feat(build): fuzz testing and git-cliff changelog automation

Add FuzzParseBytes and FuzzSanitizeYAMLEscapes in internal/model/fuzz_test.go.
Both targets run as regular seed-based tests in CI (go test ./...) and can be
run continuously via `task fuzz` (default 30 s per target; FUZZ_TIME=60s to
extend). Fuzz corpus failures are saved to testdata/fuzz/ for regression.

Add cliff.toml configuring git-cliff to generate Keep-a-Changelog-compatible
release notes from Conventional Commits. New tasks: `task changelog`
(regenerate full CHANGELOG.md) and `task changelog-next` (preview unreleased
entries without writing). Requires git-cliff (brew/cargo install git-cliff).

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

CHANGELOG.md

@@ -5,6 +5,22 @@ All notable changes to this project will be documented in this file.
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/).
 This project uses [Semantic Versioning](https://semver.org).
 
+## [0.2.27] - 2026-06-25
+
+### Added
+
+- **Fuzz testing** — `FuzzParseBytes` and `FuzzSanitizeYAMLEscapes` in `internal/model/fuzz_test.go` verify that neither the YAML parser nor the escape sanitizer panics on arbitrary input. Successful parses are also checked for structural integrity (non-nil pipeline, no empty job names). Run with `task fuzz` (default 30 s per target; set `FUZZ_TIME=60s` to extend). Found failures are saved to `testdata/fuzz/` for regression.
+
+- **Changelog automation** — `cliff.toml` configures [git-cliff](https://git-cliff.org) to generate Keep-a-Changelog–compatible release notes from Conventional Commits. `task changelog` regenerates `CHANGELOG.md` from the full git history; `task changelog-next` previews only unreleased commits without writing. Install git-cliff with `brew install git-cliff` or `cargo install git-cliff`.
+
+## [0.2.26] - 2026-06-25
+
+### Changed
+
+- **Same-stage job ordering** — when jobs within the same declared GitLab stage have `needs:` relationships between each other, the pipeline graph now splits that stage into topological sub-columns: jobs with no same-stage dependencies occupy the leftmost sub-column; jobs that depend on them are placed one sub-column to the right. Sub-columns use a narrower 20 px gap (vs. the 50 px gap between stages), and the stage header spans all sub-columns. Stages with no intra-stage `needs:` are unaffected.
+
+- **Graph links rendered behind job chips** — SVG connector lines (Bézier curves in DAG mode, bus-bar stubs in classic mode) are now drawn before the job chip rectangles, so connector lines pass behind chips rather than on top of them.
+
 ## [0.2.25] - 2026-06-25
 
 ### Added

README.md

@@ -6,7 +6,7 @@
 
 <p align="center">
   <a href="LICENSE"><img src="https://img.shields.io/badge/license-Apache%202.0-blue.svg" alt="License"></a>
-  <a href="CHANGELOG.md"><img src="https://img.shields.io/badge/release-v0.2.25-blue.svg" alt="Release"></a>
+  <a href="CHANGELOG.md"><img src="https://img.shields.io/badge/release-v0.2.27-blue.svg" alt="Release"></a>
 </p>
 
 > **Disclaimer:** This tool was built through iterative AI-assisted development with [Claude](https://claude.ai). It is experimental, incomplete, and not intended for production use. Coverage of GitLab CI keywords is best-effort and may lag behind GitLab's evolving spec. Use it at your own discretion — no correctness guarantees are made. Contributions and bug reports are welcome.
@@ -69,11 +69,18 @@ task test         # run Go unit tests
 task lint-go      # run go vet
 task validate     # run the binary against all testdata fixtures
 task ci           # full check: vet → test → build → validate
+task fuzz         # run fuzz tests for the YAML parser (Ctrl-C to stop; FUZZ_TIME=60s to set duration)
+task changelog    # regenerate CHANGELOG.md from git history via git-cliff
+task changelog-next  # preview unreleased section (dry-run, no file written)
 task build-windows  # cross-compile for Windows x64 (requires a tagged commit → glint-<tag>.exe)
 task build-linux    # cross-compile for Linux x64 (requires a tagged commit → glint-<tag>-linux-amd64)
 task clean        # remove build artifacts
 ```
 
+**Optional tools:**
+
+- [git-cliff](https://git-cliff.org) — changelog generator used by `task changelog`. Install with `brew install git-cliff` or `cargo install git-cliff`.
+
 ## Project structure
 
 ```

ROADMAP.md

@@ -77,6 +77,8 @@ The SVG renderer and terminal tree cover the basic layout. These would bring it
 - ~~**Blocked / skipped state colouring**~~ — ✓ shipped v0.2.25; `glint graph pipeline` accepts context flags (`--branch`, `--tag`, etc.); jobs evaluated as skipped are greyed out (`#868686`) with dimmed text and no icon
 - ~~**Interactive HTML output**~~ — ✓ shipped v0.2.25; `glint graph pipeline --format html` writes a self-contained `.html` file with mouse pan/zoom and a click-to-open job-detail sidebar; no external dependencies
 - ~~**Mermaid pipeline output**~~ — ✓ shipped v0.2.25; `glint graph pipeline --format mermaid` prints a Mermaid flowchart to stdout (paste into mermaid.live)
+- ~~**Same-stage job ordering**~~ — ✓ shipped v0.2.26; jobs within a stage that have `needs:` between each other are placed in topological sub-columns (left-to-right by depth); stage header spans all sub-columns
+- ~~**Graph links rendered behind job chips**~~ — ✓ shipped v0.2.26; SVG connectors (Bézier curves and bus-bar stubs) are drawn before job chips so lines pass behind rectangles
 
 ---
 
@@ -113,9 +115,9 @@ The SVG renderer and terminal tree cover the basic layout. These would bring it
 
 ## Reliability and developer experience
 
-- ~~**Structured rule IDs**~~ — ✓ shipped post-v0.2.0; GL001–GL031 assigned; GL032 added v0.2.11; GL033 added v0.2.15; GL034–GL041 added v0.2.16; output formats (--format json/sarif/junit/github) added v0.2.18; GL042–GL043 added v0.2.20; GL044 added v0.2.24; graph improvements shipped v0.2.25
+- ~~**Structured rule IDs**~~ — ✓ shipped post-v0.2.0; GL001–GL031 assigned; GL032 added v0.2.11; GL033 added v0.2.15; GL034–GL041 added v0.2.16; output formats (--format json/sarif/junit/github) added v0.2.18; GL042–GL043 added v0.2.20; GL044 added v0.2.24; graph improvements shipped v0.2.25–v0.2.26
 - ~~**`glint explain <rule-id>`**~~ — ✓ shipped v0.2.20; prints rule description, rationale, bad-YAML example, and fix; `glint explain` (no arg) lists all rules
 - ~~**Semantic versioning and first release**~~ — shipped as `v0.1.0` (2026-06-07)
 - ~~**Subcommand CLI**~~ — shipped as `v0.2.0` (2026-06-11); `glint check` / `glint graph [mode]` with ruff-style `--help`
-- **Changelog automation** — generate release notes from Conventional Commits via `git-cliff` or similar
+- ~~**Changelog automation**~~ — ✓ shipped v0.2.27; `cliff.toml` configures git-cliff to produce Keep-a-Changelog–compatible release notes from Conventional Commits; `task changelog` regenerates `CHANGELOG.md`, `task changelog-next` previews unreleased entries
-- **Fuzz testing** — add a `go test -fuzz` target for the YAML parser to harden it against malformed input
+- ~~**Fuzz testing**~~ — ✓ shipped v0.2.27; `FuzzParseBytes` and `FuzzSanitizeYAMLEscapes` in `internal/model/fuzz_test.go`; seeds run as regular tests in CI; `task fuzz` runs them continuously (default 30 s)

Taskfile.yml

@@ -175,6 +175,20 @@ tasks:
     generates:
       - "{{.BINARY}}-{{.TAG}}-linux-amd64"
 
+  fuzz:
+    desc: "Run fuzz tests for the YAML parser and sanitizer (set FUZZ_TIME=60s to control duration, default 30s)"
+    cmds:
+      - "{{.GO}} test -fuzz=FuzzParseBytes          -fuzztime=${FUZZ_TIME:-30s} ./internal/model/"
+      - "{{.GO}} test -fuzz=FuzzSanitizeYAMLEscapes -fuzztime=${FUZZ_TIME:-30s} ./internal/model/"
+
+  changelog:
+    desc: "Regenerate CHANGELOG.md from git history (requires git-cliff — see README)"
+    cmd: git cliff --config cliff.toml --output CHANGELOG.md
+
+  changelog-next:
+    desc: "Preview unreleased changelog entries without writing (requires git-cliff)"
+    cmd: git cliff --config cliff.toml --unreleased
+
   clean:
     desc: Remove build artifacts
     cmd: rm -f {{.BINARY}} {{.BINARY}}-*.exe {{.BINARY}}-*-linux-amd64

cliff.toml

@@ -0,0 +1,65 @@
+# git-cliff configuration for glint
+# Install: brew install git-cliff  OR  cargo install git-cliff
+# Usage:   task changelog            -- regenerate full CHANGELOG.md
+#          task changelog-next       -- preview unreleased section (dry-run)
+
+[changelog]
+header = """
+# Changelog
+
+All notable changes to this project will be documented in this file.
+
+The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/).
+This project uses [Semantic Versioning](https://semver.org).
+"""
+body = """
+{% if version %}\
+## [{{ version | trim_start_matches(pat="v") }}] - {{ timestamp | date(format="%Y-%m-%d") }}
+
+{% else %}\
+## [Unreleased]
+
+{% endif %}\
+{% for group, commits in commits | group_by(attribute="group") %}\
+### {{ group | upper_first }}
+
+{% for commit in commits %}\
+- {% if commit.scope %}**{{ commit.scope }}**: {% endif %}\
+{{ commit.message | upper_first }}\
+{% if commit.breaking %} **[BREAKING]**{% endif %}
+
+{% endfor %}\
+{% endfor %}\
+"""
+trim = true
+footer = ""
+postprocessors = []
+
+[git]
+conventional_commits = true
+filter_unconventional = true
+split_commits = false
+commit_preprocessors = [
+  # Drop Co-Authored-By trailers (should not appear in subjects, but guard anyway).
+  { pattern = "Co-Authored-By:.*", replace = "" },
+]
+commit_parsers = [
+  # Breaking changes (type! or scope!) — promote above everything else.
+  { message = "^[a-z]+\\([a-z-]+\\)!:|^[a-z]+!:", group = "Breaking Changes" },
+  { message = "^feat",     group = "Added" },
+  { message = "^fix",      group = "Fixed" },
+  { message = "^perf",     group = "Changed" },
+  { message = "^refactor", group = "Changed" },
+  # Maintenance commits — omit from the changelog body.
+  { message = "^docs",   skip = true },
+  { message = "^style",  skip = true },
+  { message = "^test",   skip = true },
+  { message = "^chore",  skip = true },
+  { message = "^build",  skip = true },
+  { message = "^claude", skip = true },
+]
+protect_breaking_commits = false
+filter_commits = true
+tag_pattern = "v[0-9].*"
+topo_order = false
+sort_commits = "oldest"

internal/graph/render.go

@@ -16,21 +16,124 @@ import (
 
 // Layout constants (pixels) – tuned to resemble GitLab's full pipeline graph view.
 const (
-	chipW    = 178 // job chip width
+	chipW       = 178 // job chip width
-	chipH    = 34  // job chip height
+	chipH       = 34  // job chip height
-	chipGap  = 6   // vertical gap between chips in a column
+	chipGap     = 6   // vertical gap between chips in a column
-	iconCX   = 14  // status circle: x offset from chip left edge to circle centre
+	iconCX      = 14  // status circle: x offset from chip left edge to centre
-	iconR    = 7   // status circle radius (14 px diameter)
+	iconR       = 7   // status circle radius (14 px diameter)
-	textLeft = 27  // job name text: x offset from chip left edge
+	textLeft    = 27  // job name text: x offset from chip left edge
-	labelH   = 30  // stage-name label area height (text + bottom gap)
+	labelH      = 30  // stage-name label area height (text + bottom gap)
-	topPad   = 40  // outer top padding
+	topPad      = 40  // outer top padding
-	sidePad  = 40  // outer left / right padding
+	sidePad     = 40  // outer left / right padding
-	stageGap = 50  // horizontal gap between stage columns (connector space)
+	stageGap    = 50  // horizontal gap between stage columns (connector space)
-	botPad   = 48  // outer bottom padding (legend lives here)
+	subStageGap = 20  // horizontal gap between sub-columns within the same stage
+	botPad      = 48  // outer bottom padding (legend lives here)
 )
 
 type svgPt struct{ x, y int }
 
+// column represents one vertical stack of job chips in the SVG.
+// A declared GitLab stage maps to one column unless jobs within it have
+// same-stage needs:, in which case it splits into topological sub-columns.
+type column struct {
+	stage string   // declared GitLab stage name
+	jobs  []string // job names in this column, topologically ordered
+}
+
+// computeColumns assigns jobs to SVG columns.
+// When jobs within the same declared stage have needs: relationships between
+// each other, the stage is split into topological sub-columns so that
+// depended-upon jobs appear to the left of the jobs that depend on them.
+func computeColumns(stages []string, byStage map[string][]string, jobs map[string]model.Job) []column {
+	var cols []column
+	for _, stage := range stages {
+		stageJobs := byStage[stage]
+		sort.Strings(stageJobs)
+		if len(stageJobs) == 0 {
+			continue
+		}
+
+		// Index same-stage membership.
+		stageSet := make(map[string]bool, len(stageJobs))
+		for _, j := range stageJobs {
+			stageSet[j] = true
+		}
+
+		// Check whether any job has a same-stage need.
+		hasIntraNeeds := false
+		outer:
+		for _, j := range stageJobs {
+			for _, need := range jobs[j].Needs {
+				if dep := needsJobName(need); dep != "" && stageSet[dep] {
+					hasIntraNeeds = true
+					break outer
+				}
+			}
+		}
+
+		if !hasIntraNeeds {
+			cols = append(cols, column{stage: stage, jobs: stageJobs})
+			continue
+		}
+
+		// Compute topological depth within the stage.
+		depth := make(map[string]int, len(stageJobs))
+		for _, j := range stageJobs {
+			depth[j] = -1
+		}
+		inProgress := make(map[string]bool, len(stageJobs))
+		var computeDepth func(j string) int
+		computeDepth = func(j string) int {
+			if depth[j] >= 0 {
+				return depth[j]
+			}
+			if inProgress[j] {
+				return 0 // cycle guard (cycles already rejected by GL029)
+			}
+			inProgress[j] = true
+			maxDep := -1
+			for _, need := range jobs[j].Needs {
+				dep := needsJobName(need)
+				if dep == "" || !stageSet[dep] {
+					continue
+				}
+				d := computeDepth(dep)
+				if d > maxDep {
+					maxDep = d
+				}
+			}
+			depth[j] = maxDep + 1
+			return depth[j]
+		}
+		for _, j := range stageJobs {
+			if depth[j] < 0 {
+				computeDepth(j)
+			}
+		}
+
+		// One sub-column per topological depth level.
+		maxDepth := 0
+		for _, j := range stageJobs {
+			if depth[j] > maxDepth {
+				maxDepth = depth[j]
+			}
+		}
+		for l := 0; l <= maxDepth; l++ {
+			var levelJobs []string
+			for _, j := range stageJobs {
+				if depth[j] == l {
+					levelJobs = append(levelJobs, j)
+				}
+			}
+			if len(levelJobs) > 0 {
+				sort.Strings(levelJobs)
+				cols = append(cols, column{stage: stage, jobs: levelJobs})
+			}
+		}
+	}
+	return cols
+}
+
 // RenderPipeline writes a PNG (or SVG fallback) file with a GitLab CI-style
 // pipeline layout and returns the path to the generated file.
 func RenderPipeline(p *model.Pipeline, outDir string, ctx *cicontext.Context) (string, error) {
@@ -148,28 +251,42 @@ func pipelineSVG(p *model.Pipeline, ctx *cicontext.Context) string {
 		}
 	}
 
-	// Compute per-stage column heights and job anchor points for connectors.
+	// ── Column layout ─────────────────────────────────────────────────────────
+	// Split stages into sub-columns when intra-stage needs exist.
+	cols := computeColumns(stages, byStage, p.Jobs)
+
+	// X position of each column's left edge.
+	colX := make([]int, len(cols))
+	if len(cols) > 0 {
+		colX[0] = sidePad
+		for i := 1; i < len(cols); i++ {
+			gap := stageGap
+			if cols[i].stage == cols[i-1].stage {
+				gap = subStageGap
+			}
+			colX[i] = colX[i-1] + chipW + gap
+		}
+	}
+
+	// Compute SVG dimensions and per-job connector anchor points.
 	maxColH := 0
 	rightMid := make(map[string]svgPt)
 	leftMid := make(map[string]svgPt)
 
-	for i, stage := range stages {
+	for ci, col := range cols {
-		jobs := byStage[stage]
+		n := len(col.jobs)
-		sort.Strings(jobs)
-		n := len(jobs)
 		h := n*chipH + max(0, n-1)*chipGap
 		if h > maxColH {
 			maxColH = h
 		}
-		cx := sidePad + i*(chipW+stageGap)
+		for j, name := range col.jobs {
-		for j, name := range jobs {
 			chy := topPad + labelH + j*(chipH+chipGap)
-			rightMid[name] = svgPt{cx + chipW, chy + chipH/2}
+			rightMid[name] = svgPt{colX[ci] + chipW, chy + chipH/2}
-			leftMid[name] = svgPt{cx, chy + chipH/2}
+			leftMid[name] = svgPt{colX[ci], chy + chipH/2}
 		}
 	}
 
-	svgW := sidePad*2 + len(stages)*chipW + max(0, len(stages)-1)*stageGap
+	svgW := colX[len(cols)-1] + chipW + sidePad
 	svgH := topPad + labelH + maxColH + botPad
 
 	// DAG mode: any visible job with a needs: list triggers job-to-job arrows.
@@ -199,30 +316,117 @@ func pipelineSVG(p *model.Pipeline, ctx *cicontext.Context) string {
 	// White page background.
 	wf(`  <rect width="%d" height="%d" fill="#ffffff"/>`, svgW, svgH)
 
-	// ── Stage columns ─────────────────────────────────────────────────────────
+	// ── Stage headers ─────────────────────────────────────────────────────────
-	for i, stage := range stages {
+	// Each declared stage may span multiple sub-columns; draw one header per stage.
-		cx := sidePad + i*(chipW+stageGap)
+	drawnHeader := make(map[string]bool)
-		jobs := byStage[stage]
+	for ci, col := range cols {
-		sort.Strings(jobs)
+		if drawnHeader[col.stage] {
+			continue
+		}
+		// Span all sub-columns that belong to this stage.
+		x1 := colX[ci]
+		x2 := colX[ci] + chipW
+		for j := ci + 1; j < len(cols) && cols[j].stage == col.stage; j++ {
+			x2 = colX[j] + chipW
+		}
+		centerX := (x1 + x2) / 2
 
-		// Stage name – small, gray, uppercase, centered above the chip stack.
+		// Stage name – small, gray, uppercase.
 		wf(`  <text x="%d" y="%d" text-anchor="middle" `+
 			`font-family="'GitLab Sans','Segoe UI',-apple-system,BlinkMacSystemFont,sans-serif" `+
 			`font-size="11" font-weight="600" letter-spacing="0.8" fill="#868686">%s</text>`,
-			cx+chipW/2, topPad+13, svgEsc(strings.ToUpper(stage)))
+			centerX, topPad+13, svgEsc(strings.ToUpper(col.stage)))
 
-		// Subtle separator line below stage name.
+		// Separator line spanning all sub-columns.
 		wf(`  <line x1="%d" y1="%d" x2="%d" y2="%d" stroke="#eaeaea" stroke-width="1"/>`,
-			cx, topPad+21, cx+chipW, topPad+21)
+			x1, topPad+21, x2, topPad+21)
 
-		// Job chips – each wrapped in a <g> with tooltip metadata.
+		drawnHeader[col.stage] = true
-		for j, name := range jobs {
+	}
+
+	// ── Connectors (drawn before chips so they appear behind job blocks) ───────
+	const connStroke = "#dbdbdb"
+
+	if dagMode {
+		// Job-to-job bezier curves from needs:.
+		for _, name := range visible {
+			for _, need := range p.Jobs[name].Needs {
+				dep := needsJobName(need)
+				if dep == "" {
+					continue
+				}
+				src, okS := rightMid[dep]
+				dst, okD := leftMid[name]
+				if !okS || !okD {
+					continue
+				}
+				cpX := (src.x + dst.x) / 2
+				wf(`  <path d="M%d,%d C%d,%d %d,%d %d,%d" stroke="%s" stroke-width="2" fill="none"/>`,
+					src.x, src.y, cpX, src.y, cpX, dst.y, dst.x-7, dst.y, connStroke)
+				// Small right-pointing triangle arrowhead.
+				wf(`  <polygon points="%d,%d %d,%d %d,%d" fill="%s"/>`,
+					dst.x-7, dst.y-4, dst.x, dst.y, dst.x-7, dst.y+4, connStroke)
+			}
+		}
+	} else {
+		// Classic: bus-bar connectors between adjacent stage columns.
+		// Every job in stage[i] fans to a vertical bus at the midpoint gap,
+		// then fans out to every job in stage[i+1].
+		for ci := 0; ci < len(cols)-1; ci++ {
+			x1 := colX[ci] + chipW    // right edge of current column
+			x2 := colX[ci+1]          // left edge of next column
+			midX := (x1 + x2) / 2
+
+			srcJobs := cols[ci].jobs
+			dstJobs := cols[ci+1].jobs
+
+			// Collect all Y midpoints to span the vertical bus bar.
+			var allYs []int
+			srcY := make([]int, len(srcJobs))
+			dstY := make([]int, len(dstJobs))
+			for j, name := range srcJobs {
+				srcY[j] = rightMid[name].y
+				allYs = append(allYs, srcY[j])
+			}
+			for j, name := range dstJobs {
+				dstY[j] = leftMid[name].y
+				allYs = append(allYs, dstY[j])
+			}
+			sort.Ints(allYs)
+			busMinY, busMaxY := allYs[0], allYs[len(allYs)-1]
+
+			// Vertical bus bar at midX (only when there are multiple Y levels).
+			if busMinY < busMaxY {
+				wf(`  <line x1="%d" y1="%d" x2="%d" y2="%d" stroke="%s" stroke-width="2"/>`,
+					midX, busMinY, midX, busMaxY, connStroke)
+			}
+
+			// Horizontal stubs from each source job to the bus.
+			for _, y := range srcY {
+				wf(`  <line x1="%d" y1="%d" x2="%d" y2="%d" stroke="%s" stroke-width="2"/>`,
+					x1, y, midX, y, connStroke)
+			}
+
+			// Horizontal stubs from bus to each destination job, with arrowhead.
+			for _, y := range dstY {
+				wf(`  <line x1="%d" y1="%d" x2="%d" y2="%d" stroke="%s" stroke-width="2"/>`,
+					midX, y, x2-7, y, connStroke)
+				wf(`  <polygon points="%d,%d %d,%d %d,%d" fill="%s"/>`,
+					x2-7, y-4, x2, y, x2-7, y+4, connStroke)
+			}
+		}
+	}
+
+	// ── Job chips (drawn after connectors so they appear on top) ─────────────
+	for ci, col := range cols {
+		for j, name := range col.jobs {
 			job := p.Jobs[name]
 			chy := topPad + labelH + j*(chipH+chipGap)
 			isSkipped := skippedJobs[name]
+			cx := colX[ci]
 
 			// Build <desc> content: shown in the HTML sidebar and SVG viewer tooltips.
-			desc := "stage: " + stage
+			desc := "stage: " + col.stage
 			if job.When != "" {
 				desc += "\nwhen: " + job.When
 			}
@@ -285,82 +489,6 @@ func pipelineSVG(p *model.Pipeline, ctx *cicontext.Context) string {
 		}
 	}
 
-	// ── Connectors ────────────────────────────────────────────────────────────
-	const connStroke = "#dbdbdb"
-
-	if dagMode {
-		// Job-to-job bezier curves from needs:.
-		for _, name := range visible {
-			for _, need := range p.Jobs[name].Needs {
-				dep := needsJobName(need)
-				if dep == "" {
-					continue
-				}
-				src, okS := rightMid[dep]
-				dst, okD := leftMid[name]
-				if !okS || !okD {
-					continue
-				}
-				cpX := (src.x + dst.x) / 2
-				wf(`  <path d="M%d,%d C%d,%d %d,%d %d,%d" stroke="%s" stroke-width="2" fill="none"/>`,
-					src.x, src.y, cpX, src.y, cpX, dst.y, dst.x-7, dst.y, connStroke)
-				// Small right-pointing triangle arrowhead.
-				wf(`  <polygon points="%d,%d %d,%d %d,%d" fill="%s"/>`,
-					dst.x-7, dst.y-4, dst.x, dst.y, dst.x-7, dst.y+4, connStroke)
-			}
-		}
-	} else {
-		// Classic: bus-bar connectors – every job in stage[i] fans to a vertical
-		// bus at the midpoint column gap, then fans out to every job in stage[i+1].
-		// This gives each job pair its own visual connection instead of a single
-		// center-to-center line that misses jobs at the top and bottom of columns.
-		for i := 0; i < len(stages)-1; i++ {
-			x1 := sidePad + i*(chipW+stageGap) + chipW // right edge of left column
-			x2 := sidePad + (i+1)*(chipW+stageGap)    // left edge of right column
-			midX := (x1 + x2) / 2
-
-			srcJobs := byStage[stages[i]]
-			sort.Strings(srcJobs)
-			dstJobs := byStage[stages[i+1]]
-			sort.Strings(dstJobs)
-
-			// Collect all Y midpoints to span the vertical bus bar.
-			var allYs []int
-			srcY := make([]int, len(srcJobs))
-			dstY := make([]int, len(dstJobs))
-			for j, name := range srcJobs {
-				srcY[j] = rightMid[name].y
-				allYs = append(allYs, srcY[j])
-			}
-			for j, name := range dstJobs {
-				dstY[j] = leftMid[name].y
-				allYs = append(allYs, dstY[j])
-			}
-			sort.Ints(allYs)
-			busMinY, busMaxY := allYs[0], allYs[len(allYs)-1]
-
-			// Vertical bus bar at midX (only drawn when there are multiple Y levels).
-			if busMinY < busMaxY {
-				wf(`  <line x1="%d" y1="%d" x2="%d" y2="%d" stroke="%s" stroke-width="2"/>`,
-					midX, busMinY, midX, busMaxY, connStroke)
-			}
-
-			// Horizontal stubs from each source job to the bus.
-			for _, y := range srcY {
-				wf(`  <line x1="%d" y1="%d" x2="%d" y2="%d" stroke="%s" stroke-width="2"/>`,
-					x1, y, midX, y, connStroke)
-			}
-
-			// Horizontal stubs from bus to each destination job, with arrowhead.
-			for _, y := range dstY {
-				wf(`  <line x1="%d" y1="%d" x2="%d" y2="%d" stroke="%s" stroke-width="2"/>`,
-					midX, y, x2-7, y, connStroke)
-				wf(`  <polygon points="%d,%d %d,%d %d,%d" fill="%s"/>`,
-					x2-7, y-4, x2, y, x2-7, y+4, connStroke)
-			}
-		}
-	}
-
 	// ── Legend ────────────────────────────────────────────────────────────────
 	legend := []struct{ color, label string }{
 		{"#1f75cb", "regular"},

internal/graph/render_test.go

@@ -520,6 +520,138 @@ func TestRenderHTML_WriteFileFails(t *testing.T) {
 	}
 }
 
+// ── computeColumns ────────────────────────────────────────────────────────────
+
+func TestComputeColumns_NoIntraNeeds(t *testing.T) {
+	// No intra-stage needs → one column per stage.
+	stages := []string{"build", "test"}
+	byStage := map[string][]string{
+		"build": {"build-a", "build-b"},
+		"test":  {"test-a"},
+	}
+	jobs := map[string]model.Job{
+		"build-a": {Name: "build-a", Stage: "build"},
+		"build-b": {Name: "build-b", Stage: "build"},
+		"test-a":  {Name: "test-a", Stage: "test"},
+	}
+	cols := computeColumns(stages, byStage, jobs)
+	if len(cols) != 2 {
+		t.Fatalf("expected 2 columns (one per stage), got %d", len(cols))
+	}
+	if cols[0].stage != "build" || len(cols[0].jobs) != 2 {
+		t.Errorf("col[0]: got stage=%q jobs=%v", cols[0].stage, cols[0].jobs)
+	}
+	if cols[1].stage != "test" || len(cols[1].jobs) != 1 {
+		t.Errorf("col[1]: got stage=%q jobs=%v", cols[1].stage, cols[1].jobs)
+	}
+}
+
+func TestComputeColumns_IntraNeeds(t *testing.T) {
+	// job-b depends on job-a in the same stage → split into 2 sub-columns.
+	stages := []string{"build"}
+	byStage := map[string][]string{
+		"build": {"job-a", "job-b"},
+	}
+	jobs := map[string]model.Job{
+		"job-a": {Name: "job-a", Stage: "build"},
+		"job-b": {Name: "job-b", Stage: "build", Needs: []any{"job-a"}},
+	}
+	cols := computeColumns(stages, byStage, jobs)
+	if len(cols) != 2 {
+		t.Fatalf("expected 2 sub-columns, got %d", len(cols))
+	}
+	// Both belong to the same stage.
+	if cols[0].stage != "build" || cols[1].stage != "build" {
+		t.Error("both sub-columns should be in stage 'build'")
+	}
+	// job-a has no same-stage deps → depth 0 → first sub-column.
+	if len(cols[0].jobs) != 1 || cols[0].jobs[0] != "job-a" {
+		t.Errorf("col[0] should contain job-a, got %v", cols[0].jobs)
+	}
+	// job-b depends on job-a → depth 1 → second sub-column.
+	if len(cols[1].jobs) != 1 || cols[1].jobs[0] != "job-b" {
+		t.Errorf("col[1] should contain job-b, got %v", cols[1].jobs)
+	}
+}
+
+func TestComputeColumns_EmptyStageSkipped(t *testing.T) {
+	// Stages with no jobs are silently skipped.
+	stages := []string{"build", "empty", "test"}
+	byStage := map[string][]string{
+		"build": {"j1"},
+		"test":  {"j2"},
+	}
+	jobs := map[string]model.Job{
+		"j1": {Name: "j1", Stage: "build"},
+		"j2": {Name: "j2", Stage: "test"},
+	}
+	cols := computeColumns(stages, byStage, jobs)
+	if len(cols) != 2 {
+		t.Fatalf("expected 2 columns (empty stage skipped), got %d", len(cols))
+	}
+}
+
+func TestComputeColumns_CrossStageNeedIgnored(t *testing.T) {
+	// job-b has needs: for both job-a (same stage) and prev-job (different stage).
+	// The cross-stage need must be ignored when computing topological depth.
+	stages := []string{"build"}
+	byStage := map[string][]string{
+		"build": {"job-a", "job-b"},
+	}
+	jobs := map[string]model.Job{
+		"job-a": {Name: "job-a", Stage: "build"},
+		"job-b": {Name: "job-b", Stage: "build", Needs: []any{"job-a", "prev-job"}},
+		"prev-job": {Name: "prev-job", Stage: "prepare"},
+	}
+	cols := computeColumns(stages, byStage, jobs)
+	// Still split into 2 sub-columns; the cross-stage need is ignored.
+	if len(cols) != 2 {
+		t.Fatalf("expected 2 sub-columns, got %d", len(cols))
+	}
+	if cols[0].jobs[0] != "job-a" {
+		t.Errorf("expected job-a in first sub-column, got %v", cols[0].jobs)
+	}
+	if cols[1].jobs[0] != "job-b" {
+		t.Errorf("expected job-b in second sub-column, got %v", cols[1].jobs)
+	}
+}
+
+func TestComputeColumns_CycleGuard(t *testing.T) {
+	// Intra-stage cycle must not hang (cycle guard returns depth 0).
+	stages := []string{"build"}
+	byStage := map[string][]string{
+		"build": {"a", "b"},
+	}
+	jobs := map[string]model.Job{
+		"a": {Name: "a", Stage: "build", Needs: []any{"b"}},
+		"b": {Name: "b", Stage: "build", Needs: []any{"a"}},
+	}
+	cols := computeColumns(stages, byStage, jobs)
+	// Should return without hanging; exact column count is implementation-defined.
+	if len(cols) == 0 {
+		t.Error("expected at least one column")
+	}
+}
+
+func TestPipelineSVG_IntraStageDAG(t *testing.T) {
+	// job-b depends on job-a in the same stage → SVG should render both,
+	// placed as sub-columns (sub-stage gap between them).
+	p := &model.Pipeline{
+		Stages: []string{"build"},
+		Jobs: map[string]model.Job{
+			"job-a": {Name: "job-a", Stage: "build"},
+			"job-b": {Name: "job-b", Stage: "build", Needs: []any{"job-a"}},
+		},
+	}
+	svg := pipelineSVG(p, nil)
+	if !strings.Contains(svg, "job-a") { t.Error("expected job-a in SVG") }
+	if !strings.Contains(svg, "job-b") { t.Error("expected job-b in SVG") }
+	// DAG mode: bezier connectors between the two jobs.
+	if !strings.Contains(svg, "<path") {
+		t.Error("expected bezier <path> connector between intra-stage jobs")
+	}
+}
+
 // ── htmlPage ──────────────────────────────────────────────────────────────────
 
 func TestHtmlPage(t *testing.T) {

internal/model/fuzz_test.go

@@ -0,0 +1,77 @@
+package model
+
+import "testing"
+
+// FuzzParseBytes ensures the YAML parser never panics on arbitrary input and
+// that successful parses return a structurally sound Pipeline.
+// Run with: go test -fuzz=FuzzParseBytes ./internal/model/
+// Found failures are saved to testdata/fuzz/FuzzParseBytes/.
+func FuzzParseBytes(f *testing.F) {
+	// Seed corpus: representative inputs covering the main code paths in
+	// ParseBytes, including the sanitizeYAMLEscapes pre-processing step.
+	seeds := [][]byte{
+		{},
+		[]byte("null"),
+		[]byte("stages: [build]\nbuild-job:\n  stage: build\n  script: echo ok\n"),
+		[]byte(".base:\n  script: [make]\nchild:\n  extends: .base\n  stage: test\n"),
+		[]byte("stages: [a, b]\njob-a:\n  stage: a\n  script: run\njob-b:\n  stage: b\n  needs: [job-a]\n  script: run\n"),
+		[]byte("*undefined_anchor"),
+		[]byte("- item1\n- item2\n"),
+		[]byte("my-job: \"just a string\"\n"),
+		[]byte("my-job:\n  stage: [build, test]\n"),
+		[]byte("# glint: ignore GL007\nlegacy:\n  only: [main]\n  script: ok\n"),
+		[]byte("workflow:\n  rules:\n    - if: '$CI_COMMIT_BRANCH == \"main\"'\n      when: always\njob:\n  script: echo ok\n"),
+		[]byte("include:\n  - local: other.yml\njob:\n  script: echo ok\n"),
+		[]byte("job:\n  script: echo ok\n  when: on_failure\n  rules:\n    - if: '$VAR =~ /^us\\//'\n"),
+		[]byte("job:\n  stage: test\n  image:\n    name: golang:1.21\n    entrypoint: ['']\n  parallel:\n    matrix:\n      - PLATFORM: [linux, darwin]\n  script: go build\n"),
+		[]byte("default:\n  retry: 2\n  timeout: 1h30m\nvariables:\n  ENV: production\nstages: [build, test, deploy]\n"),
+		[]byte("&anchor\n  script: [echo ok]\njob:\n  <<: *anchor\n  stage: build\n"),
+	}
+	for _, s := range seeds {
+		f.Add(s)
+	}
+
+	f.Fuzz(func(t *testing.T, data []byte) {
+		p, err := ParseBytes(data)
+		if err != nil {
+			return // errors are acceptable; panics are not
+		}
+		if p == nil {
+			t.Fatal("ParseBytes returned nil pipeline with nil error")
+		}
+		for name := range p.Jobs {
+			if name == "" {
+				t.Fatal("ParseBytes produced a job with an empty name")
+			}
+		}
+	})
+}
+
+// FuzzSanitizeYAMLEscapes ensures the escape sanitizer never panics and never
+// produces output shorter than its input (it can only expand \/ to \\/).
+func FuzzSanitizeYAMLEscapes(f *testing.F) {
+	seeds := [][]byte{
+		{},
+		[]byte("stage: build"),
+		[]byte(`if: "$CI_BRANCH =~ /^us\//"`),
+		[]byte(`"pattern: /^us\//"`),
+		[]byte(`'single quoted \/ unchanged'`),
+		[]byte(`"\n\t\r"`),
+		[]byte(`"nested \"quote\" inside"`),
+		[]byte(`'it''s fine'`),
+		[]byte(`"unclosed`),
+		{'"', '\\'},      // double-quoted string ending with a lone backslash
+		{'"', '\\', '/'}, // the exact sequence being rewritten
+	}
+	for _, s := range seeds {
+		f.Add(s)
+	}
+
+	f.Fuzz(func(t *testing.T, data []byte) {
+		out := sanitizeYAMLEscapes(data)
+		if len(out) < len(data) {
+			t.Fatalf("sanitizeYAMLEscapes shrank output: input len=%d output len=%d\ninput: %q",
+				len(data), len(out), data)
+		}
+	})
+}