Fix core simulation and modernize PyG environment - #52

Summary

• validate virtual-network lifecycle inputs and make event/accounting behavior deterministic
• add versioned metrics semantics with legacy aggregate-summary compatibility
• make deployment accounting transactional and failure classification exhaustive
• register joint-PR solvers and roll back partial mappings on failure
• upgrade the supported stack to PyTorch 2.11.0 and PyG 2.8.0.post1
• make optional PyG compiled extensions non-mandatory at import time
• support CPU installs on Linux/macOS and supported CUDA builds on Linux
• test Python 3.10 and 3.11 in CI and update installation documentation

Root cause

The core simulation accepted invalid lifecycle values, used fragile aggregation and floating-point resource accounting, and left some failed requests under-classified. The joint-PR solvers were not imported into the registry and could return after partial failures without undoing earlier mappings.

The package also pinned platform-specific PyG wheel URLs in project metadata and attempted package installation during module import. That made installation dependent on one Python/OS/accelerator combination and caused imports to mutate the environment.

Validation

• .venv/bin/python -m pytest -q — 251 passed
• wheel build from pyproject.toml
• python -m pip check
• shell syntax and YAML parsing checks
• imports verified with optional torch_sparse, torch_scatter, and torch_cluster deliberately unavailable
• exhaustive audit over 80 small graph cases: no invalid successes, exceptions, input mutation, or oracle disagreement
• deterministic 30-request end-to-end comparison across order, random, and FFD joint-PR solvers, including resource restoration and failure-breakdown invariants