smoothquant: fix istupakov int8 size note comparing output to itself

smoothquant: honor --alpha (work around neural-compressor dropping SmoothQuant* extra_options)

quantize-int8-smoothquant: use one fixed append-mode log file

quantize-int8-smoothquant: route output through loguru + log to file

scripts: move quantize/shard tools into scripts/ subfolder

smoothquant: point wer-quants.py suggestion at the parakeet_web repo

smoothquant: strip orphaned folded smooth-scale initializers

quantize-fp16: make it a uv run script (PEP 723 header)

docs: reflect standalone script defaults (./ and ./calibration_audio)

scripts: default quantize-fp16 and shard-fp32 to the current directory

smoothquant: normalize usage examples to standalone invocation

smoothquant: default to . for models and ./candidates, ./calibration_audio for calib

smoothquant: default alpha to per-layer auto search

docs: fill in generalization table with comprehensive benchmark numbers

Add browser-friendly sharded fp32 encoder + shard-fp32.py

docs: expand fp16 intro (naive cast, equal to fp32, half-size packaging win)

docs: merge fp16/fp32 WER column (identical values) with italic note

docs: correct calibration framing (bilingual audio used) + TODO recomputed generalization values

docs: note murmure contribution / upstreaming in README

Add .gitignore for calibration audio and local logs

Recalibrate int8 encoder on held-out speeches (fix eval contamination)

docs: language-agnostic note + FLEURS-fr check + parakeet_web origin

docs: frame the long-audio result as not-reproduced, not a guarantee

Parakeet TDT 0.6B v3 (Multilingual) ONNX with a SmoothQuant int8 encoder