Normalize correctness for the Capsule Network second wave

[TennisGazelle]

Second-wave normalization

This PR is intentionally stacked on top of the preservation/revival branch from PR #6.

Base: revival/repository-foundation
Boundary tag: second-wave-start -> 0e9428abbede0d8e960d4a6b467c810926e0e4cd

The preservation PR documents the 2018 artifact without silently changing behavior. This PR begins the corrected second wave: characterize/fix known correctness and reproducibility defects, establish host-testable algorithm boundaries, modernize the build, and leave CUDA compile/runtime verification as the final explicit handoff for a GPU box.

Completed so far

• normalized seedable RNG semantics and weight initialization
• removed double-epsilon vector norm behavior
• hardened capsule flatten/reshape validation
• corrected strict Pareto dominance, crowding preference, rank selection, population stats, and nondominated sorting
• removed the development fakeEvaluate() call from the real GA evaluation path
• separated host-testable GA core from CUDA-dependent fitness evaluation
• corrected offspring generation so crossover never mutates the parent population
• added host tests for utility, NSGA-II, and generation invariants

Validation policy

Each host-verifiable goal is committed separately and must pass the GitHub-hosted make ci workflow before the next correctness layer is considered complete. CUDA-specific compile/runtime claims will remain explicitly unverified until run on a real CUDA host.

The final version of this PR body will include:

• goal-by-goal commit and behavior ledger
• historical-compatibility notes for every corrected behavior
• exact host validation commands
• exact contemporary CUDA configure/build/test commands
• GPU parity commands
• failure-triage commands
• copy/paste continuation prompts for a future AI agent if CUDA compilation or numerical parity fails

Do not merge this PR based only on host CI if the CUDA section is still marked unverified.