feat: finish #28 (safetensors, config.json, latent_length) + implement #38 (per-sample conditioning)

[Mec-iS]

Summary

Finishes the remaining #28 items and implements #38 (per-sample conditioning). Single branch, separate commits per issue.

Issue #28 remaining items

#3 — safetensors export

Added safetensors dependency. Both training scripts now export .safetensors alongside .pt files after training:

• train_audio_decoder.py: decoder.safetensors
• train_audio_diffusion.py: dit.safetensors
  Prior excluded (full pickled object, not a state_dict).

#4 — derive latent_length from audio_length

--latent-length now defaults to None (sentinel). When not explicitly set, derived as audio_length // 320 (EnCodec stride). Applied to train_audio_diffusion.py and sample_audio.py. Backward compatible: explicit --latent-length overrides derivation.

#5 — config.json alongside checkpoints

Both training scripts now write config.json with full architecture hyperparams:

• train_audio_decoder.py: decoder_type, latent_channels, base_channels, feature_dim, q, k, upsample_strides, audio_length, sample_rate
• train_audio_diffusion.py: dim, depth, num_heads, patch_size, latent_channels, latent_length, spec_dim, q, k, cfg_dropout, num_steps, audio_length

The Rust generation worker can now reconstruct exact architecture without hard-coding.

Issue #38 — per-sample conditioning

sample_audio.py now supports:

• --num-samples (default 1): multi-sample generation
• --per-sample-conditioning (flag): resample c_spec per sample for varied banks

When active, each sample gets a fresh c_spec probe with a deterministic per-sample seed (base_seed + idx + 1000), steering toward different spectral regions. When off (default), shared c_spec produces a coherent bank.

Multi-sample outputs: sample_00.wav, sample_01.wav, etc. Single sample keeps original filename.

Also added --audio-length to sample_audio.py for latent_length derivation.

Tests (TDD, 13 new)

• test_config_export_28.py (8 tests): config.json export from both scripts, safetensors export + loadability, latent_length derivation
• test_per_sample_conditioning_38.py (5 tests): CLI flags, per-sample diversity > shared (guidance=50, perturbed AdaLN), shared condenses vs unconditional, per-sample deterministic

Full suite: 238 passed, ruff clean, pre-commit hook passed.

Version

0.7.0 to 0.8.0

Closes #28, closes #38.