Remove exploratory/branched implementations from main (VQ-VAE, original ResNets, EMA, pre-transformer UNet)

[jlotthammer]

Summary

Removes exploratory / branched implementations from main that were not part of the deployed STARLING system, leaving only the production model code (VAE distance-map autoencoder + ViT/DiT latent diffusion) and the training code used to produce it.

Full pre-cleanup state is preserved on the archive/pre-cleanup branch (tag archive/pre-cleanup-2026-07-16).

Net change: −1937 lines, no production code paths altered.

What was removed

Orphaned model modules (zero importers in production or training):

• models/quantize.py — VQ-VAE VectorQuantizer2
• models/resnets_original.py — pre-vae_components ResNet enc/dec
• models/ema.py — EMA, never wired into training or diffusion

Pre-transformer UNet backbone (superseded by the ViT/DiT transformer):

• models/unet.py, configs/unet/unet.yaml (+ its configs.yaml defaults entry)
• Its only remaining tie to production was vit.py importing SinusoidalPosEmb; that class is byte-identical to the one already in transformer.py, so the import was redirected there.
• Dropped the unused UNetConditional construction in diffusion_train.py (the real backbone passed to the diffusion model was already the ViT).

Broken references / bit-rot:

• pyproject.toml: removed starling-sample and ae-train console scripts (targeted non-existent modules training/vae_generate.py, training/ae_train.py).
• diffusion_train.py: removed import starling.data.ddpm_loader (module does not exist — only ddpm_loader_tar is present; this broke starling-ddpm-train at import time).

Docs: removed the Sphinx autosummary stubs and api.rst entries for all deleted modules.

What was deliberately kept

• The core VAE (vae.py, vae_components.py, distributions.py) — this is the production distance-map autoencoder, loaded at inference. Only the VQ-VAE variant was exploratory.
• evaluate_vae.py + benchmark_mds.py — wired to the live starling-benchmark CLI.
• devtools/scripts/ — left untouched.

Verification

• All tracked .py byte-compile; import starling + all affected modules import at runtime; ViT(12,512,8,512) constructs.
• All 14 remaining pyproject console scripts resolve to real module:function targets.
• No dangling references to any removed module remain.
• Fast test suite: 138 passed, 2 skipped.

Out of scope (pre-existing, not caused by this PR)

8 stochastic generation/reconstruction tests in test_starling.py fail because ~19/100 generated conformations contain NaN on this environment (torch 2.13.0). Confirmed identical failures on a clean main worktree, so this is unrelated to the cleanup and tracked separately.