[sfloess]

Summary

Implements Phase 1 of #5: runtime-neutral hardware discovery and a versioned, serializable profile.

Included

• Standard-library profile dataclasses for host, CPU, and GPU facts.
• Linux CPU/memory discovery from /proc and GPU enumeration from DRM sysfs.
• PCI vendor/device IDs, bound kernel driver, and dedicated VRAM when exposed.
• Explicit shared-or-unknown memory semantics for integrated GPUs.
• Best-effort discovery of selected diagnostic/runtime executables on PATH.
• JSON serialization and an explicit CLI:
  • python -m evaluation_ai.hardware discover
  • python -m evaluation_ai.hardware discover --json
• Injectable proc/sys/etc roots for deterministic tests.
• Documentation of schema semantics and the distinction between detection and backend qualification.

Design boundaries

• No new runtime dependencies.
• No hardware inspection at import time.
• No driver installation, service launch, model download, or runtime configuration mutation.
• GPU detection is not reported as runtime support or execution qualification.
• Missing facts remain unknown rather than inferred.

Validation

Added tests for CPU/memory/distribution discovery, multiple GPUs, shared-memory reporting, and JSON serialization. The GitHub Actions matrix will provide the execution results; I have not run the suite in a local checkout.

Closes the Phase 1 implementation portion of #5; later qualification and runtime configuration phases remain open.

[sonarqubecloud]

Quality Gate passed

Issues
3 New issues
0 Accepted issues

Measures
0 Security Hotspots
0.0% Coverage on New Code
0.0% Duplication on New Code

See analysis details on SonarQube Cloud