Replace the recursive SQL CTE in DetectCycles with Go-layer DFS using
shared visited set. The previous implementation enumerated all paths
through the dependency graph, causing exponential blowup with diamond
patterns (multiple issues depending on the same target).
Changes:
- Add loadDependencyGraph() to load deps as adjacency list in one query
- Implement DFS cycle detection with recStack for back-edge detection
- Add normalizeCycle() for consistent cycle deduplication
- Add DetectCycles-specific benchmarks (Linear, Dense, Tree graphs)
- Use direct SQL INSERT in benchmarks to bypass AddDependency overhead
Performance improvement on dense graph (500 nodes, 2500 edges):
- Before: >120s timeout
- After: 1.6ms
Benchmarks:
- DetectCycles_Linear_1000: 0.84ms (1000 nodes, 999 edges)
- DetectCycles_Dense_500: 1.59ms (500 nodes, ~2500 edges)
- DetectCycles_Tree_1000: 0.85ms (1000 nodes, 999 edges)