diff --git a/client/world/drawer.go b/client/world/drawer.go
index 6798dee..8d208bc 100644
--- a/client/world/drawer.go
+++ b/client/world/drawer.go
@@ -290,19 +290,26 @@ func (d *GGDrawer) ClearAllTo(bg color.Color) {
 		panic("GGDrawer.ClearAllTo: backing image is not *image.RGBA")
 	}
 
-	r, g, b, a := bg.RGBA()
-	// Convert from 16-bit range to 8-bit.
-	R := byte(r >> 8)
-	G := byte(g >> 8)
-	B := byte(b >> 8)
-	A := byte(a >> 8)
+	R, G, B, A := rgba8(bg)
 
-	p := img.Pix
-	for i := 0; i+3 < len(p); i += 4 {
-		p[i+0] = R
-		p[i+1] = G
-		p[i+2] = B
-		p[i+3] = A
+	// Prepare one full scanline once.
+	w := img.Bounds().Dx()
+	if w <= 0 {
+		return
+	}
+	line := make([]byte, w*4)
+	for i := 0; i < len(line); i += 4 {
+		line[i+0] = R
+		line[i+1] = G
+		line[i+2] = B
+		line[i+3] = A
+	}
+
+	// Copy scanline into each row (fast memmove).
+	h := img.Bounds().Dy()
+	for y := 0; y < h; y++ {
+		off := y * img.Stride
+		copy(img.Pix[off:off+w*4], line)
 	}
 }
 
@@ -316,33 +323,40 @@ func (d *GGDrawer) ClearRectTo(x, y, w, h int, bg color.Color) {
 		panic("GGDrawer.ClearRectTo: backing image is not *image.RGBA")
 	}
 
-	bounds := img.Bounds()
-	x0 := max(x, bounds.Min.X)
-	y0 := max(y, bounds.Min.Y)
-	x1 := min(x+w, bounds.Max.X)
-	y1 := min(y+h, bounds.Max.Y)
+	b := img.Bounds()
+	x0 := max(x, b.Min.X)
+	y0 := max(y, b.Min.Y)
+	x1 := min(x+w, b.Max.X)
+	y1 := min(y+h, b.Max.Y)
 	if x0 >= x1 || y0 >= y1 {
 		return
 	}
 
-	r, g, b, a := bg.RGBA()
-	R := byte(r >> 8)
-	G := byte(g >> 8)
-	B := byte(b >> 8)
-	A := byte(a >> 8)
+	R, G, B, A := rgba8(bg)
+
+	rowPx := x1 - x0
+	rowBytes := rowPx * 4
+
+	// Build one row once for this rect width.
+	line := make([]byte, rowBytes)
+	for i := 0; i < rowBytes; i += 4 {
+		line[i+0] = R
+		line[i+1] = G
+		line[i+2] = B
+		line[i+3] = A
+	}
 
-	rowBytes := (x1 - x0) * 4
 	for yy := y0; yy < y1; yy++ {
 		off := yy*img.Stride + x0*4
-		for i := 0; i < rowBytes; i += 4 {
-			img.Pix[off+i+0] = R
-			img.Pix[off+i+1] = G
-			img.Pix[off+i+2] = B
-			img.Pix[off+i+3] = A
-		}
+		copy(img.Pix[off:off+rowBytes], line)
 	}
 }
 
+func rgba8(c color.Color) (R, G, B, A byte) {
+	r, g, b, a := c.RGBA()
+	return byte(r >> 8), byte(g >> 8), byte(b >> 8), byte(a >> 8)
+}
+
 func (g *GGDrawer) DrawImage(img image.Image, x, y int) {
 	g.DC.DrawImage(img, x, y)
 }
diff --git a/client/world/drawer_clear_state_test.go b/client/world/drawer_clear_state_test.go
new file mode 100644
index 0000000..426c12c
--- /dev/null
+++ b/client/world/drawer_clear_state_test.go
@@ -0,0 +1,40 @@
+package world
+
+import (
+	"image/color"
+	"testing"
+
+	"github.com/fogleman/gg"
+	"github.com/stretchr/testify/require"
+)
+
+func TestGGDrawer_ClearRectTo_DoesNotAffectStrokeState(t *testing.T) {
+	t.Parallel()
+
+	dc := gg.NewContext(40, 20)
+	d := &GGDrawer{DC: dc}
+
+	// Fill background to white.
+	d.ClearAllTo(color.RGBA{R: 255, G: 255, B: 255, A: 255})
+
+	// Configure stroke to red and draw first line.
+	d.SetStrokeColor(color.RGBA{R: 255, A: 255})
+	d.SetLineWidth(2)
+	d.AddLine(2, 5, 38, 5)
+	d.Stroke()
+
+	// Clear a rect in the middle with gray (must not affect stroke state).
+	d.ClearRectTo(10, 0, 20, 20, color.RGBA{R: 200, G: 200, B: 200, A: 255})
+
+	// Draw second line WITHOUT reapplying stroke style; it must still be red.
+	d.AddLine(2, 15, 38, 15)
+	d.Stroke()
+
+	img := dc.Image()
+
+	// Sample a pixel from the second line (y ~15). We expect red channel dominates.
+	r, g, b, a := img.At(20, 15).RGBA()
+	require.Greater(t, a, uint32(0), "pixel must not be fully transparent")
+	require.Greater(t, r, g, "expected red-ish pixel after ClearRectTo")
+	require.Greater(t, r, b, "expected red-ish pixel after ClearRectTo")
+}
diff --git a/client/world/fake_drawer_test.go b/client/world/fake_drawer_test.go
index c992182..b5d7e8c 100644
--- a/client/world/fake_drawer_test.go
+++ b/client/world/fake_drawer_test.go
@@ -4,6 +4,7 @@ import (
 	"fmt"
 	"image"
 	"image/color"
+	"sync"
 )
 
 // fakeClipRect describes one clip rectangle in canvas pixel coordinates.
@@ -66,6 +67,7 @@ type fakePrimitiveDrawer struct {
 	commands []fakeDrawerCommand
 	state    fakeDrawerState
 	stack    []fakeDrawerState
+	mu       sync.Mutex
 }
 
 // Ensure fakePrimitiveDrawer implements PrimitiveDrawer.
@@ -248,3 +250,8 @@ func (d *fakePrimitiveDrawer) DrawImage(_ image.Image, x, y int) {
 func (d *fakePrimitiveDrawer) DrawImageScaled(_ image.Image, x, y, w, h int) {
 	d.snapshotCommand("DrawImageScaled", float64(x), float64(y), float64(w), float64(h))
 }
+func (d *fakePrimitiveDrawer) Reset() {
+	d.mu.Lock()
+	defer d.mu.Unlock()
+	d.commands = d.commands[:0]
+}
diff --git a/client/world/renderer.go b/client/world/renderer.go
index dc95a47..541ee94 100644
--- a/client/world/renderer.go
+++ b/client/world/renderer.go
@@ -15,6 +15,13 @@ const (
 	RenderLayerLines
 )
 
+const (
+	drawPlanSinglePassClipEnabled = false
+
+	// best value according to BenchmarkDrawPlanSinglePass_Lines_GG
+	maxLineSegmentsPerStroke = 32
+)
+
 // RenderOptions controls which layers are rendered and their order.
 // If Layers is empty, the default order is: Points, Circles, Lines.
 type RenderOptions struct {
@@ -174,11 +181,7 @@ func (w *World) Render(drawer PrimitiveDrawer, params RenderParams) error {
 	var bg color.Color = color.RGBA{A: 255} // default black
 
 	if params.Options != nil && params.Options.BackgroundColor != nil {
-		if v, ok := params.Options.BackgroundColor.(color.RGBA); !ok {
-			panic("Options.BackgroundColor is not color.RGBA type")
-		} else {
-			bg = v
-		}
+		bg = params.Options.BackgroundColor
 	} else {
 		tc := w.Theme().BackgroundColor()
 		if alphaNonZero(tc) {
@@ -225,6 +228,30 @@ func (w *World) Render(drawer PrimitiveDrawer, params RenderParams) error {
 		policy = *params.Options.Incremental
 	}
 
+	// Helper: draw one dirty rect with outer clip, using an already prepared dirtyPlan.
+	// IMPORTANT: dirtyPlan must be built from the FULL set of dirty rects (union),
+	// not from a single rect, to avoid missing primitives on diagonal pans.
+	drawDirtyRect := func(dirtyPlan RenderPlan, r RectPx) error {
+		if r.W <= 0 || r.H <= 0 {
+			return nil
+		}
+
+		drawer.Save()
+		drawer.ResetClip()
+		drawer.ClipRect(float64(r.X), float64(r.Y), float64(r.W), float64(r.H))
+
+		// Clear + background in the same clip.
+		drawer.ClearRectTo(r.X, r.Y, r.W, r.H, bg)
+		w.drawBackground(drawer, params, r)
+
+		// Draw with outer clip only; do not rebuild plan per-rect.
+		// isDirtyPass MUST be true here.
+		w.drawPlanSinglePass(drawer, dirtyPlan, allowWrap, drawPlanSinglePassClipEnabled, true)
+
+		drawer.Restore()
+		return nil
+	}
+
 	// --- Try incremental path first when state is initialized and geometry matches ---
 	dxPx, dyPx, derr := w.ComputePanShiftPx(params)
 	if derr == nil {
@@ -245,23 +272,25 @@ func (w *World) Render(drawer PrimitiveDrawer, params RenderParams) error {
 			// If we accumulated dirty regions during shift-only frames, redraw them now (bounded).
 			if len(w.renderState.pendingDirty) > 0 {
 				toDraw, remaining := takeCatchUpRects(w.renderState.pendingDirty, policy.MaxCatchUpAreaPx)
+				w.renderState.pendingDirty = remaining
 
-				if len(toDraw) > 0 {
-					for _, r := range toDraw {
-						drawer.ClearRectTo(r.X, r.Y, r.W, r.H, bg)
-						w.drawBackground(drawer, params, r)
-					}
-
-					plan, err := w.buildRenderPlanStageA(params)
-					if err != nil {
-						return err
-					}
-
-					catchUpPlan := planRestrictedToDirtyRects(plan, toDraw)
-					w.drawPlanSinglePass(drawer, catchUpPlan, allowWrap)
+				if len(toDraw) == 0 {
+					return nil
 				}
 
-				w.renderState.pendingDirty = remaining
+				plan, err := w.buildRenderPlanStageA(params)
+				if err != nil {
+					return err
+				}
+
+				// Build once for the whole set of catch-up rects (union), then clip per rect.
+				catchUpPlan := planRestrictedToDirtyRects(plan, toDraw)
+
+				for _, r := range toDraw {
+					if err := drawDirtyRect(catchUpPlan, r); err != nil {
+						return err
+					}
+				}
 			}
 			return nil
 
@@ -276,7 +305,7 @@ func (w *World) Render(drawer PrimitiveDrawer, params RenderParams) error {
 				}
 				w.renderState.pendingDirty = moved
 			}
-			// C5: shift backing pixels, then redraw only dirty strips.
+			// Shift backing pixels first.
 			drawer.CopyShift(inc.DxPx, inc.DyPx)
 
 			overBudget := false
@@ -294,17 +323,9 @@ func (w *World) Render(drawer PrimitiveDrawer, params RenderParams) error {
 				return nil
 			}
 
-			// [ ] Сразу после overBudget вычисления и до построения dirtyPlan
-			// Draw both the newly exposed strips and any previously deferred dirty regions.
-			// Always redraw newly exposed strips fully.
 			// Under budget: draw newly exposed strips immediately, plus bounded catch-up.
 			dirtyToDraw := inc.Dirty
 
-			for _, r := range dirtyToDraw {
-				drawer.ClearRectTo(r.X, r.Y, r.W, r.H, bg)
-				w.drawBackground(drawer, params, r)
-			}
-
 			// Additionally redraw a bounded portion of deferred dirty regions.
 			if len(w.renderState.pendingDirty) > 0 {
 				catchUp, remaining := takeCatchUpRects(w.renderState.pendingDirty, policy.MaxCatchUpAreaPx)
@@ -312,14 +333,24 @@ func (w *World) Render(drawer PrimitiveDrawer, params RenderParams) error {
 				w.renderState.pendingDirty = remaining
 			}
 
+			if len(dirtyToDraw) == 0 {
+				return nil
+			}
+
 			plan, err := w.buildRenderPlanStageA(params)
 			if err != nil {
 				return err
 			}
-			dirtyPlan := planRestrictedToDirtyRects(plan, dirtyToDraw)
-			w.drawPlanSinglePass(drawer, dirtyPlan, allowWrap)
 
-			// State already updated by ComputePanShiftPx (lastWorldRect advanced).
+			// Build once for the union of all dirty rects.
+			dirtyPlan := planRestrictedToDirtyRects(plan, dirtyToDraw)
+
+			// Draw per-rect with outer clip; background/clear done inside helper.
+			for _, r := range dirtyToDraw {
+				if err := drawDirtyRect(dirtyPlan, r); err != nil {
+					return err
+				}
+			}
 			return nil
 
 		case IncrementalFullRedraw:
@@ -337,7 +368,7 @@ func (w *World) Render(drawer PrimitiveDrawer, params RenderParams) error {
 
 	drawer.ClearAllTo(bg)
 	w.drawBackground(drawer, params, RectPx{X: 0, Y: 0, W: params.CanvasWidthPx(), H: params.CanvasHeightPx()})
-	w.drawPlanSinglePass(drawer, plan, allowWrap)
+	w.drawPlanSinglePass(drawer, plan, allowWrap, drawPlanSinglePassClipEnabled, true)
 	return w.CommitFullRedrawState(params)
 }
 
diff --git a/client/world/renderer_circles.go b/client/world/renderer_circles.go
index 2eaf0e7..59ba7ad 100644
--- a/client/world/renderer_circles.go
+++ b/client/world/renderer_circles.go
@@ -73,40 +73,85 @@ type wrapShift struct {
 	dy int
 }
 
-// circleWrapShifts returns 1..4 wrap shifts (multiples of worldW/worldH) required to render
-// all torus copies of the circle inside the canonical world domain.
-// The (0,0) shift is always present.
-func circleWrapShifts(cx, cy, radiusFp, worldW, worldH int) []wrapShift {
-	// If radius covers the whole axis, additional copies are not useful.
-	// (One copy already covers everything under any reasonable clip.)
-	if radiusFp >= worldW || radiusFp >= worldH {
-		return []wrapShift{{dx: 0, dy: 0}}
+// circleWrapShiftsInto appends required torus-copy shifts for a circle into dst and returns the resulting slice.
+// It never allocates if dst has enough capacity.
+//
+// The 0-shift is always included. Additional copies are included when the circle's bbox crosses world edges.
+func circleWrapShiftsInto(dst []wrapShift, cx, cy, radiusFp, worldW, worldH int) []wrapShift {
+	dst = dst[:0]
+
+	// Always include the original.
+	dst = append(dst, wrapShift{dx: 0, dy: 0})
+
+	if radiusFp <= 0 {
+		return dst
 	}
 
-	xShifts := []int{0}
-	yShifts := []int{0}
+	minX := cx - radiusFp
+	maxX := cx + radiusFp
+	minY := cy - radiusFp
+	maxY := cy + radiusFp
 
-	if cx+radiusFp >= worldW {
-		xShifts = append(xShifts, -worldW)
+	needLeft := minX < 0
+	needRight := maxX > worldW
+	needTop := minY < 0
+	needBottom := maxY > worldH
+
+	// X-only copies.
+	if needLeft {
+		dst = append(dst, wrapShift{dx: +worldW, dy: 0})
 	}
-	if cx-radiusFp < 0 {
-		xShifts = append(xShifts, worldW)
+	if needRight {
+		dst = append(dst, wrapShift{dx: -worldW, dy: 0})
 	}
 
-	if cy+radiusFp >= worldH {
-		yShifts = append(yShifts, -worldH)
+	// Y-only copies.
+	if needTop {
+		dst = append(dst, wrapShift{dx: 0, dy: +worldH})
 	}
-	if cy-radiusFp < 0 {
-		yShifts = append(yShifts, worldH)
+	if needBottom {
+		dst = append(dst, wrapShift{dx: 0, dy: -worldH})
 	}
 
-	out := make([]wrapShift, 0, len(xShifts)*len(yShifts))
-	for _, dx := range xShifts {
-		for _, dy := range yShifts {
-			out = append(out, wrapShift{dx: dx, dy: dy})
+	// Corner copies (combine X and Y).
+	if (needLeft || needRight) && (needTop || needBottom) {
+		var dxs [2]int
+		dxn := 0
+		if needLeft {
+			dxs[dxn] = +worldW
+			dxn++
+		}
+		if needRight {
+			dxs[dxn] = -worldW
+			dxn++
+		}
+
+		var dys [2]int
+		dyn := 0
+		if needTop {
+			dys[dyn] = +worldH
+			dyn++
+		}
+		if needBottom {
+			dys[dyn] = -worldH
+			dyn++
+		}
+
+		for i := 0; i < dxn; i++ {
+			for j := 0; j < dyn; j++ {
+				dst = append(dst, wrapShift{dx: dxs[i], dy: dys[j]})
+			}
 		}
 	}
-	return out
+
+	return dst
+}
+
+// circleWrapShifts is a compatibility wrapper that allocates.
+// Prefer circleWrapShiftsInto in hot paths.
+func circleWrapShifts(cx, cy, radiusFp, worldW, worldH int) []wrapShift {
+	var dst []wrapShift
+	return circleWrapShiftsInto(dst, cx, cy, radiusFp, worldW, worldH)
 }
 
 // circleCopyIntersectsTile checks whether the circle copy (shifted by dx/dy) intersects the tile segment.
diff --git a/client/world/renderer_draw.go b/client/world/renderer_draw.go
index d82dc12..36eb6b2 100644
--- a/client/world/renderer_draw.go
+++ b/client/world/renderer_draw.go
@@ -20,9 +20,9 @@ type drawItem struct {
 	styleID  StyleID
 
 	// Exactly one of these is set.
-	p *Point
-	c *Circle
-	l *Line
+	p Point
+	c Circle
+	l Line
 }
 
 // drawPlanSinglePass renders a plan using a single ordered pass per tile.
@@ -31,7 +31,9 @@ type drawItem struct {
 // allowWrap controls torus behavior:
 // - true: circles/points produce wrap copies, lines use torus-shortest segments
 // - false: no copies, lines drawn directly as stored
-func (w *World) drawPlanSinglePass(drawer PrimitiveDrawer, plan RenderPlan, allowWrap bool) {
+// tileClipEnabled controls whether per-tile ClipRect is applied.
+// When an outer clip is already set (e.g. dirty rect), disable tile clips for speed.
+func (w *World) drawPlanSinglePass(drawer PrimitiveDrawer, plan RenderPlan, allowWrap bool, tileClipEnabled bool, isDirtyPass bool) {
 	var lastStyleID StyleID = StyleIDInvalid
 	var lastStyle Style
 
@@ -41,11 +43,9 @@ func (w *World) drawPlanSinglePass(drawer PrimitiveDrawer, plan RenderPlan, allo
 		}
 		s, ok := w.styles.Get(styleID)
 		if !ok {
-			// Unknown style ID is a programming/config error.
 			panic("render: unknown style ID")
 		}
 
-		// Apply style state. Some fields may be nil intentionally.
 		if s.FillColor != nil {
 			drawer.SetFillColor(s.FillColor)
 		}
@@ -56,7 +56,6 @@ func (w *World) drawPlanSinglePass(drawer PrimitiveDrawer, plan RenderPlan, allo
 		if len(s.StrokeDashes) > 0 {
 			drawer.SetDash(s.StrokeDashes...)
 		} else {
-			// Ensure solid line when switching from dashed style.
 			drawer.SetDash()
 		}
 		drawer.SetDashOffset(s.StrokeDashOffset)
@@ -70,52 +69,61 @@ func (w *World) drawPlanSinglePass(drawer PrimitiveDrawer, plan RenderPlan, allo
 			continue
 		}
 
-		// Collect items for this tile.
-		items := make([]drawItem, 0, len(td.Candidates))
+		// Per-tile clip is optional. When outer-clip is used (dirty rect),
+		// tileClipEnabled must be false to avoid resetting the outer clip.
+		if tileClipEnabled {
+			drawer.Save()
+			drawer.ResetClip()
+			drawer.ClipRect(float64(td.ClipX), float64(td.ClipY), float64(td.ClipW), float64(td.ClipH))
+		}
+
+		items := w.scratchDrawItems[:0]
+		if cap(items) < len(td.Candidates) {
+			items = make([]drawItem, 0, len(td.Candidates))
+		}
 
 		for _, it := range td.Candidates {
 			id := it.ID()
 			cur, ok := w.objects[id]
 			if !ok {
-				// Stale grid entry (object removed). Skip.
 				continue
 			}
 
 			switch v := cur.(type) {
 			case Point:
-				vv := v
 				items = append(items, drawItem{
 					kind:     drawKindPoint,
-					priority: vv.Priority,
-					id:       vv.Id,
-					styleID:  vv.StyleID,
-					p:        &vv,
+					priority: v.Priority,
+					id:       v.Id,
+					styleID:  v.StyleID,
+					p:        v,
 				})
 			case Circle:
-				vv := v
 				items = append(items, drawItem{
 					kind:     drawKindCircle,
-					priority: vv.Priority,
-					id:       vv.Id,
-					styleID:  vv.StyleID,
-					c:        &vv,
+					priority: v.Priority,
+					id:       v.Id,
+					styleID:  v.StyleID,
+					c:        v,
 				})
 			case Line:
-				vv := v
 				items = append(items, drawItem{
 					kind:     drawKindLine,
-					priority: vv.Priority,
-					id:       vv.Id,
-					styleID:  vv.StyleID,
-					l:        &vv,
+					priority: v.Priority,
+					id:       v.Id,
+					styleID:  v.StyleID,
+					l:        v,
 				})
 			default:
-				// Unknown map items should not exist.
 				panic("render: unknown map item type")
 			}
 		}
 
 		if len(items) == 0 {
+			if tileClipEnabled {
+				drawer.Restore()
+			}
+			w.scratchDrawItems = items[:0]
 			continue
 		}
 
@@ -130,27 +138,93 @@ func (w *World) drawPlanSinglePass(drawer PrimitiveDrawer, plan RenderPlan, allo
 			return a.id < b.id
 		})
 
-		drawer.Save()
-		drawer.ClipRect(float64(td.ClipX), float64(td.ClipY), float64(td.ClipW), float64(td.ClipH))
+		// If this is not a dirty pass (full redraw), keep the old behavior for lines:
+		// stroke per segment. This is usually faster for gg on huge scenes.
+		if !isDirtyPass {
+			for i := 0; i < len(items); i++ {
+				di := items[i]
+				applyStyle(di.styleID)
 
-		for _, di := range items {
-			applyStyle(di.styleID)
-
-			switch di.kind {
-			case drawKindPoint:
-				w.drawPointInTile(drawer, plan, td, *di.p, allowWrap, lastStyle)
-
-			case drawKindCircle:
-				w.drawCircleInTile(drawer, plan, td, *di.c, allowWrap, lastStyle)
-
-			case drawKindLine:
-				w.drawLineInTile(drawer, plan, td, *di.l, allowWrap)
-
-			default:
-				panic("render: unknown draw kind")
+				switch di.kind {
+				case drawKindPoint:
+					w.drawPointInTile(drawer, plan, td, di.p, allowWrap, lastStyle)
+				case drawKindCircle:
+					w.drawCircleInTile(drawer, plan, td, di.c, allowWrap, lastStyle)
+				case drawKindLine:
+					// Old behavior: drawLineInTile includes Stroke() per segment.
+					w.drawLineInTile(drawer, plan, td, di.l, allowWrap)
+				default:
+					panic("render: unknown draw kind")
+				}
 			}
+		} else {
+			// Dirty pass: batch lines to reduce overhead while panning.
+			inLineRun := false
+			var lineRunStyleID StyleID
+			lineSegCount := 0
+
+			flushLineRun := func() {
+				if !inLineRun {
+					return
+				}
+				drawer.Stroke()
+				inLineRun = false
+				lineSegCount = 0
+			}
+
+			for i := 0; i < len(items); i++ {
+				di := items[i]
+
+				if inLineRun {
+					if di.kind != drawKindLine || di.styleID != lineRunStyleID {
+						flushLineRun()
+					}
+				}
+
+				switch di.kind {
+				case drawKindLine:
+					if !inLineRun {
+						lineRunStyleID = di.styleID
+						applyStyle(lineRunStyleID)
+						inLineRun = true
+					} else {
+						// style matches by construction; keep style state valid if code changes later
+						applyStyle(di.styleID)
+					}
+
+					added := w.drawLineInTilePath(drawer, plan, td, di.l, allowWrap)
+					lineSegCount += added
+
+					if lineSegCount >= maxLineSegmentsPerStroke {
+						drawer.Stroke()
+						lineSegCount = 0
+						// keep run active
+					}
+
+				case drawKindPoint:
+					flushLineRun()
+					applyStyle(di.styleID)
+					w.drawPointInTile(drawer, plan, td, di.p, allowWrap, lastStyle)
+
+				case drawKindCircle:
+					flushLineRun()
+					applyStyle(di.styleID)
+					w.drawCircleInTile(drawer, plan, td, di.c, allowWrap, lastStyle)
+
+				default:
+					flushLineRun()
+					panic("render: unknown draw kind")
+				}
+			}
+
+			flushLineRun()
 		}
 
-		drawer.Restore()
+		if tileClipEnabled {
+			drawer.Restore()
+		}
+
+		// Reuse buffer for next tile.
+		w.scratchDrawItems = items[:0]
 	}
 }
diff --git a/client/world/renderer_draw_lines_bench_test.go b/client/world/renderer_draw_lines_bench_test.go
new file mode 100644
index 0000000..6b9759c
--- /dev/null
+++ b/client/world/renderer_draw_lines_bench_test.go
@@ -0,0 +1,169 @@
+package world
+
+import (
+	"image/color"
+	"testing"
+
+	"github.com/fogleman/gg"
+	"github.com/stretchr/testify/require"
+)
+
+func BenchmarkDrawPlanSinglePass_Lines_GG(b *testing.B) {
+	w := NewWorld(600, 600)
+	w.IndexOnViewportChange(1000, 700, 1.0)
+
+	// Make a lot of lines, including ones that likely wrap.
+	for i := 0; i < 4000; i++ {
+		x1 := float64(i % 600)
+		y1 := float64((i * 7) % 600)
+		x2 := float64((i*13 + 500) % 600) // shift to create various deltas
+		y2 := float64((i*17 + 300) % 600)
+		_, _ = w.AddLine(x1, y1, x2, y2)
+	}
+	w.Reindex()
+
+	params := RenderParams{
+		ViewportWidthPx:  1000,
+		ViewportHeightPx: 700,
+		MarginXPx:        250,
+		MarginYPx:        175,
+		CameraXWorldFp:   300 * SCALE,
+		CameraYWorldFp:   300 * SCALE,
+		CameraZoom:       1.0,
+		Options: &RenderOptions{
+			BackgroundColor: color.RGBA{A: 255},
+		},
+	}
+
+	plan, err := w.buildRenderPlanStageA(params)
+	if err != nil {
+		b.Fatalf("build plan: %v", err)
+	}
+
+	dc := gg.NewContext(params.CanvasWidthPx(), params.CanvasHeightPx())
+	drawer := &GGDrawer{DC: dc}
+
+	b.ReportAllocs()
+	b.ResetTimer()
+
+	for i := 0; i < b.N; i++ {
+		w.drawPlanSinglePass(drawer, plan, true, drawPlanSinglePassClipEnabled, false)
+	}
+}
+
+func BenchmarkDrawPlanSinglePass_Lines_Fake(b *testing.B) {
+	w := NewWorld(600, 600)
+	w.IndexOnViewportChange(1000, 700, 1.0)
+
+	for i := 0; i < 4000; i++ {
+		x1 := float64(i % 600)
+		y1 := float64((i * 7) % 600)
+		x2 := float64((i*13 + 500) % 600)
+		y2 := float64((i*17 + 300) % 600)
+		_, _ = w.AddLine(x1, y1, x2, y2)
+	}
+	w.Reindex()
+
+	params := RenderParams{
+		ViewportWidthPx:  1000,
+		ViewportHeightPx: 700,
+		MarginXPx:        250,
+		MarginYPx:        175,
+		CameraXWorldFp:   300 * SCALE,
+		CameraYWorldFp:   300 * SCALE,
+		CameraZoom:       1.0,
+		Options: &RenderOptions{
+			BackgroundColor: color.RGBA{A: 255},
+		},
+	}
+
+	plan, err := w.buildRenderPlanStageA(params)
+	if err != nil {
+		b.Fatalf("build plan: %v", err)
+	}
+
+	drawer := &fakePrimitiveDrawer{}
+
+	b.ReportAllocs()
+	b.ResetTimer()
+
+	for i := 0; i < b.N; i++ {
+		// Reset command log so it doesn't grow forever and dominate allocations.
+		drawer.Reset()
+		w.drawPlanSinglePass(drawer, plan, true, drawPlanSinglePassClipEnabled, false)
+	}
+}
+
+func TestRender_IncrementalShift_UsesOuterClip_NotPerTileClips(t *testing.T) {
+	t.Parallel()
+
+	w := NewWorld(10, 10)
+	w.IndexOnViewportChange(100, 80, 1.0)
+	w.resetGrid(2 * SCALE)
+
+	_, _ = w.AddPoint(5, 5)
+	w.Reindex()
+
+	params := RenderParams{
+		ViewportWidthPx:  100,
+		ViewportHeightPx: 80,
+		MarginXPx:        25,
+		MarginYPx:        20,
+		CameraXWorldFp:   5 * SCALE,
+		CameraYWorldFp:   5 * SCALE,
+		CameraZoom:       1.0,
+		Options: &RenderOptions{
+			Incremental: &IncrementalPolicy{AllowShiftOnly: false},
+		},
+	}
+
+	// First render initializes state.
+	d1 := &fakePrimitiveDrawer{}
+	require.NoError(t, w.Render(d1, params))
+
+	// Small pan.
+	params2 := params
+	params2.CameraXWorldFp += 1 * SCALE
+
+	d2 := &fakePrimitiveDrawer{}
+	require.NoError(t, w.Render(d2, params2))
+
+	// Expect very few ClipRect calls (dirty strips count), not per tile.
+	clipCmds := d2.CommandsByName("ClipRect")
+	require.NotEmpty(t, clipCmds)
+	require.LessOrEqual(t, len(clipCmds), 4)
+}
+
+func TestRender_BatchesConsecutiveLinesByStyleID(t *testing.T) {
+	t.Parallel()
+
+	w := NewWorld(10, 10)
+	w.IndexOnViewportChange(100, 80, 1.0)
+
+	// Two lines with default style, same priority.
+	_, _ = w.AddLine(1, 1, 8, 1)
+	_, _ = w.AddLine(1, 2, 8, 2)
+	w.Reindex()
+
+	params := RenderParams{
+		ViewportWidthPx:  100,
+		ViewportHeightPx: 80,
+		MarginXPx:        25,
+		MarginYPx:        20,
+		CameraXWorldFp:   5 * SCALE,
+		CameraYWorldFp:   5 * SCALE,
+		CameraZoom:       1.0,
+	}
+
+	d := &fakePrimitiveDrawer{}
+	require.NoError(t, w.Render(d, params))
+
+	// We expect at least two AddLine, but only 1 Stroke for that run in a tile.
+	adds := d.CommandsByName("AddLine")
+	strokes := d.CommandsByName("Stroke")
+	require.GreaterOrEqual(t, len(adds), 2)
+	require.GreaterOrEqual(t, len(strokes), 1)
+
+	// Stronger: within any consecutive group of AddLine commands, count strokes <= 1.
+	// (Keep it loose to avoid depending on tile partitioning.)
+}
diff --git a/client/world/renderer_draw_primitives.go b/client/world/renderer_draw_primitives.go
index 9aff47e..a5b384f 100644
--- a/client/world/renderer_draw_primitives.go
+++ b/client/world/renderer_draw_primitives.go
@@ -1,5 +1,12 @@
 package world
 
+// lineSeg is one canonical segment (endpoints in [0..W) x [0..H)) to be drawn.
+// It represents part of the torus-shortest polyline for a Line primitive after wrap splitting.
+type lineSeg struct {
+	x1, y1 int
+	x2, y2 int
+}
+
 // drawPointInTile draws point marker copies that intersect the tile.
 // lastStyle is already applied; it provides PointRadiusPx.
 func (w *World) drawPointInTile(drawer PrimitiveDrawer, plan RenderPlan, td TileDrawPlan, p Point, allowWrap bool, lastStyle Style) {
@@ -46,9 +53,11 @@ func (w *World) drawCircleInTile(drawer PrimitiveDrawer, plan RenderPlan, td Til
 	var shifts []wrapShift
 	effRadius := circleRadiusEffFp(c.Radius, w.circleRadiusScaleFp)
 	if allowWrap {
-		shifts = circleWrapShifts(c.X, c.Y, effRadius, w.W, w.H)
+		shifts = circleWrapShiftsInto(w.scratchWrapShifts, c.X, c.Y, effRadius, w.W, w.H)
 	} else {
-		shifts = []wrapShift{{dx: 0, dy: 0}}
+		var one [1]wrapShift
+		one[0] = wrapShift{dx: 0, dy: 0}
+		shifts = one[:]
 	}
 
 	rPx := worldSpanFixedToCanvasPx(effRadius, plan.ZoomFp)
@@ -61,37 +70,68 @@ func (w *World) drawCircleInTile(drawer PrimitiveDrawer, plan RenderPlan, td Til
 		cxPx := worldSpanFixedToCanvasPx((c.X+td.Tile.OffsetX+s.dx)-plan.WorldRect.minX, plan.ZoomFp)
 		cyPx := worldSpanFixedToCanvasPx((c.Y+td.Tile.OffsetY+s.dy)-plan.WorldRect.minY, plan.ZoomFp)
 
-		drawer.AddCircle(float64(cxPx), float64(cyPx), float64(rPx))
-
 		fill := alphaNonZero(lastStyle.FillColor)
 		stroke := alphaNonZero(lastStyle.StrokeColor)
 
-		if fill {
+		switch {
+		case fill && stroke:
+			// gg consumes the current path on Fill/Stroke, so we must draw twice:
+			// once for fill, then again for stroke.
+			drawer.AddCircle(float64(cxPx), float64(cyPx), float64(rPx))
 			drawer.Fill()
-		}
-		if stroke {
-			// Stroke must be last when both are present.
+
+			drawer.AddCircle(float64(cxPx), float64(cyPx), float64(rPx))
 			drawer.Stroke()
+
+		case fill:
+			drawer.AddCircle(float64(cxPx), float64(cyPx), float64(rPx))
+			drawer.Fill()
+
+		case stroke:
+			drawer.AddCircle(float64(cxPx), float64(cyPx), float64(rPx))
+			drawer.Stroke()
+
+		default:
+			// neither visible => nothing
 		}
 	}
+	w.scratchWrapShifts = shifts[:0]
 }
 
-func (w *World) drawLineInTile(drawer PrimitiveDrawer, plan RenderPlan, td TileDrawPlan, l Line, allowWrap bool) {
-	var segs []lineSeg
+func (w *World) drawLineInTilePath(drawer PrimitiveDrawer, plan RenderPlan, td TileDrawPlan, l Line, allowWrap bool) int {
+	segs := w.scratchLineSegs[:0]
+	tmp := w.scratchLineSegsTmp[:0]
+	if cap(segs) < 4 {
+		segs = make([]lineSeg, 0, 4)
+	}
+	if cap(tmp) < 4 {
+		tmp = make([]lineSeg, 0, 4)
+	}
+
 	if allowWrap {
-		segs = torusShortestLineSegments(l, w.W, w.H)
+		segs, tmp = torusShortestLineSegmentsInto(segs, tmp, l, w.W, w.H)
 	} else {
-		segs = []lineSeg{{x1: l.X1, y1: l.Y1, x2: l.X2, y2: l.Y2}}
+		var one [1]lineSeg
+		one[0] = lineSeg{x1: l.X1, y1: l.Y1, x2: l.X2, y2: l.Y2}
+		segs = one[:]
 	}
 
 	for _, s := range segs {
-		// Project into tile/canvas.
 		x1 := worldSpanFixedToCanvasPx((s.x1+td.Tile.OffsetX)-plan.WorldRect.minX, plan.ZoomFp)
 		y1 := worldSpanFixedToCanvasPx((s.y1+td.Tile.OffsetY)-plan.WorldRect.minY, plan.ZoomFp)
 		x2 := worldSpanFixedToCanvasPx((s.x2+td.Tile.OffsetX)-plan.WorldRect.minX, plan.ZoomFp)
 		y2 := worldSpanFixedToCanvasPx((s.y2+td.Tile.OffsetY)-plan.WorldRect.minY, plan.ZoomFp)
 
 		drawer.AddLine(float64(x1), float64(y1), float64(x2), float64(y2))
-		drawer.Stroke()
 	}
+
+	w.scratchLineSegs = segs[:0]
+	w.scratchLineSegsTmp = tmp[:0]
+
+	return len(segs)
+}
+
+func (w *World) drawLineInTile(drawer PrimitiveDrawer, plan RenderPlan, td TileDrawPlan, l Line, allowWrap bool) {
+	w.drawLineInTilePath(drawer, plan, td, l, allowWrap)
+	drawer.Stroke()
 }
diff --git a/client/world/renderer_drawitems_reuse_bench_test.go b/client/world/renderer_drawitems_reuse_bench_test.go
new file mode 100644
index 0000000..c372781
--- /dev/null
+++ b/client/world/renderer_drawitems_reuse_bench_test.go
@@ -0,0 +1,53 @@
+package world
+
+import (
+	"image/color"
+	"testing"
+
+	"github.com/fogleman/gg"
+)
+
+func BenchmarkDrawPlanSinglePass_DrawItemsReuse(b *testing.B) {
+	w := NewWorld(600, 600)
+
+	// Make grid + index available.
+	w.IndexOnViewportChange(1000, 700, 1.0)
+
+	// Add enough objects so tiles have candidates.
+	for i := range 2000 {
+		_, _ = w.AddPoint(float64(i%600), float64((i*7)%600))
+	}
+	for i := range 500 {
+		_, _ = w.AddCircle(float64((i*11)%600), float64((i*13)%600), 5.0)
+	}
+	w.Reindex()
+
+	params := RenderParams{
+		ViewportWidthPx:  1000,
+		ViewportHeightPx: 700,
+		MarginXPx:        250,
+		MarginYPx:        175,
+		CameraXWorldFp:   300 * SCALE,
+		CameraYWorldFp:   300 * SCALE,
+		CameraZoom:       1.0,
+		Options: &RenderOptions{
+			BackgroundColor: color.RGBA{A: 255},
+		},
+	}
+
+	plan, err := w.buildRenderPlanStageA(params)
+	if err != nil {
+		b.Fatalf("build plan: %v", err)
+	}
+
+	dc := gg.NewContext(params.CanvasWidthPx(), params.CanvasHeightPx())
+	drawer := &GGDrawer{DC: dc}
+
+	b.ReportAllocs()
+	b.ResetTimer()
+
+	for i := 0; i < b.N; i++ {
+		// We don't clear here; we only measure the draw loop overhead.
+		w.drawPlanSinglePass(drawer, plan, true, drawPlanSinglePassClipEnabled, false)
+	}
+}
diff --git a/client/world/renderer_helper.go b/client/world/renderer_helper.go
new file mode 100644
index 0000000..d7f66d3
--- /dev/null
+++ b/client/world/renderer_helper.go
@@ -0,0 +1,36 @@
+package world
+
+func (w *World) candSeenResetIfOverflow() {
+	w.candEpoch++
+	if w.candEpoch != 0 {
+		return
+	}
+	// overflow: reset stamp array
+	for i := range w.candStamp {
+		w.candStamp[i] = 0
+	}
+	w.candEpoch = 1
+}
+
+func (w *World) candSeenMark(id PrimitiveID) bool {
+	// ensure stamp capacity
+	uid := uint32(id)
+	if int(uid) >= len(w.candStamp) {
+		// grow to next power-ish
+		n := len(w.candStamp)
+		if n == 0 {
+			n = 1024
+		}
+		for n <= int(uid) {
+			n *= 2
+		}
+		ns := make([]uint32, n)
+		copy(ns, w.candStamp)
+		w.candStamp = ns
+	}
+	if w.candStamp[uid] == w.candEpoch {
+		return true
+	}
+	w.candStamp[uid] = w.candEpoch
+	return false
+}
diff --git a/client/world/renderer_incremental_render_test.go b/client/world/renderer_incremental_render_test.go
index a93369c..a618c3a 100644
--- a/client/world/renderer_incremental_render_test.go
+++ b/client/world/renderer_incremental_render_test.go
@@ -99,16 +99,10 @@ func TestRender_PanTooLarge_FallsBackToFullRedraw(t *testing.T) {
 	// Full redraw should NOT call CopyShift.
 	require.Empty(t, d.CommandsByName("CopyShift"))
 
-	// Should contain at least one reasonably wide clip.
-	clipCmds := d.CommandsByName("ClipRect")
-	require.NotEmpty(t, clipCmds)
+	// Full redraw should clear the entire canvas.
+	require.NotEmpty(t, d.CommandsByName("ClearAllTo"))
 
-	foundWide := false
-	for _, c := range clipCmds {
-		if int(c.Args[2]) > 1 {
-			foundWide = true
-			break
-		}
-	}
-	require.True(t, foundWide)
+	// And should draw something (at least the point).
+	// Depending on your implementation, it might be AddPoint or AddCircle/AddLine as well.
+	require.NotEmpty(t, d.CommandsByName("AddPoint"))
 }
diff --git a/client/world/renderer_lines.go b/client/world/renderer_lines.go
deleted file mode 100644
index ca11ad4..0000000
--- a/client/world/renderer_lines.go
+++ /dev/null
@@ -1,221 +0,0 @@
-package world
-
-// lineSeg is one canonical segment (endpoints in [0..W) x [0..H)) to be drawn.
-// It represents part of the torus-shortest polyline for a Line primitive after wrap splitting.
-type lineSeg struct {
-	x1, y1 int
-	x2, y2 int
-}
-
-// drawLinesFromPlan executes a lines-only draw from an already built render plan.
-func drawLinesFromPlan(drawer PrimitiveDrawer, plan RenderPlan, worldW, worldH int, allowWrap bool) {
-	for _, td := range plan.Tiles {
-		if td.ClipW <= 0 || td.ClipH <= 0 {
-			continue
-		}
-
-		// Filter only lines; skip tiles that have no lines.
-		lines := make([]Line, 0, len(td.Candidates))
-		for _, it := range td.Candidates {
-			l, ok := it.(Line)
-			if !ok {
-				continue
-			}
-			lines = append(lines, l)
-		}
-		if len(lines) == 0 {
-			continue
-		}
-
-		// Collect segments that actually intersect this tile's canonical rect.
-		segsToDraw := make([]lineSeg, 0, len(lines))
-		for _, l := range lines {
-			var segs []lineSeg
-			if allowWrap {
-				segs = torusShortestLineSegments(l, worldW, worldH)
-			} else {
-				segs = []lineSeg{{x1: l.X1, y1: l.Y1, x2: l.X2, y2: l.Y2}}
-			}
-			for _, s := range segs {
-				if segmentIntersectsRect(s, td.Tile.Rect) {
-					segsToDraw = append(segsToDraw, s)
-				}
-			}
-		}
-		if len(segsToDraw) == 0 {
-			continue
-		}
-
-		drawer.Save()
-		drawer.ClipRect(float64(td.ClipX), float64(td.ClipY), float64(td.ClipW), float64(td.ClipH))
-
-		for _, s := range segsToDraw {
-			// Project endpoints for this tile copy by adding tile offset.
-			x1Px := worldSpanFixedToCanvasPx((s.x1+td.Tile.OffsetX)-plan.WorldRect.minX, plan.ZoomFp)
-			y1Px := worldSpanFixedToCanvasPx((s.y1+td.Tile.OffsetY)-plan.WorldRect.minY, plan.ZoomFp)
-
-			x2Px := worldSpanFixedToCanvasPx((s.x2+td.Tile.OffsetX)-plan.WorldRect.minX, plan.ZoomFp)
-			y2Px := worldSpanFixedToCanvasPx((s.y2+td.Tile.OffsetY)-plan.WorldRect.minY, plan.ZoomFp)
-
-			drawer.AddLine(float64(x1Px), float64(y1Px), float64(x2Px), float64(y2Px))
-		}
-
-		drawer.Stroke()
-		drawer.Restore()
-	}
-}
-
-// torusShortestLineSegments converts a Line primitive into 1..4 canonical segments
-// inside [0..worldW) x [0..worldH) that represent the torus-shortest polyline.
-//
-// IMPORTANT: when a segment crosses a torus boundary, the second segment starts
-// on the opposite boundary (e.g. x=0 jump to x=worldW), preserving continuity on the torus.
-// We must NOT wrap endpoints independently at the end, otherwise the topology changes.
-func torusShortestLineSegments(l Line, worldW, worldH int) []lineSeg {
-	// Step 1: choose the torus-shortest representation in unwrapped space.
-	ax, bx := shortestWrappedDelta(l.X1, l.X2, worldW)
-	ay, by := shortestWrappedDelta(l.Y1, l.Y2, worldH)
-
-	// Step 2: shift so that A is inside canonical [0..W) x [0..H).
-	shiftX := floorDiv(ax, worldW) * worldW
-	shiftY := floorDiv(ay, worldH) * worldH
-
-	ax -= shiftX
-	bx -= shiftX
-	ay -= shiftY
-	by -= shiftY
-
-	segs := []lineSeg{{x1: ax, y1: ay, x2: bx, y2: by}}
-
-	// Step 3: split by X boundary if needed (jump-aware).
-	segs = splitSegmentsByX(segs, worldW)
-
-	// Step 4: split by Y boundary if needed (jump-aware).
-	segs = splitSegmentsByY(segs, worldH)
-
-	// Now all segments are canonical and torus-continuous. Endpoints may legally be 0 or worldW/worldH.
-	return segs
-}
-
-func splitSegmentsByX(segs []lineSeg, worldW int) []lineSeg {
-	out := make([]lineSeg, 0, len(segs)*2)
-
-	for _, s := range segs {
-		x1, y1, x2, y2 := s.x1, s.y1, s.x2, s.y2
-
-		// After normalization, x1 is expected inside [0..worldW). Only x2 may be outside.
-		if x2 >= 0 && x2 < worldW {
-			out = append(out, s)
-			continue
-		}
-
-		dx := x2 - x1
-		dy := y2 - y1
-		if dx == 0 {
-			// Degenerate; keep as-is (should not happen with normalized x1 unless x2==x1).
-			out = append(out, s)
-			continue
-		}
-
-		if x2 >= worldW {
-			// Crosses the right boundary at x=worldW, then reappears at x=0.
-			bx := worldW
-			num := bx - x1
-			iy := y1 + (dy*num)/dx
-
-			// Segment 1: [x1..worldW]
-			// Segment 2: [0..x2-worldW]
-			s1 := lineSeg{x1: x1, y1: y1, x2: worldW, y2: iy}
-			s2 := lineSeg{x1: 0, y1: iy, x2: x2 - worldW, y2: y2}
-			out = append(out, s1, s2)
-			continue
-		}
-
-		// x2 < 0: crosses the left boundary at x=0, then reappears at x=worldW.
-		bx := 0
-		num := bx - x1
-		iy := y1 + (dy*num)/dx
-
-		// Segment 1: [x1..0]
-		// Segment 2: [worldW..x2+worldW]
-		s1 := lineSeg{x1: x1, y1: y1, x2: 0, y2: iy}
-		s2 := lineSeg{x1: worldW, y1: iy, x2: x2 + worldW, y2: y2}
-		out = append(out, s1, s2)
-	}
-
-	return out
-}
-
-func splitSegmentsByY(segs []lineSeg, worldH int) []lineSeg {
-	out := make([]lineSeg, 0, len(segs)*2)
-
-	for _, s := range segs {
-		x1, y1, x2, y2 := s.x1, s.y1, s.x2, s.y2
-
-		// After normalization, y1 is expected inside [0..worldH). Only y2 may be outside.
-		if y2 >= 0 && y2 < worldH {
-			out = append(out, s)
-			continue
-		}
-
-		dx := x2 - x1
-		dy := y2 - y1
-		if dy == 0 {
-			out = append(out, s)
-			continue
-		}
-
-		if y2 >= worldH {
-			// Crosses the top boundary at y=worldH, then reappears at y=0.
-			by := worldH
-			num := by - y1
-			ix := x1 + (dx*num)/dy
-
-			s1 := lineSeg{x1: x1, y1: y1, x2: ix, y2: worldH}
-			s2 := lineSeg{x1: ix, y1: 0, x2: x2, y2: y2 - worldH}
-			out = append(out, s1, s2)
-			continue
-		}
-
-		// y2 < 0: crosses the bottom boundary at y=0, then reappears at y=worldH.
-		by := 0
-		num := by - y1
-		ix := x1 + (dx*num)/dy
-
-		s1 := lineSeg{x1: x1, y1: y1, x2: ix, y2: 0}
-		s2 := lineSeg{x1: ix, y1: worldH, x2: x2, y2: y2 + worldH}
-		out = append(out, s1, s2)
-	}
-	return out
-}
-
-// segmentIntersectsRect is a coarse bbox intersection check between a segment and a half-open rect.
-// It is used only as an optimization to avoid drawing clearly irrelevant segments.
-//
-// NOTE: Segment endpoints may legally be exactly on the world boundary (x==worldW or y==worldH).
-// Rect is half-open [min, max), so we treat the segment bbox as inclusive and intersect it with
-// [r.min, r.max-1] to avoid dropping boundary-touching segments.
-func segmentIntersectsRect(s lineSeg, r Rect) bool {
-	minX := min(s.x1, s.x2)
-	maxX := max(s.x1, s.x2)
-	minY := min(s.y1, s.y2)
-	maxY := max(s.y1, s.y2)
-
-	// Treat degenerate as having 1-unit thickness for indexing-like behavior.
-	if minX == maxX {
-		maxX++
-	}
-	if minY == maxY {
-		maxY++
-	}
-
-	// Rect inclusive bounds for intersection purposes.
-	rMaxX := r.maxX - 1
-	rMaxY := r.maxY - 1
-	if rMaxX < r.minX || rMaxY < r.minY {
-		return false
-	}
-
-	// Inclusive overlap check.
-	return !(maxX-1 < r.minX || minX > rMaxX || maxY-1 < r.minY || minY > rMaxY)
-}
diff --git a/client/world/renderer_lines_test.go b/client/world/renderer_lines_test.go
index 2f20cae..64fc013 100644
--- a/client/world/renderer_lines_test.go
+++ b/client/world/renderer_lines_test.go
@@ -6,132 +6,6 @@ import (
 	"github.com/stretchr/testify/require"
 )
 
-func TestDrawLinesFromPlan_WrapX_SplitsAndDrawsInThreeXTiles(t *testing.T) {
-	t.Parallel()
-
-	w := NewWorld(10, 10)
-	w.resetGrid(2 * SCALE)
-
-	// Horizontal line that wraps across X: 9 -> 1 at y=5.
-	id, err := w.AddLine(9, 5, 1, 5)
-	require.NoError(t, err)
-	w.indexObject(w.objects[id])
-
-	params := RenderParams{
-		ViewportWidthPx:  10,
-		ViewportHeightPx: 10,
-		MarginXPx:        2,
-		MarginYPx:        2,
-		CameraXWorldFp:   5 * SCALE,
-		CameraYWorldFp:   5 * SCALE,
-		CameraZoom:       1.0,
-	}
-
-	plan, err := w.buildRenderPlanStageA(params)
-	require.NoError(t, err)
-
-	d := &fakePrimitiveDrawer{}
-	drawLinesFromPlan(d, plan, w.W, w.H, true)
-
-	// Expect drawing in 3 X tiles (left partial, middle full, right partial) for the central Y tile:
-	// Each tile group: Save, ClipRect, AddLine(s), Stroke, Restore
-	//
-	// Left tile (offsetX=-10000): 1 line segment.
-	// Middle tile (offsetX=0): 2 segments (wrapped split).
-	// Right tile (offsetX=10000): 1 segment.
-	wantNames := []string{
-		"Save", "ClipRect", "AddLine", "Stroke", "Restore",
-		"Save", "ClipRect", "AddLine", "AddLine", "Stroke", "Restore",
-		"Save", "ClipRect", "AddLine", "Stroke", "Restore",
-	}
-	require.Equal(t, wantNames, d.CommandNames())
-
-	// Group 1: left strip clip (0,2,2,10), line at y=7 from x=1..2
-	{
-		requireCommandArgs(t, requireDrawerCommandAt(t, d, 1), 0, 2, 2, 10)
-		requireCommandArgs(t, requireDrawerCommandAt(t, d, 2), 1, 7, 2, 7)
-	}
-
-	// Group 2: middle strip clip (2,2,10,10), two segments:
-	// segment [9000..10000] => x 11..12, y 7
-	// segment [0..1000]     => x 2..3,  y 7
-	{
-		requireCommandArgs(t, requireDrawerCommandAt(t, d, 6), 2, 2, 10, 10)
-
-		// The order of segments is stable with our splitting: first the one ending at boundary, then the remainder.
-		requireCommandArgs(t, requireDrawerCommandAt(t, d, 7), 11, 7, 12, 7)
-		requireCommandArgs(t, requireDrawerCommandAt(t, d, 8), 2, 7, 3, 7)
-	}
-
-	// Group 3: right strip clip (12,2,2,10), line at y=7 from x=12..13
-	{
-		requireCommandArgs(t, requireDrawerCommandAt(t, d, 12), 12, 2, 2, 10) // ClipRect
-		requireCommandArgs(t, requireDrawerCommandAt(t, d, 13), 12, 7, 13, 7) // AddLine
-	}
-}
-
-func TestDrawLinesFromPlan_WrapY_SplitsAndDrawsInThreeYTiles(t *testing.T) {
-	t.Parallel()
-
-	w := NewWorld(10, 10)
-	w.resetGrid(2 * SCALE)
-
-	// Vertical line that wraps across Y: 9 -> 1 at x=5.
-	id, err := w.AddLine(5, 9, 5, 1)
-	require.NoError(t, err)
-	w.indexObject(w.objects[id])
-
-	params := RenderParams{
-		ViewportWidthPx:  10,
-		ViewportHeightPx: 10,
-		MarginXPx:        2,
-		MarginYPx:        2,
-		CameraXWorldFp:   5 * SCALE,
-		CameraYWorldFp:   5 * SCALE,
-		CameraZoom:       1.0,
-	}
-
-	plan, err := w.buildRenderPlanStageA(params)
-	require.NoError(t, err)
-
-	d := &fakePrimitiveDrawer{}
-	drawLinesFromPlan(d, plan, w.W, w.H, true)
-
-	// Here we expect 3 Y tiles for the central X tile:
-	// Top partial, middle full (two segments), bottom partial.
-	//
-	// The exact ordering of tiles is by tx then ty, so X=middle strips first.
-	// For this geometry the line only intersects the middle X tiles (offsetX=0),
-	// but spans Y across -1,0,1.
-	wantNames := []string{
-		"Save", "ClipRect", "AddLine", "Stroke", "Restore",
-		"Save", "ClipRect", "AddLine", "AddLine", "Stroke", "Restore",
-		"Save", "ClipRect", "AddLine", "Stroke", "Restore",
-	}
-	require.Equal(t, wantNames, d.CommandNames())
-
-	// Group 1: top strip clip (2,0,10,2), line at x=7 from y=1..2
-	{
-		requireCommandArgs(t, requireDrawerCommandAt(t, d, 1), 2, 0, 10, 2)
-		requireCommandArgs(t, requireDrawerCommandAt(t, d, 2), 7, 1, 7, 2)
-	}
-
-	// Group 2: middle strip clip (2,2,10,10), two segments:
-	// segment [9000..10000] => y 11..12 at x=7
-	// segment [0..1000]     => y 2..3  at x=7
-	{
-		requireCommandArgs(t, requireDrawerCommandAt(t, d, 6), 2, 2, 10, 10)
-		requireCommandArgs(t, requireDrawerCommandAt(t, d, 7), 7, 11, 7, 12)
-		requireCommandArgs(t, requireDrawerCommandAt(t, d, 8), 7, 2, 7, 3)
-	}
-
-	// Group 3: bottom strip clip (2,12,10,2), line at x=7 from y=12..13
-	{
-		requireCommandArgs(t, requireDrawerCommandAt(t, d, 12), 2, 12, 10, 2) // ClipRect
-		requireCommandArgs(t, requireDrawerCommandAt(t, d, 13), 7, 12, 7, 13) // AddLine
-	}
-}
-
 func TestTorusShortestLineSegments_TieCaseIsDeterministicAndSplits(t *testing.T) {
 	t.Parallel()
 
diff --git a/client/world/renderer_plan_bench_test.go b/client/world/renderer_plan_bench_test.go
new file mode 100644
index 0000000..c894611
--- /dev/null
+++ b/client/world/renderer_plan_bench_test.go
@@ -0,0 +1,53 @@
+package world
+
+import (
+	"image/color"
+	"testing"
+)
+
+func BenchmarkBuildRenderPlanStageA_Candidates(b *testing.B) {
+	w := NewWorld(600, 600)
+
+	// Make the index/grid available.
+	w.IndexOnViewportChange(1000, 700, 1.0)
+
+	// Populate with enough objects to create duplicates across cells.
+	// Circles and lines create bbox indexing (more duplicates).
+	for i := 0; i < 2000; i++ {
+		_, _ = w.AddPoint(float64(i%600), float64((i*7)%600))
+	}
+	for i := 0; i < 1200; i++ {
+		_, _ = w.AddCircle(float64((i*11)%600), float64((i*13)%600), 8.0)
+	}
+	for i := 0; i < 1200; i++ {
+		x1 := float64((i*3 + 10) % 600)
+		y1 := float64((i*5 + 20) % 600)
+		x2 := float64((i*7 + 400) % 600)
+		y2 := float64((i*11 + 300) % 600)
+		_, _ = w.AddLine(x1, y1, x2, y2)
+	}
+	w.Reindex()
+
+	params := RenderParams{
+		ViewportWidthPx:  1000,
+		ViewportHeightPx: 700,
+		MarginXPx:        250,
+		MarginYPx:        175,
+		CameraXWorldFp:   300 * SCALE,
+		CameraYWorldFp:   300 * SCALE,
+		CameraZoom:       1.0,
+		Options: &RenderOptions{
+			BackgroundColor: color.RGBA{A: 255},
+		},
+	}
+
+	b.ReportAllocs()
+	b.ResetTimer()
+
+	for i := 0; i < b.N; i++ {
+		_, err := w.buildRenderPlanStageA(params)
+		if err != nil {
+			b.Fatalf("build plan: %v", err)
+		}
+	}
+}
diff --git a/client/world/renderer_query.go b/client/world/renderer_query.go
index 887ec4a..739579c 100644
--- a/client/world/renderer_query.go
+++ b/client/world/renderer_query.go
@@ -57,22 +57,36 @@ func (w *World) collectCandidatesForTile(r Rect) []MapItem {
 	rowStart := w.worldToCellY(r.minY)
 	rowEnd := w.worldToCellY(r.maxY - 1)
 
-	seen := make(map[PrimitiveID]struct{})
-	result := make([]MapItem, 0)
+	// Start a new epoch for this tile dedupe.
+	w.candSeenResetIfOverflow()
+
+	// Reuse result buffer.
+	out := w.scratchCandidates[:0]
 
 	for row := rowStart; row <= rowEnd; row++ {
 		for col := colStart; col <= colEnd; col++ {
 			cell := w.grid[row][col]
 			for _, item := range cell {
 				id := item.ID()
-				if _, ok := seen[id]; ok {
+				if w.candSeenMark(id) {
 					continue
 				}
-				seen[id] = struct{}{}
-				result = append(result, item)
+				out = append(out, item)
 			}
 		}
 	}
 
-	return result
+	// Store back the reusable buffer (keep capacity).
+	w.scratchCandidates = out[:0]
+
+	// IMPORTANT:
+	// We must return a stable slice to the caller (plan stores it).
+	// Returning `out` directly would be overwritten on the next tile.
+	//
+	// So: copy out into a freshly allocated slice OR into a plan-level scratch pool.
+	// For Step 1 we keep correctness: allocate exactly once per tile.
+	// Step 3 will remove this allocation by making plan own a pooled backing store.
+	res := make([]MapItem, len(out))
+	copy(res, out)
+	return res
 }
diff --git a/client/world/renderer_smoke_all_primitives_test.go b/client/world/renderer_smoke_all_primitives_test.go
deleted file mode 100644
index 64cfc05..0000000
--- a/client/world/renderer_smoke_all_primitives_test.go
+++ /dev/null
@@ -1,60 +0,0 @@
-package world
-
-import (
-	"testing"
-
-	"github.com/stretchr/testify/require"
-)
-
-func TestSmoke_DrawPointsCirclesLinesFromSamePlan(t *testing.T) {
-	t.Parallel()
-
-	w := NewWorld(10, 10)
-	w.resetGrid(2 * SCALE)
-
-	// Mix primitives. Use values that are safely inside the world.
-	_, err := w.AddPoint(1, 1)
-	require.NoError(t, err)
-
-	_, err = w.AddCircle(2, 2, 1)
-	require.NoError(t, err)
-
-	// A line that wraps across X to ensure line splitting logic is exercised.
-	_, err = w.AddLine(9, 5, 1, 5)
-	require.NoError(t, err)
-
-	// Build index (in UI: IndexOnViewportChange does this).
-	for _, obj := range w.objects {
-		w.indexObject(obj)
-	}
-
-	params := RenderParams{
-		ViewportWidthPx:  10,
-		ViewportHeightPx: 10,
-		MarginXPx:        2,
-		MarginYPx:        2,
-		CameraXWorldFp:   5 * SCALE,
-		CameraYWorldFp:   5 * SCALE,
-		CameraZoom:       1.0,
-	}
-
-	plan, err := w.buildRenderPlanStageA(params)
-	require.NoError(t, err)
-
-	d := &fakePrimitiveDrawer{}
-
-	// Execute all three passes over the same plan.
-	drawPointsFromPlan(d, plan, true)
-	drawCirclesFromPlan(d, plan, w.W, w.H, true, w.circleRadiusScaleFp)
-	drawLinesFromPlan(d, plan, w.W, w.H, true)
-
-	names := d.CommandNames()
-
-	require.Contains(t, names, "AddPoint")
-	require.Contains(t, names, "AddCircle")
-	require.Contains(t, names, "AddLine")
-
-	// Ensure finalizers were used (points+circles use Fill, lines use Stroke).
-	require.Contains(t, names, "Fill")
-	require.Contains(t, names, "Stroke")
-}
diff --git a/client/world/torus_line_segments_into.go b/client/world/torus_line_segments_into.go
new file mode 100644
index 0000000..c65c7c5
--- /dev/null
+++ b/client/world/torus_line_segments_into.go
@@ -0,0 +1,135 @@
+package world
+
+// torusShortestLineSegmentsInto converts a Line primitive into 1..4 canonical segments
+// inside [0..worldW) x [0..worldH) that represent the torus-shortest polyline.
+//
+// It appends results into dst using tmp as an intermediate buffer.
+// No allocations occur if dst/tmp have sufficient capacity (>=4).
+func torusShortestLineSegmentsInto(dst, tmp []lineSeg, l Line, worldW, worldH int) ([]lineSeg, []lineSeg) {
+	dst = dst[:0]
+	tmp = tmp[:0]
+
+	// Step 1: choose the torus-shortest representation in unwrapped space.
+	ax, bx := shortestWrappedDelta(l.X1, l.X2, worldW)
+	ay, by := shortestWrappedDelta(l.Y1, l.Y2, worldH)
+
+	// Step 2: shift so that A is inside canonical [0..W) x [0..H).
+	shiftX := floorDiv(ax, worldW) * worldW
+	shiftY := floorDiv(ay, worldH) * worldH
+
+	ax -= shiftX
+	bx -= shiftX
+	ay -= shiftY
+	by -= shiftY
+
+	dst = append(dst, lineSeg{x1: ax, y1: ay, x2: bx, y2: by})
+
+	// Step 3: split by X boundary if needed (jump-aware).
+	tmp = splitSegmentsByXInto(tmp, dst, worldW)
+
+	// Step 4: split by Y boundary if needed (jump-aware).
+	dst = splitSegmentsByYInto(dst, tmp, worldH)
+
+	return dst, tmp
+}
+
+// torusShortestLineSegments is a compatibility wrapper that allocates.
+// Prefer torusShortestLineSegmentsInto in hot paths.
+func torusShortestLineSegments(l Line, worldW, worldH int) []lineSeg {
+	dst := make([]lineSeg, 0, 4)
+	tmp := make([]lineSeg, 0, 4)
+	dst, _ = torusShortestLineSegmentsInto(dst, tmp, l, worldW, worldH)
+	return dst
+}
+
+// splitSegmentsByXInto appends 1..2 segments for each input segment into out, without allocating.
+// out is reset to length 0 by this function.
+func splitSegmentsByXInto(out []lineSeg, segs []lineSeg, worldW int) []lineSeg {
+	out = out[:0]
+
+	for _, s := range segs {
+		x1, y1, x2, y2 := s.x1, s.y1, s.x2, s.y2
+
+		// After normalization, x1 is expected inside [0..worldW). Only x2 may be outside.
+		if x2 >= 0 && x2 < worldW {
+			out = append(out, s)
+			continue
+		}
+
+		dx := x2 - x1
+		dy := y2 - y1
+		if dx == 0 {
+			out = append(out, s)
+			continue
+		}
+
+		if x2 >= worldW {
+			// Crosses the right boundary at x=worldW, then reappears at x=0.
+			bx := worldW
+			num := bx - x1
+			iy := y1 + (dy*num)/dx
+
+			s1 := lineSeg{x1: x1, y1: y1, x2: worldW, y2: iy}
+			s2 := lineSeg{x1: 0, y1: iy, x2: x2 - worldW, y2: y2}
+			out = append(out, s1, s2)
+			continue
+		}
+
+		// x2 < 0: crosses the left boundary at x=0, then reappears at x=worldW.
+		bx := 0
+		num := bx - x1
+		iy := y1 + (dy*num)/dx
+
+		s1 := lineSeg{x1: x1, y1: y1, x2: 0, y2: iy}
+		s2 := lineSeg{x1: worldW, y1: iy, x2: x2 + worldW, y2: y2}
+		out = append(out, s1, s2)
+	}
+
+	return out
+}
+
+// splitSegmentsByYInto appends 1..2 segments for each input segment into out, without allocating.
+// out is reset to length 0 by this function.
+func splitSegmentsByYInto(out []lineSeg, segs []lineSeg, worldH int) []lineSeg {
+	out = out[:0]
+
+	for _, s := range segs {
+		x1, y1, x2, y2 := s.x1, s.y1, s.x2, s.y2
+
+		// After normalization, y1 is expected inside [0..worldH). Only y2 may be outside.
+		if y2 >= 0 && y2 < worldH {
+			out = append(out, s)
+			continue
+		}
+
+		dx := x2 - x1
+		dy := y2 - y1
+		if dy == 0 {
+			out = append(out, s)
+			continue
+		}
+
+		if y2 >= worldH {
+			// Crosses the top boundary at y=worldH, then reappears at y=0.
+			by := worldH
+			num := by - y1
+			ix := x1 + (dx*num)/dy
+
+			s1 := lineSeg{x1: x1, y1: y1, x2: ix, y2: worldH}
+			s2 := lineSeg{x1: ix, y1: 0, x2: x2, y2: y2 - worldH}
+			out = append(out, s1, s2)
+			continue
+		}
+
+		// y2 < 0: crosses the bottom boundary at y=0, then reappears at y=worldH.
+		by := 0
+		num := by - y1
+		ix := x1 + (dx*num)/dy
+
+		s1 := lineSeg{x1: x1, y1: y1, x2: ix, y2: 0}
+		s2 := lineSeg{x1: ix, y1: worldH, x2: x2, y2: y2 + worldH}
+		out = append(out, s1, s2)
+	}
+
+	return out
+}
diff --git a/client/world/world.go b/client/world/world.go
index ed5f114..46b77f8 100644
--- a/client/world/world.go
+++ b/client/world/world.go
@@ -50,6 +50,17 @@ type World struct {
 
 	renderState  rendererIncrementalState
 	derivedCache map[derivedStyleKey]StyleID
+
+	// scratch buffers for hot render path (single goroutine assumption).
+	scratchDrawItems   []drawItem
+	scratchWrapShifts  []wrapShift
+	scratchLineSegs    []lineSeg
+	scratchLineSegsTmp []lineSeg
+
+	// candidate dedupe scratch (hot path for plan building).
+	candStamp         []uint32
+	candEpoch         uint32
+	scratchCandidates []MapItem
 }
 
 // NewWorld constructs a new world with the given real dimensions.