v0.6.0

Added

• CLI training scripts for all three model types, each with optional W&B integration:
  • desisky-train-broadband — Train broadband MLP on moon-contaminated data with per-band scatter+residual visualization callbacks
  • desisky-train-vae — Train VAE (InfoVAE-MMD) with reconstruction, latent corner, broadband CDF, and airglow CDF visualization callbacks
  • desisky-train-ldm — Train LDM (EDM framework) for dark/moon/twilight variants with broadband CDF, airglow CDF, latent corner comparison, and conditional validation grid callbacks
• CLI inference scripts for all three model types:
  • desisky-infer-broadband — Run broadband MLP inference, output CSV or npz
  • desisky-infer-vae — VAE encode+reconstruct with subset selection (full/dark/moon/twilight)
  • desisky-infer-ldm — Generate spectra with LDM; conditioning from real data, user file, or inline JSON
• Multi-format data loading for LDM training (--data-path):
  • .npz — Pre-processed data (flux, conditioning, wavelength keys)
  • .fits — SkySpecVAC format with automatic variant-specific quality filtering
  • .csv + --flux-path — Metadata CSV with separate .npy flux array
• plot_broadband_band_panel() — New 2x2 scatter+residual diagnostic panel for broadband training visualization
• BroadbandTrainer now supports wandb_config and on_epoch_end callback parameters (backward-compatible, both default to None)
• 6 new console entry points registered in pyproject.toml
• docs/CLI_GUIDE.md — Documents all CLI commands, data formats, output formats, and wandb integration
• Integration tests for all 6 CLI scripts (3 training + 3 inference)

Changed

• train_vae.py — Full rewrite to use VAETrainer with wandb callbacks; uses raw tensors instead of TensorDataset for correct VAETrainer iteration
• BroadbandTrainer restructured with train() / _train_loop() pattern matching VAETrainer / LatentDiffusionTrainer

Fixed

• LDM training encoding now uses VAE full forward pass (vae(batch, key)["latent"]) with reparameterization, matching the notebook pattern, instead of vae.encode() which returns a raw tuple
• SkySpecVAC method name mapping: load_dark → load_dark_time, load_moon → load_moon_contaminated, load_twilight → load_sun_contaminated
• LDM training and inference now apply manual metadata enrichment (attach_solar_flux, add_galactic_coordinates, add_ecliptic_coordinates) since enrich=True only adds SKY_MAG_V_SPEC and ECLIPSE_FRAC
• NaN/Inf rows in conditioning columns are now detected and removed during data loading
• plot_conditional_validation_grid() no longer crashes with small validation sets (e.g., 29 twilight samples): adaptive bin count and proper empty-bin handling
• plot_cdf_comparison() histograms now use shared bin edges between real and generated data for accurate visual comparison
• plot_broadband_band_panel() histogram bins reduced from 80 to 30; combined +/- sigma into single legend entry; relabeled zero-line as "0 (perfect)"
• Suppressed harmless speclite log10 warnings from negative flux values in early-training generated spectra