cathar v0.5.2

Changed

• De-clip now reconstructs clipped peaks with least-squares autoregressive
  interpolation (LSAR) — the classical audio-restoration method (Janssen,
  Veldhuis & Vries, 1986) — instead of a cubic fill. An AR model is fit to the
  reliable audio either side of each clipped run (two-sided autocorrelation →
  Levinson-Durbin) and the gap samples that minimise its prediction error are
  solved for (banded normal equations via Cholesky), so a clipped peak is rebuilt
  toward its true amplitude rather than flattened to the shoulder level. A
  stability guard falls back to the previous smooth fill when a solve rings or
  overshoots, so badly-clipped material softens gracefully. On a +8 dB-clipped
  voice the old fill could only reach the 0.977 plateau; LSAR rebuilds the true
  peaks to ~1.53 (normalise afterwards).

Fixed

• Mono WAV output played in the left speaker only. hound writes 32-bit float
  WAV as WAVE_FORMAT_EXTENSIBLE and tags a single channel as FRONT_LEFT, so
  layout-aware players (CoreAudio / afplay) routed it hard-left. Mono output is
  now tagged FRONT_CENTER and plays centred. Stereo and FLAC/AIFF were
  unaffected.

Documentation

• Added "Cleaning Up Sound" (book/) — a from-first-principles book on the
  concepts this toolkit uses, for readers new to DAWs/DSP, with diagrams, a cover
  page, and a GitHub Pages build.
• Documented every remaining public item (Error + variants, AudioData +
  fields, the Denoiser trait, NoisePrint/SpectralDenoiser fields,
  with_noise_print, generate_wave, variance) and added
  #![deny(missing_docs)] to the cathar crate so public docs can't regress.

Internal

• Split the ~2,400-line cathar/src/lib.rs into focused modules (audio,
  denoise, restore, enhance, loudness, resample, error, util),
  re-exported flat so the public API is unchanged. No behaviour change.