perf: replace manual AST walk with tree-sitter Query API for JS/TS/TSX extraction

WASM build speed regressed from 5.0 ms/file (v2.0.0) to 6.6 ms/file (v2.1.0+)
due to excessive WASM-JS boundary crossings in the JS extractor — every
node.type, .text, .childForFieldName() call crosses the boundary.

Phase 1 — cache redundant boundary crossings:
- Fix findChild() calling node.child(i) twice per match
- Cache node.type, node.startPosition.row, prop.text into locals across
  class_declaration, export_statement, extractCallInfo, walkCallChain,
  findParentClass, and variable_declaration cases

Phase 2 — tree-sitter Query API extraction:
- Import Query from web-tree-sitter, compile S-expression patterns per
  language (JS vs TS/TSX) during createParsers(), cache in module-level Map
- New extractSymbolsQuery() uses query.matches(rootNode) to run pattern
  matching entirely inside WASM, reducing boundary crossings from
  O(nodes × properties) to O(matches × captures)
- Post-processing only runs on matched nodes (tens per file) not all AST
  nodes (thousands per file)
- Manual tree walk preserved as fallback for backward compatibility when
  query is not provided

Result: 6.6 → 4.4 ms/file (−33%), below v2.0.0 baseline of 5.0 ms/file.
All 536 tests pass, node/edge counts identical.

Impact: 12 functions changed, 20 affected

Author: github-actions[bot]

src/extractors/helpers.js

@@ -4,7 +4,8 @@ export function nodeEndLine(node) {
 
 export function findChild(node, type) {
   for (let i = 0; i < node.childCount; i++) {
-    if (node.child(i).type === type) return node.child(i);
+    const child = node.child(i);
+    if (child.type === type) return child;
   }
   return null;
 }