perf(metrics): optimize LakosMetrics transitive dependency graph

[ThanosTsiamis]

Description

This Pull Request addresses the performance hotspot in LakosMetrics by overhauling how transitive reachability is computed within MetricsComponentDependencyGraph.

The Problem

Previously, getTransitiveDependenciesOf() executed a fresh Depth-First Search (DFS) from scratch for every single component. On complex architecture graphs with shared downstream subgraphs or deep/cyclic dependencies, this caused heavily redundant traversals, driving the runtime complexity up to O(V * (V + E)). It also introduced significant garbage collection pressure by instantiating new HashSet collections on every query.

The Solution

Instead of query-time DFS traversals, this PR introduces a robust, primitive-backed precomputation engine executed once during the graph's initialization:

1. Cycle Detection (Kosaraju's Algorithm): determineStronglyConnectedComponents() runs a two-pass iterative DFS (forward to find finishing times, reverse on the transposed graph) to safely group cyclic dependencies into Strongly Connected Components (SCCs).
2. DAG Condensation: The component graph is condensed into a Directed Acyclic Graph (DAG) where each node represents an isolated SCC.
3. Topological BitSet Propagation: The DAG is sorted topologically, and downstream reachability is propagated using fast bitwise operations (BitSet.or()). This ensures that shared downstream paths are evaluated exactly once.
4. O(1) Queries: The final reachability matrices are mapped back to their respective components and cached, changing getTransitiveDependenciesOf() from an expensive graph traversal into an O(1) map lookup.

Additional Robustness & Safeguards

• Stack Safety: The algorithm completely avoids recursion in favor of an explicit iterative stack approach (ArrayDeque<TraversalFrame>), protecting the JVM against StackOverflowError on deep, linear dependency chains.
• Cache Locality: Internal graph representations are flattened into primitive int[][], int[], and boolean[] arrays to optimize CPU cache locality and reduce memory footprint.
• Duplicate Resilience: Rather than utilizing a strict ImmutableMap.Builder (which crashes with an IllegalArgumentException if duplicate elements are fed from upstream), the precomputation safely compiles results into an intermediate LinkedHashMap before generating the final immutable copy.
• Semantic Parity: Existing cycle and self-reachability rules are perfectly preserved. A component only transitively reaches itself if it belongs to an SCC of size > 1, or contains a direct self-loop.

---

Related Issue

Closes #1628

[hankem]

Thank you for this improvement!

On my system, it reduces the time to compute

ArchitectureMetrics.lakosMetrics(MetricsComponents.fromClasses(classes))

for hibernate-core:7.3.6 (with 37,281 classes)

• from (101±2) s with ArchUnit 1.4.2
• by a factor of more than 16 to (6.2±0.3) s. 🎉