[Naeemkh]

Summary

Rewrites KempnerPulse from a single script into a clean, standalone
four-layer package, with a test suite, a documentation site, and CI.

Architecture

• Read (reader/): pluggable backends (dcgmi dmon, dcgm-exporter, replay) that emit raw per-GPU records in the source's own vocabulary.
• Translate (translate/): maps raw records to a versioned, vendor-neutral canonical schema with normalized units.
• Compute (compute/): pure-functional Real Utilization, the 12-category workload classification, and health states.
• Present (present/): the Rich terminal UI (fleet / focus / plot / job views) and the CSV exporter.
• Cross-cutting: configuration, hardware identification, GPU selection, system queries, and the lifecycle / entry point.
• Removes the legacy single-file implementation and the temporary compatibility shim so the package stands on its own.

Packaging & CLI

• Moves the project to a src/ layout with an updated pyproject.toml.
• Adds a short kp alias alongside the kempnerpulse command.
• Adds a --backend replay mode that replays a saved capture with no GPU required.
• Moves the kempnerpulse-on-cnode.sh helper into scripts/.

Tests & CI

• Adds a 156-test unit + integration suite covering every layer (GPU-free, via the replay backend).
• Adds a GitHub Actions workflow that runs the suite across Python 3.9–3.12.
• Validated live on H100 hardware (CSV export, dashboard render, and the interactive TUI).

Documentation

• Adds a Sphinx + Furo documentation site (install, usage, architecture, and an auto-generated API reference) with a GitHub Pages deploy workflow.
• Trims the README to a concise front page that links into the docs.