perf(parser): Phase B — TreeCursor migration for parser.Walk

[aksOps]

Summary

Phase B of the enrich OOM fix plan. Single task: rewrite `parser.Walk` from recursive `Node.Child(i)` to iterative TreeCursor traversal. Public API unchanged; all callers untouched.

Honest accounting

The plan estimated 90%+ allocation reduction from TreeCursor based on a generic claim that cursors avoid per-node allocation. In smacker/go-tree-sitter v0.x, `TreeCursor.CurrentNode()` still routes through `Tree.cachedNode` which heap-allocates a `*Node` on first visit per node. So Phase B does not drop allocations further from where Phase A already landed.

What this PR DOES deliver:

• Removes Go-level recursion frames per descent (stack discipline).
• Matches the canonical tree-sitter walking idiom (code clarity).
• Preserves determinism — pre-order DFS visitation order matches the recursive form exactly.

The 91%-of-allocations `cachedNode` hot spot pprof flagged on airflow was driven by per-node re-parse. Task A1 fixed that by parsing once per file; the additional allocation savings from Phase B alone are within noise.

Why ship it anyway

• Plan structure requires 4 phase PRs; this is Phase B.
• Future binding upgrades (or a switch to a binding that exposes a non-allocating cursor API) would benefit from already being on the cursor path.
• Code is shorter and matches tree-sitter's documented traversal idiom.

Test plan

• `go test ./internal/parser/... ./internal/intelligence/extractor/...` — 39 pass across 6 packages
• `go test ./... -count=1` — 877 pass (unchanged)
• `go vet ./...` clean
• CI on this PR