v0.3.0

Added

• nam feature and module (on by default): load and run Neural Amp Modeler .nam models (a
  pure-Rust port of NeuralAmpModelerCore's WaveNet, plain and SlimmableContainer; matches it to
  ≤ 3e-5 on its example models, which are now test vectors), and nam::convert, the official
  app's NAM → pedal converter: a fundamental/3rd-harmonic tone test and a small-signal linear
  response give two minimum-phase FIRs. With the app's own excitation the result matches its
  upload to ≤ 0.03 dB; with ours (the default) within ~1–1.5 dB rms, the app's own measurement
  noise. Device::import_nam_file(slot, path) converts and uploads. Examples nam_convert,
  nam_render, nam_check. Python: open_stomp.nam, Device.import_nam_file. Confirmed on
  hardware: an official and an open conversion of the same bass model sound identical.
• nam::convert::Measure::Clean (convert_measured; Python nam.convert(..., measure="clean")),
  experimental: the same converter with a steady-state small-signal measurement instead of the
  app's transient-affected one. Re-amp tests on hardware against the real model (Ampeg SVT bass
  model): closer mid scoop, top-end roll-off and upper-note distortion than the official
  algorithm, but still short on the 63–90 Hz resonance and about 5 dB quieter. Not the default.
• nam::convert::PEDAL_TAPS, upload_data, pedal_irs: the pedal runs only the first 512 taps
  of each uploaded filter (found by re-amping; docs/05). Measure::Clean now uploads exactly
  those taps, with the ignored halves zeroed; Measure::App keeps the app's layout.
• nam::optimise and Measure::Optimised, the default for Device::import_nam_file (about a
  minute per model): fit G1/G2 so a simulation of the pedal (y = G2 ⊛ f(G1 ⊛ x), its saturation
  curve fitted from re-amp recordings) matches the model on sweeps, tones and pink noise, starting
  from the clean conversion. Hand-written gradients, Adam, no ML framework. On hardware (Ampeg SVT
  bass model) it scores 6.72–6.85 dB rms error against the real model, against 9.32 for the
  official conversion in the same session. Device::import_nam_file_with(.., Measure::App) keeps
  the official algorithm (a few seconds). Progress and cancel: convert::convert_optimised,
  optimise::optimise_with_progress. Also optimise::simulate, example nam_optimise.
  Python: Device.import_nam_file(..., measure="optimised") (default) and nam.convert(..., measure="optimised").
• nam::convert::official_excitation / find_official_excitation: read the official app's test
  signal from the user's own Stomp Manager install (located by fingerprint; no vendor data in the
  source). With it the conversion equals the official one (error energy −51 to −93 dB on ten
  models). Strictly opt-in: Device::import_nam_file and everything else never read the app.
  Python: nam.official_excitation().
• upload module: user AMP (NAM, SysEx 0x58) and user IR (0x10 + 0x09) upload framing:
  chunking, the 128-byte slot-info record (UserAmpInfo: name, factor), delete, and a
  Reassembler that turns captured chunks back into floats. Every captured import rebuilds
  byte for byte. Device::import_user_amp / delete_user_amp / import_user_ir send the
  official app's sequences (2048 finite floats, slots 0–9, else UploadError). Also in the
  Python bindings (open_stomp.upload, Device.import_user_amp …).
• DeviceState follows user AMP imports and deletes (slot names, log events with progress).

Changed

• docs/05: corrected chunk header (plain byte offset + total length, not 8-byte units), the
  slot-info record layout, and the app selecting the slot's model before uploading.
• docs/05, docs/09: the NAM converter algorithm, recovered from the vendor code and verified stage
  by stage: a small-signal best linear approximation and a sine tone test (fundamental and 3rd
  harmonic) through the network give two minimum-phase FIRs (Wiener–Hammerstein G1/G2) whose first
  1024 taps, reversed, are the upload; factor is always 1.0.