Drop the timestamp suffix from the run name (the run_id already makes each
execution unique) and log the start time as a `run_started` param instead, so
run titles in the UI stay legible.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
mlflow_utils: resolve tracking store + git provenance from CWD
Now that mlflow_utils is shared by kdv-pinn and the kdv-foundation companion that
imports it, resolve the mlflow.db and git provenance from the working directory
(the project being run) rather than this module's package location. Each repo logs
to its own store; worktrees still share their main repo's db via the common git dir.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Promote the flat module set into an installable package (src/kdv_pinn/ +
pyproject.toml), add device.py and mlflow_utils.py (sqlite tracking, git
tagging, model registry), move run_training/parameter_sweep into scripts/, and
add soliton config files + dev requirements.
The IST physics-foundation-model work that was prototyped under
src/kdv_pinn/foundation/ has been extracted into a standalone companion repo
(kdv-foundation) that imports this package; it is intentionally not part of
kdv-pinn core.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
