Add FlowUNet1D generative bridge flow-matching denoiser

[will-rice]

Summary

Adds the repo's first generative denoiser: FlowUNet1D, a waveform-domain Schrödinger-bridge-style flow-matching model with a data-prediction objective. The full research survey and design rationale (diffusion vs flow/bridge vs GAN vs latent/codec, with citations) is committed as docs/generative_denoiser_proposal.md.

Key idea: training corrupts clean audio along a Brownian bridge between the clean and noisy waveforms (x_t = (1−t)·x₀ + t·y + σ_max√(t(1−t))·ε) and the network regresses the clean endpoint using the existing L1 + multi-resolution STFT losses. Sampling starts at the noisy waveform and re-projects onto the bridge for num_inference_steps iterations — at 1 step it reduces exactly to a single predictive forward pass (verified by test), and more steps buy generative refinement.

Changes

• src/denoisers/modeling/flowunet1d/ — FlowUNet1DConfig (model_type="flowunet1d"), sinusoidal TimeEmbedding, zero-initialized FiLM1D conditioning, FlowUNet1D backbone reusing the existing unet1d Down/Mid/Up blocks with a 2-channel input (bridge state ‖ noisy), and FlowUNet1DModel with predict_clean() (training) and a deterministic-by-default bridge sampler in forward() (optional stochastic sampling).
• FlowMatchingLightningModule — subclass of DenoisersLightningModule inheriting metrics, DNSMOS, EMA, and W&B logging; overrides train/val steps for the bridge objective. Parent now reads autoencoder via getattr(..., True) so configs without the field work.
• Wiring — registered in MODELS/CONFIGS, exported from the package root, and train.py selects the flow module for the flowunet1d key (train flowunet1d <name> <data_root>).
• Tests — config round-trip, sampler output shape, bridge-sigma endpoint behavior, single-step ≡ predictive-call equivalence, and Lightning train/val steps.
• Docs — cited survey + design proposal, including 48 kHz risk analysis and fallbacks (StoRM-style regeneration seeded by the existing pretrained unet1d-vctk-48khz, complex-STFT front-end).

Test plan

• pytest — new model + lightning tests pass (29 passed locally; the 3 test_pretrained HF-download tests require network and pass in CI)
• ruff check . clean
• mypy . clean (40 files)

🤖 Generated with Claude Code

https://claude.ai/code/session_01CnieBtBKfRzKAYpc7N7hWP

---

Generated by Claude Code

[cursor]

Bugbot is not enabled for your account, so this pull request was not reviewed.

Enable Bugbot in the Cursor dashboard to get automatic reviews on future PRs.