feat: audit phase 6 (T2) — on-canvas labels and Cmd/Ctrl-K search

Even with the HUD from T1, the tree itself was 57 wordless icons: no way to
answer "does he know NestJS?" without clicking around.

Search palette (Cmd/Ctrl-K):
- New utils/search.ts — pure, 9 unit tests. Subsequence matching on short
  fields ("wbrtc" finds WebRTC), plus tech-stack and company matching so
  "Prisma" surfaces the projects using it and "Peraton" finds the role.
- Descriptions deliberately use substring, NOT subsequence matching. Over
  long prose a subsequence match is almost always spurious — the first
  version matched "Prisma" as p...r...i...s...m...a spread across an
  unrelated sentence, which a test caught.
- Selecting a result flies the camera to that node and opens its modal.
- Built to the ARIA combobox pattern: input keeps focus and owns
  aria-activedescendant, results are a listbox. Options intentionally have no
  keyboard handler — making them focusable would break the pattern — so the
  svelte-ignore there is documented rather than incidental.

On-canvas labels:
- A Text per node, hidden below LABEL_ZOOM_THRESHOLD and counter-scaled by
  1/zoom so they stay a constant size on screen rather than ballooning.
- Threshold sits just above the settled desktop zoom, so the default view
  stays a clean constellation and zooming in reveals labels. 'L' forces them
  on at any zoom.
- Long labels truncate at 18 chars; full text is still in the tooltip/modal.
- Collision culling: labels are placed biggest-node-first, and any whose box
  would overlap an already-placed label or someone else's node is dropped.
  Dense clusters simply show fewer labels until you zoom further in. This is
  purely presentational — no node coordinates changed.

Three bugs found during review:
- The entrance zoom tween only called updateParallax, so the label
  counter-scale stayed frozen at its pre-zoom value (labels ended up at 1/4
  size and effectively vanished). It now updates labels too, as does the
  skip/reduced-motion path that sets scale directly.
- updateLabels forced alpha = 1, so labels appeared over nodes that had not
  faded in yet — visible during the splash screen. Alpha now belongs to the
  entrance and labels arrive with their node.
- animation.ts imported updateParallax only; it needed updateLabels as well.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

src/lib/utils/search.ts

@@ -0,0 +1,93 @@
+// ==========================================================================
+// SEARCH — subsequence matching over the tree data. Pure, no DOM, no PixiJS.
+// ==========================================================================
+
+import { treeNodes, type TreeNode } from '$lib/data';
+
+export interface SearchResult {
+	node: TreeNode;
+	score: number;
+	/** Why it matched, for display ("React" vs "tech: React"). */
+	context?: string;
+}
+
+/**
+ * Case-insensitive subsequence match: "wbrtc" matches "WebRTC".
+ *
+ * Returns a score where lower is better, or null for no match. Consecutive
+ * runs and an early first match are rewarded, so "react" ranks the React node
+ * above a project that merely lists React in its stack.
+ */
+function subsequenceScore(haystack: string, needle: string): number | null {
+	if (!needle) return 0;
+	const h = haystack.toLowerCase();
+	const n = needle.toLowerCase();
+
+	// Exact and prefix matches short-circuit to the best scores.
+	if (h === n) return -1000;
+	if (h.startsWith(n)) return -500 + h.length;
+
+	let score = 0;
+	let hi = 0;
+	let lastHit = -1;
+	for (let ni = 0; ni < n.length; ni++) {
+		const found = h.indexOf(n[ni], hi);
+		if (found === -1) return null;
+		// Penalise gaps; reward adjacency.
+		score += found === lastHit + 1 ? 0 : found - hi + 1;
+		lastHit = found;
+		hi = found + 1;
+	}
+	// Prefer shorter haystacks — "Git" should beat "GitHub Actions" for "git".
+	return score + h.length * 0.1;
+}
+
+/** Search every node by label, then category, then description/tech stack. */
+export function searchNodes(query: string, nodes: TreeNode[] = treeNodes): SearchResult[] {
+	const q = query.trim();
+	if (!q) return [];
+
+	const results: SearchResult[] = [];
+
+	for (const node of nodes) {
+		const labelScore = subsequenceScore(node.label, q);
+		if (labelScore !== null) {
+			results.push({ node, score: labelScore });
+			continue;
+		}
+
+		// Tech stack is a strong signal — searching "prisma" should surface the
+		// projects that use it even though no node is named Prisma.
+		if (node.category === 'projects' && node.techStack) {
+			const tech = node.techStack.find((t) => subsequenceScore(t, q) !== null);
+			if (tech) {
+				results.push({ node, score: 200, context: `tech: ${tech}` });
+				continue;
+			}
+		}
+
+		if (node.category === 'experience' && node.company) {
+			const companyScore = subsequenceScore(node.company, q);
+			if (companyScore !== null) {
+				results.push({ node, score: companyScore + 100, context: node.company });
+				continue;
+			}
+		}
+
+		// Descriptions use plain substring matching, not subsequence. Over long
+		// prose a subsequence match is almost always spurious — "Prisma" found a
+		// p...r...i...s...m...a spread across an unrelated sentence.
+		const haystack = node.description.toLowerCase();
+		const needle = q.toLowerCase();
+		const at = haystack.indexOf(needle);
+		if (at !== -1) {
+			results.push({
+				node,
+				score: 400 + at,
+				context: node.description.slice(Math.max(0, at - 20), at + 45).trim()
+			});
+		}
+	}
+
+	return results.sort((a, b) => a.score - b.score);
+}