Make CPU SHM pool importable without NumPy and modernize core deps/CI/telemetry

[nanocubit]

Motivation

• Allow basic CPU shared-memory smoke checks (allocate/free/coalescing) to run in constrained environments where installing NumPy or heavy GPU deps is not possible.
• Reduce heavy imports at module-import time and split expensive GPU/Ray dependencies into optional profiles to make core installs and CI CPU jobs lighter.
• Improve observability and robustness of runtime/server/worker paths with structured logging and telemetry hooks for allocation/lease events and failures.

Description

• Made NumPy optional in zerolink/core/cpu/shm_pool.py by using try/except ImportError and added a runtime guard in get_numpy_array that raises a clear RuntimeError if NumPy is absent, and replaced np.dtype typing with Any to avoid import-time failures.
• Reworked dependency layout: moved heavy deps out of base install, added dependency profile files (requirements-gpu.txt, requirements-ray.txt, requirements-cgpu.txt, requirements-dev.txt), added numpy>=1.24.0 to core requirements.txt, and updated pyproject.toml and setup.py to support extras (gpu, ray, cgpu, dev, full) and to allow installation without Torch/CUDA toolchain.
• Updated CI (.github/workflows/ci.yml) to split test-core, test-gpu-profile, and test-ray-profile jobs, pin newer Python versions and run CPU-only tests with --noconftest where appropriate.
• Introduced structured logging and telemetry improvements: renamed pynexus metrics to zerolink_*, added new metrics and helper functions (observe_runtime_latency, record_alloc_failure, record_lease_event) in zerolink/monitoring/telemetry.py, and wired telemetry/logging calls into runtime, server and worker code paths (zerolink/runtime/unified.py, zerolink/server/main_server.py, zerolink/workers/gpu_worker.py).
• Hardened runtime/server/worker behavior: safer SHM/socket defaults (/tmp/zerolink.sock, zerolink_cpu_pool), JSON-structured log messages, better MainServer/MainIPCLeaseManager2P lifecycle and lease bookkeeping, safer handling of missing GPU components, and improved graceful shutdown logic.
• Cleanups and minor refactors: small docs/README updates, fix name replacements (pynexus -> zerolink), defensive parsing of requirements in setup.py, optional CUDA extension build when Torch is available, and tests conftest adjustments to be CPU-safe.

Testing

• Compiled the modified module with python -m py_compile zerolink/core/cpu/shm_pool.py which succeeded.
• Ran a CPU SHM coalescing smoke script using SharedMemoryPool (create pool, allocate two blocks, free both, assert coalescing, cleanup) which printed shm coalescing ok and returned successfully.
• Executed protocol unit tests with pytest tests/test_protocol.py -q and again with pytest --noconftest tests/test_protocol.py -q, both runs passed (3 passed).

---

Codex Task

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 58ab5b38aa

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Fix Python 3.8 annotation that breaks setup import

The new return annotation list[str] requires PEP 585 (Python ≥3.9). Because setup.py is executed at import time during installation, this will raise TypeError: 'type' object is not subscriptable on Python 3.8, even though python_requires='>=3.8' still advertises 3.8 support. Use typing.List[str] or postpone evaluation with from __future__ import annotations to keep 3.8 installs working.

Useful? React with 👍 / 👎.