refactor(checkpoints): HF-only writes; configurable W&B read alias

[mkeeler43]

Summary

Stacks on #63 to deliver the policy we agreed on: HuggingFace is the only place model checkpoints are stored. W&B continues to handle scalars, media, and run history.

What the new UX looks like

One-time setup:

wandb login
huggingface-cli login   # or: export HF_TOKEN=...

Training:

python engiopt/cgan_cnn_2d/cgan_cnn_2d.py --problem-id beams2d --seed 1 --track --save-model

• --track logs scalars + media to W&B (default project: engiopt)
• --save-model uploads the checkpoint to HF (default repo: IDEALLab/engiopt-cgan-cnn-2d)
• Override --hf-entity / --hf-repo-prefix for personal experiments — same shape as --wandb-entity / --wandb-project

Evaluation / inference:

python engiopt/cgan_cnn_2d/evaluate_cgan_cnn_2d.py --problem-id beams2d --seed 1

• Pulls the checkpoint from HF automatically
• Falls back silently to the legacy W&B artifact for runs trained before this cutover

Flags removed (cleanup)

These flags from #63 are no longer surfaced to users; behavior is folded into sensible defaults:

Removed flag	New behavior
--checkpoint-backend	Always HF. --save-model is the on/off.
--hf-private	Always public. Create a private repo manually if needed.
--model-source	Always auto: HF first, W&B fallback.
--local-model-dir	Power-user knob; can be re-added later.

The W&B read fallback inside checkpoint_store.py stays as a safety net while existing artifacts are still on W&B. A later cleanup PR will remove it once everything is migrated.

What's verified

• HF train → eval round-trip on beams2d / cgan_cnn_2d (1 epoch, seed 999):
  • Checkpoint uploaded to IDEALLab/engiopt-cgan-cnn-2d/beams2d/seed_999/
  • Immutable run-scoped path …/seed_999/run_<wandb_run_id>/ also created
  • W&B run summary recorded the HF repo, package path, and revision
  • Eval auto-pulled the checkpoint, ran inference, wrote the metrics CSV
  • Test artifacts deleted afterward
• CI: ruff, pre-commit, mypy all green

Forward-compat for the upcoming LVAE work

The future LVAE-minimal branch saves with custom subpaths (rec{r}/perf{p}/{filter}/seed_{s}) and reads legacy W&B artifacts under the LVAE alias scheme (seed_{s}_rec{r}_perf{p}…). Both hooks — extra_path_parts and wandb_artifact_alias — are already wired into save_checkpoint_package and resolve_named_checkpoint, so the LVAE migration is per-script editing only. No further changes to checkpoint_store.py are needed.

Commits

1. HF-only writes, configurable W&B read alias — drop the "wandb"/"both" write modes; add wandb_artifact_alias so legacy reads work for non-standard alias schemes.
2. Collapse CLI surface to W&B parallel — drop the 4 flags listed above; bake their old defaults into the call sites.
3. mypy fix — drop two now-invalid model_source= kwargs on load_model_from_reference call sites that the earlier flag cleanup left behind.