fix: stop using InfoNode.index as subnetwork scratch state

[antoneri]

Summary

Stop using InfoNode.index as scratch storage while generating sub-Infomap networks.

InfomapBase::generateSubNetwork(InfoNode&) previously overwrote each child node's index with its temporary position in the new subnetwork, then reused that mutated state to clone internal edges. That worked as a local shortcut, but it also leaked a write to a field that carries real optimizer and tree state elsewhere in the codebase.

This change replaces that side-channel with a local InfoNode* -> position map and adds a regression test that proves subnetwork generation no longer clobbers parent node indices.

Problem

The old implementation did this during subnetwork construction:

• clone each child node
• write node.index = childIndex
• later use edge.source->index / edge.target->index to find the cloned endpoints

That meant generateSubNetwork(...) mutated caller-visible InfoNode.index state as an implementation detail and never restored it.

Why that is a bug:

• InfoNode.index is not scratch-only state
• it is used elsewhere for optimizer module assignment, consolidation, and other tree/partition bookkeeping
• a helper that silently rewrites it creates hidden coupling and makes the surrounding code harder to reason about

Even when the mutation does not immediately surface as a user-visible partition bug, it is still state corruption at the function boundary.

Fix

In src/core/InfomapBase.cpp, replace the temporary node.index write with a local std::unordered_map<InfoNode*, unsigned int>:

• map original child node pointer -> cloned subnetwork position
• use that map when cloning internal edges
• leave the original nodes' index values untouched

Tests

Add a regression in test/cpp/test_infomap_wrapper.cpp that:

• builds a small partitioned fixture
• records sentinel index values on the parent module's children
• generates the subnetwork
• verifies those sentinel indices are preserved
• verifies the generated subnetwork clones the module's internal edge set exactly

Also add the required -fno-access-control compile option for the lifecycle test target in CMakeLists.txt so the regression can exercise the internal path directly.

Risk

Medium.

This changes C++ core logic, but the surface is tightly bounded:

• one helper function
• no public API changes
• behavior is preserved except for removing the leaked side effect on InfoNode.index

The new mapping is local to subnetwork generation and does not change the cloned graph semantics.

Verification

Ran:

• PATH="/opt/homebrew/bin:$PATH" make test-native JOBS=1

Why This Should Merge Independently

This fix stands on its own.

It is not inherently tied to the larger CSR experimentation branch; that work only made the hidden coupling around InfoNode.index more obvious. Removing the side-channel is independently valuable because it:

• makes generateSubNetwork(...) respect its caller's state
• removes a hidden dependency on a semantically overloaded field
• adds a focused regression for a fragile internal path