7));
+    assert_eq!(back.layers[1].link_group, Some(7));
+    assert_eq!(back.layers[2].link_group, None);
+    let LayerContent::Group(gb) = &back.layers[2].content else { panic!() };
+    assert_eq!(gb.children[0].link_group, Some(3));
+    assert_eq!(back.layers[0].effects.reference, Some((-201.0, 47.5)));
+    assert_eq!(back.layers[1].effects.reference, None);
+    // Unchanged documents write the same resource bytes again.
+    let r1 = export(&back, "x.psd", &ExportOptions::default()).unwrap().bytes;
+    let f1 = PsdFile::from_bytes(&r1).unwrap();
+    let res = |f: &PsdFile| f.resources.iter().find(|r| r.id == 1026).map(|r| r.data.clone());
+    let f0 = PsdFile::from_bytes(&export(&d, "x.psd", &ExportOptions::default()).unwrap().bytes).unwrap();
+    assert_eq!(res(&f1), res(&f0));
+    // One u16 per record: a, b, divider, inner, group (bottom first).
+    assert_eq!(res(&f0).unwrap(), vec![0, 7, 0, 7, 0, 0, 0, 3, 0, 0]);
+    // Large engine ids are renumbered by first appearance.
+    let mut big = d.clone();
+    big.layers[0].link_group = Some(1 << 40);
+    big.layers[1].link_group = Some(1 << 40);
+    let fb = PsdFile::from_bytes(&export(&big, "x.psd", &ExportOptions::default()).unwrap().bytes).unwrap();
+    assert_eq!(res(&fb).unwrap(), vec![0, 1, 0, 1, 0, 0, 0, 2, 0, 0]);
+    // Unlinked documents write no 1026.
+    let mut none = d.clone();
+    none.layers[0].link_group = None;
+    none.layers[1].link_group = None;
+    if let LayerContent::Group(g) = &mut none.layers[2].content {
+        g.children[0].link_group = None;
+    }
+    let fnone = PsdFile::from_bytes(&export(&none, "x.psd", &ExportOptions::default()).unwrap().bytes).unwrap();
+    assert!(res(&fnone).is_none());
+}
diff --git a/crates/ui-egui/examples/fx_bench.rs b/crates/ui-egui/examples/fx_bench.rs
index 3484bc34..cc02f582 100644
--- a/crates/ui-egui/examples/fx_bench.rs
+++ b/crates/ui-egui/examples/fx_bench.rs
@@ -135,18 +135,53 @@ fn cpu_time(doc: &Document, region: Rect) -> f64 {
     t.elapsed().as_secs_f64() * 1000.0
 }
 
-fn report(what: &str, region: Rect, cpu: Option<f64>, gpu: Result<f64, String>) {
-    let mp = region.width() as f64 * region.height() as f64 / 1e6;
-    let c = cpu.map_or("-".to_string(), |v| format!("{v:9.1} ms"));
-    let g = match gpu {
-        Ok(v) => format!("{v:9.1} ms"),
-        Err(e) => format!("unsupported ({e})"),
-    };
-    println!("{what:<40} {mp:7.2} MP   cpu {c:>12}   gpu {g}");
+/// `--compare a.psd b.psd …`: GPU vs CPU max difference (premultiplied, /255) per file.
+fn compare(files: &[String]) {
+    let Some(mut g) = gpu() else { return };
+    let mut worst_all = 0.0f32;
+    for p in files {
+        let Ok(bytes) = std::fs::read(p) else { continue };
+        let Ok(r) = photocraft_io::import(p, &bytes) else {
+            println!("{p}: import failed");
+            continue;
+        };
+        let doc = r.document;
+        let fx = doc.walk().iter().filter(|(_, _, l)| photocraft_compose::effects::has_effects(l)).count();
+        let cpu = photocraft_compose::flatten(&doc);
+        match photocraft_gpu::render_to_vec(&mut g.comp, &g.device, &g.queue, &doc, doc.bounds()) {
+            Ok(out) => {
+                let mut worst = (0.0f32, 0usize);
+                for (i, (c, o)) in cpu.px.iter().zip(&out).enumerate() {
+                    for k in 0..4 {
+                        let (a, b) = if k == 3 { (c[3], o[3]) } else { (c[k] * c[3], o[k] * o[3]) };
+                        let d = (a - b).abs();
+                        if d > worst.0 || d.is_nan() {
+                            worst = (if d.is_nan() { 9.0 } else { d }, i);
+                        }
+                    }
+                }
+                let w = doc.size.width as usize;
+                worst_all = worst_all.max(worst.0);
+                println!("{:<70} {fx:>2} fx layers  max diff {:6.2}/255 at ({},{})", p, worst.0 * 255.0, worst.1 % w, worst.1 / w);
+                if std::env::var_os("FX_DUMP").is_some() && worst.0 > 1.0 / 255.0 {
+                    println!("    cpu {:?} gpu {:?}", cpu.px[worst.1], out[worst.1]);
+                    for (_, depth, l) in doc.walk() {
+                        println!("    {}{} {:?} op {} fill {} clipped {} visible {} fx {:?}", "  ".repeat(depth), l.name, l.blend, l.opacity, l.fill_opacity, l.clipped, l.visible, l.effects.items.iter().map(|e| e.label()).collect::<Vec<_>>());
+                    }
+                }
+            }
+            Err(e) => println!("{p:<70} {fx:>2} fx layers  CPU fallback: {e}"),
+        }
+    }
+    println!("worst over all files: {:.2}/255", worst_all * 255.0);
 }
 
 fn main() {
     let args: Vec<String> = std::env::args().collect();
+    if let Some(i) = args.iter().position(|a| a == "--compare") {
+        compare(&args[i + 1..]);
+        return;
+    }
     let no_cpu = args.iter().any(|a| a == "--no-cpu");
     let mut doc = if let Some(p) = arg(&args, "--psd") {
         let bytes = std::fs::read(&p).expect("read --psd");
@@ -162,38 +197,88 @@ fn main() {
         Some(g) => gpu_time(g, doc, r),
         None => Err("no adapter".into()),
     };
-    let cpu = |doc: &Document, r: Rect| if no_cpu { None } else { Some(cpu_time(doc, r)) };
+
+    let reps: usize = arg(&args, "--reps").and_then(|v| v.parse().ok()).unwrap_or(5);
+    let stats = |v: &mut Vec<f64>| -> Option<(f64, f64)> {
+        v.sort_by(|a, b| a.partial_cmp(b).unwrap_or(std::cmp::Ordering::Equal));
+        (!v.is_empty()).then(|| (v[0], v[v.len() / 2]))
+    };
+    let show = |what: &str, r: Rect, c: Option<(f64, f64)>, g: Result<(f64, f64), String>| {
+        let mp = r.width() as f64 * r.height() as f64 / 1e6;
+        let c = c.map_or("-".to_string(), |(m, med)| format!("{m:8.1} / {med:8.1} ms"));
+        let g = match g {
+            Ok((m, med)) => format!("{m:7.1} / {med:7.1} ms"),
+            Err(e) => format!("unsupported ({e})"),
+        };
+        println!("{what:<42} {mp:6.2} MP   cpu {c:>22}   gpu {g}   (min / median)");
+    };
+    // Each case: `edit` mutates the document, then both compositors refresh `region`.
+    let mut case = |doc: &mut Document, what: &str, n: usize, edit: &mut dyn FnMut(&mut Document, usize) -> Rect, cpu_too: bool| {
+        let (mut cs, mut gs, mut gerr) = (Vec::new(), Vec::new(), None);
+        let mut region = full;
+        for i in 0..n {
+            region = edit(doc, i).intersect(&full);
+            if cpu_too && !no_cpu {
+                cs.push(cpu_time(doc, region));
+            }
+            match gt(doc, region) {
+                Ok(v) => gs.push(v),
+                Err(e) => gerr = Some(e),
+            }
+        }
+        show(what, region, stats(&mut cs), match gerr {
+            Some(e) => Err(e),
+            None => Ok(stats(&mut gs).unwrap_or((0.0, 0.0))),
+        });
+    };
 
     if args.iter().any(|a| a == "--baseline") {
         let mut plain = doc.clone();
         for l in &mut plain.layers {
             l.effects.items.clear();
         }
-        let _ = gt(&plain, full);
-        report("baseline: no effects, full refresh", full, None, gt(&plain, full));
+        case(&mut plain, "baseline: no effects, full refresh", reps, &mut |_, _| full, false);
     }
-    photocraft_compose::purge_effect_cache();
-    report("full refresh, cold effect caches", full, cpu(&doc, full), gt(&doc, full));
-    report("full refresh, warm caches", full, cpu(&doc, full), gt(&doc, full));
-
+    case(&mut doc, "full refresh, cold effect caches", 1, &mut |_, _| {
+        photocraft_compose::purge_effect_cache();
+        full
+    }, true);
+    case(&mut doc, "full refresh, warm caches", reps.min(3), &mut |_, _| full, true);
     // Adjustment tweak (maps unaffected).
-    if let Some(LayerContent::Adjustment(Adjustment::HueSaturation { hue, .. })) = doc.layers.last_mut().map(|l| &mut l.content) {
-        *hue += 5.0;
-    }
-    report("adjustment tweak (full refresh)", full, cpu(&doc, full), gt(&doc, full));
-
-    // Brush dab on a plain layer.
+    case(&mut doc, "adjustment tweak (full refresh)", reps.min(3), &mut |d, _| {
+        if let Some(LayerContent::Adjustment(Adjustment::HueSaturation { hue, .. })) = d.layers.last_mut().map(|l| &mut l.content) {
+            *hue += 5.0;
+        }
+        full
+    }, true);
+    // Brush dabs on a plain layer and on the effect layer: the canvas refreshes the damage rect
+    // grown by the effect reach (`canvas::effect_reach`).
     let margin = doc.walk().iter().map(|(_, _, l)| photocraft_compose::effects::margin(l)).max().unwrap_or(0);
-    let dab = Rect::new(full.width() as i32 / 3, full.height() as i32 / 3, full.width() as i32 / 3 + 64, full.height() as i32 / 3 + 64);
-    if let Some(l) = doc.layers.iter_mut().find(|l| l.name == "paint") {
-        l.surface_mut().unwrap().fill_rect(dab, &[0.1, 0.9, 0.1, 1.0]);
-        report("dab on a plain layer (damage rect)", dab, cpu(&doc, dab), gt(&doc, dab));
-    }
-    // Brush dab on the effect layer: its effects reach `margin` px beyond the dab.
-    let blob_dab = Rect::new(full.width() as i32 / 2 - 32, full.height() as i32 / 2 - 32, full.width() as i32 / 2 + 32, full.height() as i32 / 2 + 32);
-    if let Some(l) = doc.layers.iter_mut().find(|l| l.name == "blob") {
-        l.surface_mut().unwrap().fill_rect(blob_dab, &[0.9, 0.1, 0.1, 1.0]);
-        let r = blob_dab.inflate(margin).intersect(&full);
-        report("dab on the effect layer (damage + reach)", r, cpu(&doc, r), gt(&doc, r));
-    }
+    let (w, h) = (full.width() as i32, full.height() as i32);
+    case(&mut doc, "dab on a plain layer (damage + reach)", reps, &mut |d, i| {
+        let r = Rect::new(w / 3 + i as i32 * 70, h / 3, w / 3 + i as i32 * 70 + 64, h / 3 + 64);
+        if let Some(l) = d.layers.iter_mut().find(|l| l.name == "paint") {
+            l.surface_mut().unwrap().fill_rect(r, &[0.1, 0.9, 0.1, 1.0]);
+        }
+        r.inflate(margin)
+    }, true);
+    // Moving a text layer by whole pixels: its effect maps move with it.
+    case(&mut doc, "move a text layer 7 px (old + new bounds)", reps, &mut |d, _| {
+        let Some(l) = d.layers.iter_mut().find(|l| matches!(l.content, LayerContent::Text(_))) else { return Rect::EMPTY };
+        let LayerContent::Text(t) = &mut l.content else { return Rect::EMPTY };
+        let Some(src) = t.cache.as_ref() else { return Rect::EMPTY };
+        let b = src.content_bounds();
+        let to = Rect::new(b.x0 + 7, b.y0, b.x1 + 7, b.y1);
+        let mut moved = photocraft_raster::Surface::new(src.format());
+        moved.write_region(to, &src.read_region(b));
+        t.cache = Some(moved);
+        b.union(&to).inflate(margin)
+    }, true);
+    case(&mut doc, "dab on the effect layer (damage + reach)", reps, &mut |d, i| {
+        let r = Rect::new(w / 2 - 32 + i as i32 * 300, h / 2 - 32, w / 2 + 32 + i as i32 * 300, h / 2 + 32);
+        if let Some(l) = d.layers.iter_mut().find(|l| l.name == "blob") {
+            l.surface_mut().unwrap().fill_rect(r, &[0.9, 0.1, 0.1, 1.0]);
+        }
+        r.inflate(margin)
+    }, true);
 }
diff --git a/crates/ui-egui/src/canvas.rs b/crates/ui-egui/src/canvas.rs
index 1183e1e6..deeeff9f 100644
--- a/crates/ui-egui/src/canvas.rs
+++ b/crates/ui-egui/src/canvas.rs
@@ -179,7 +179,7 @@ fn display_doc(app: &PhotocraftApp, idx: usize) -> (std::sync::Arc<Document>, u6
 pub fn ensure_texture(app: &mut PhotocraftApp, ctx: &egui::Context, idx: usize) -> Option<(egui::TextureId, f32)> {
     let (revision, last_damage, id) = {
         let st = app.session.documents().get(idx)?;
-        (st.revision, st.last_damage, st.doc.id)
+        (st.revision, st.last_damage.map(|r| if r.is_empty() { r } else { r.inflate(effect_reach(&st.doc.layers)) }), st.doc.id)
     };
     let (doc, preview_key) = display_doc(app, idx);
     let cache = app.canvases.entry(id).or_insert(CanvasCache { revision: 0, texture: None, scale: 1.0, preview_key: 0, on_gpu: false, tex_revision: 0, tex_preview_key: 0 });
@@ -215,12 +215,30 @@ pub fn ensure_texture(app: &mut PhotocraftApp, ctx: &egui::Context, idx: usize)
     Some((cache.texture.as_ref()?.id(), cache.scale))
 }
 
+/// How far beyond an edit's damage rect the composite can change: layer effects (shadows, glows,
+/// strokes, …) on the edited layer and on the groups around it reach that far.
+fn effect_reach(layers: &[photocraft_doc::Layer]) -> i32 {
+    layers
+        .iter()
+        .filter(|l| l.visible)
+        .map(|l| {
+            let own = if photocraft_compose::effects::has_effects(l) { photocraft_compose::effects::margin(l) } else { 0 };
+            let inner = match &l.content {
+                LayerContent::Group(g) => effect_reach(&g.children),
+                _ => 0,
+            };
+            own + inner
+        })
+        .max()
+        .unwrap_or(0)
+}
+
 /// GPU path: make sure document `idx` is current in the GPU canvas. Brush strokes re-composite
 /// and upload only their damage rect; everything else re-composites the whole document.
 /// Returns false if there is no GPU canvas.
 fn ensure_gpu(app: &mut PhotocraftApp, idx: usize) -> bool {
     let Some(gpu) = app.gpu.clone() else { return false };
-    let Some((revision, last_damage, id)) = app.session.documents().get(idx).map(|st| (st.revision, st.last_damage, st.doc.id)) else { return false };
+    let Some((revision, last_damage, id)) = app.session.documents().get(idx).map(|st| (st.revision, st.last_damage.map(|r| if r.is_empty() { r } else { r.inflate(effect_reach(&st.doc.layers)) }), st.doc.id)) else { return false };
     let (doc, preview_key) = display_doc(app, idx);
     let size = [doc.size.width, doc.size.height];
     let cache = app.canvases.entry(id).or_insert(CanvasCache { revision: 0, texture: None, scale: 1.0, preview_key: 0, on_gpu: false, tex_revision: 0, tex_preview_key: 0 });
@@ -1436,4 +1454,21 @@ mod tests {
         let p = d.px[0];
         assert!((p[0] - 1.0).abs() < 1e-6 && (p[3] - 0.25).abs() < 1e-6, "{p:?}");
     }
+
+    #[test]
+    fn damage_grows_by_nested_effect_reach() {
+        use photocraft_doc::{Effect, Layer};
+        let fmt = photocraft_color::PixelFormat::RGBA8;
+        assert_eq!(effect_reach(&[Layer::raster("plain", fmt)]), 0);
+        let mut inner = Layer::raster("inner", fmt);
+        inner.effects.items.push(Effect::default_drop_shadow());
+        let m = photocraft_compose::effects::margin(&inner);
+        let mut group = Layer::group("g", vec![inner.clone()]);
+        group.effects.items.push(Effect::default_drop_shadow());
+        assert_eq!(effect_reach(std::slice::from_ref(&inner)), m);
+        // A child's edit moves the group's shape, whose effects reach further.
+        assert_eq!(effect_reach(&[group.clone()]), 2 * m);
+        group.visible = false;
+        assert_eq!(effect_reach(&[group]), 0);
+    }
 }
diff --git a/crates/ui-egui/src/liquify_ui.rs b/crates/ui-egui/src/liquify_ui.rs
index 1d3e9c66..f6de33b7 100644
--- a/crates/ui-egui/src/liquify_ui.rs
+++ b/crates/ui-egui/src/liquify_ui.rs
@@ -623,7 +623,7 @@ fn draw_mesh(d: &LiquifyDialog, painter: &egui::Painter, area: ERect) {
     }
     let stroke = Stroke::new(0.75, Color32::from_rgba_unmultiplied(90, 90, 90, 210));
     let pos = |i: usize, j: usize| [x0 + i as f64 * h, y0 + j as f64 * h];
-    for axis in 0..2 {
+    for axis in [0usize, 1] {
         let origin = if axis == 0 { x0 } else { y0 };
         for jj in 0..ny - 1 {
             for ii in 0..nx - 1 {
diff --git a/docs/development.md b/docs/development.md
index daced0c2..8649d9eb 100644
--- a/docs/development.md
+++ b/docs/development.md
@@ -25,7 +25,9 @@ Image code is slow at `opt-level 0`, so the workspace profile builds dependencie
 | `PHOTOCRAFT_CONTROL_PORT` | Same as `--control <port>` |
 | `PHOTOCRAFT_CPU_CANVAS=1` | Force the CPU canvas path instead of the wgpu shader canvas |
 | `PHOTOCRAFT_GPU_TILE=2048` | Force GPU canvas tiling (tests tile seams) |
-| `PHOTOCRAFT_FX_NOCACHE=1` | Bypass the layer-effect map cache (`compose::effect_maps`) |
+| `PHOTOCRAFT_FX_NOCACHE=1` | Bypass the CPU layer-effect map cache (`compose::effect_maps`) |
+| `PHOTOCRAFT_CPU_COMPOSE=1` | Keep the wgpu canvas but composite on the CPU (compare GPU vs CPU renders, e.g. with the snapshot example) |
+| `PHOTOCRAFT_FX_TRACE=1` | Print the CPU time spent on GPU effect shapes and distance fields per rebuild |
 | `PHOTOCRAFT_THEME_FILE=tokens.json` | **Debug builds only:** live design-token overrides, re-read on change |
 
 ### Live design tokens
@@ -164,8 +166,11 @@ the live command registry (`menus::is_live`) and rewrites [`docs/parity.md`](par
 ## Performance notes
 
 - The canvas is presented by a custom WGSL shader (`ui-egui/src/gpu_canvas.rs`): mip-mapped/nearest sampling, procedural checkerboard, pixel grid, tiling. Brush strokes upload only their damage rect.
-- **The canvas composites on the GPU** (`photocraft-gpu`, driven from `gpu_canvas.rs`). Anything the GPU planner can't express (layer effects, layers clipped to pass-through groups, documents over the texture limit) returns `Unsupported` and falls back to the CPU compositor (`photocraft-compose`), which is also the reference for export and tests.
-- **Layer effects** are expensive. `compose::effect_maps` caches shadow, glow, bevel and satin maps per layer state (LRU, 768 MB budget).
+- **The canvas composites on the GPU** (`photocraft-gpu`, driven from `gpu_canvas.rs`), layer effects included. What the planner can't express returns `Unsupported` and the canvas falls back to the CPU compositor (`photocraft-compose`, also the reference for export and tests): vector masks, layers clipped to pass-through groups, stroked shapes with clipped layers, pattern fills without Photoshop's cached pixels, artboards, and documents or effect regions over the texture limit. `ui.inspect` → `perf.timings.gpuFallback` names the reason (`null` on the GPU path).
+- **Layer effects on the GPU** (`gpu/src/fx.rs`, kernels in `gpu/src/compose.wgsl`). Every enabled effect becomes a *map program* over the layer's effect region (shift, dilate, Gaussian blur, glow ramp, bevel height and shading, contour, stroke band), mirroring `compose::effects` step by step; the chunked composite then paints through the maps, clipped to that region, and copies the result back into the backdrop in place, so a small text layer costs only its own pixels. The layer's shape (`compose::layer_shape`) and its distance fields (`compose::effects::distance_field`: a sequential transform whose tie-breaking a parallel GPU pass can't reproduce bit for bit) come from compose on the CPU, computed in parallel bands. Everything is cached per layer state: an unrelated edit, or an effect's colour or opacity, rebuilds nothing; a brush dab recomputes the touched 256² tiles grown by the effect reach; changing one effect's geometry rebuilds that effect only. Cache budget `gpu::FX_BUDGET` (1.5 GB, least recently drawn layers evicted first). `PHOTOCRAFT_FX_TRACE=1` prints the CPU time of each rebuild.
+- The canvas grows a stroke's damage rect by the effect reach of the layers around it (`canvas::effect_reach`), so effects beyond the dab refresh too (on both paths).
+- **Numbers** (7360 × 4912, 8 text layers + one painted layer with drop shadow + stroke + bevel, Hue/Saturation on top; M4 Pro; `cargo run --release -p photocraft-ui-egui --example fx_bench`): FX_NUMBERS. `fx_bench --compare corpus/psd/…/*.psd` reports the GPU vs CPU difference on real files.
+- On the CPU path `compose::effect_maps` caches shadow, glow, bevel and satin maps per layer state (LRU, 768 MB budget).
 - Live adjustment previews on large documents use a downsampled proxy (`ui-egui/src/proxy.rs`).
 - `ui.inspect` returns `perf` timings (UI ms per frame, composite ms, upload ms).
 - Never scan full surfaces per frame. Cache per document revision (`PhotocraftApp::cached_bounds`). An uncached `content_bounds()` on a 36 MP layer once cost 77 ms per frame.
@@ -220,8 +225,12 @@ corpus PSD with Photoshop's own merged image (PASS ≤ 2/255). To dig into one f
 cargo run --release -p photocraft-io --example oracle_diff -- corpus/psd/<file>.psd 0 png /tmp/diff.png
 ```
 
-writes ours | Photoshop | a diff heatmap side by side; `col`, `row` and `DUMP_FX=1` print samples and raw
+writes ours | Photoshop | a diff heatmap side by side; `col`, `row`, `worst [n]`, `grid x0 y0 x1 y1 [ch]`,
+`layerpx x y` and `DUMP_FX=1` print samples, the worst pixels, value grids, per-layer pixels and raw
 effect descriptors. Findings so far: fill-layer gradients are framed by the layer's mask bounds;
 Photoshop's gradient Smoothness is a Catmull-Rom blend and "Perceptual" interpolation is Oklab
 (baked into dense stops on import, `crates/io/src/gradient_bake.rs`); a shape layer's vector
-stroke is drawn above its clipped layers.
+stroke is drawn above its clipped layers; linked effect patterns tile from the layer's `fxrp`
+reference point; stroke distances follow a 5 × 5 chamfer metric (1, √2, √5) seeded at sub-pixel
+edge offsets; interior effects keep the layer's alpha; outside strokes blend onto the backdrop with
+their own modes, an upper stroke covering lower ones.
diff --git a/docs/roadmap.md b/docs/roadmap.md
index c701aa72..54756799 100644
--- a/docs/roadmap.md
+++ b/docs/roadmap.md
@@ -13,11 +13,11 @@ composite oracle and the test count.
 | M2 PSD v1 | ✅ | photocraft-psd: 134/135 real files byte-exact round trip |
 | M3 Viewer app | ✅ | egui shell (Pro / Studio / Classic themes), 13+ codecs, native and web (trunk) builds |
 | M4 Native format + engine | ✅ | 500+ commands, `.pcraft` (incremental, autosave, crash recovery), CLI, persistent preferences |
-| M5 GPU compositor | 🟡 | wgpu compositor drives the canvas; layer effects, pattern fills and a few clip cases fall back to the CPU |
+| M5 GPU compositor | 🟡 | wgpu compositor drives the canvas, layer effects included (≤1/255 vs CPU); vector masks, uncached pattern fills, artboards and a few clip cases fall back to the CPU |
 | M6 Paint + select | 🟡 | brush engine, all selection tools, multi-layer selection, snapping + smart guides, free transform + warp; native tablet backends pending |
 | M7 Adjust + filters | 🟡 | 16 adjustment layers + destructive-only adjustments, 70+ filters incl. Blur Gallery, Actions record/replay, Fade |
-| M8 PSD v2 | 🟡 | adjustments (incl. Selective Color, Color Lookup), fills, effects, patterns, text, shapes, smart objects, alpha channels; oracle 102/170 |
-| M9 Text, vector, styles | 🟡 | type engine + Warp Text, shapes / pen / paths, all 10 effects on CPU; GPU effects pending |
+| M8 PSD v2 | 🟡 | adjustments (incl. Selective Color, Color Lookup), fills, effects, patterns, text, shapes, smart objects, alpha channels; oracle 111/170 |
+| M9 Text, vector, styles | 🟡 | type engine + Warp Text, shapes / pen / paths, all 10 effects on CPU and GPU (parity ≤1/255, 30/31 corpus effect files on the GPU) |
 | M10 Smart features | 🟡 | classical Select Subject / Object, content-aware fill and scale, healing, auto-align / auto-blend; ML backend not started |
 | M11 Automation + formats | 🟡 | MCP (headless + live bridge), batch, Image Processor, prefs over MCP; DoD test passes (10 agent tasks over MCP, `automation/tests/agent_tasks.rs`); JP2 / DICOM / DPX / C2PA pending |
 | M12 Pro parity | 🟡 | CMYK / Lab / Indexed / Bitmap / Duotone, ICC + soft proofing, channels + Quick Mask, smart-object stack modes; print, HDR, photomerge, timeline, artboards, layer comps pending |
@@ -30,8 +30,9 @@ Landed on 2026-10-01: multi-layer selection, live smart objects + smart filters,
 Quick Mask, patterns, Warp, preferences, snapping, ~33 filters, the remaining core adjustments.
 
 Next:
-1. **Fidelity**: the PSD oracle (102/170). Linked-pattern phase origin, layer effects (bevel,
-   satin, multi-stroke), grayscale adjustments; PSD link groups; Photoshop smart filters in `SoLd`.
+1. **Fidelity**: the PSD oracle (111/170). Bevel/emboss, satin, inner glow; modern
+   Brightness/Contrast and grayscale Levels curves; channel restrictions (`brst`); Lab-space
+   blending; Photoshop smart filters in `SoLd`.
 2. **Layer Comps and Artboards** (File/Layer/View items and export depend on them).
 3. **GPU layer effects** (the last big CPU fallback on the canvas).
 4. **Puppet Warp, Perspective Warp, Liquify, Vanishing Point, Camera Raw / Lens Correction**.