Fix train.py CLI defaults, optional Piper, and validation batch size

[zain1806481]

Summary

• Replace default="False" on store_true CLI flags with default=False (the string is truthy).
• Only import Piper sample generation when piper_sample_generator_path is set.
• Cap false-positive validation DataLoader batch size at 256 to avoid loading the entire validation set onto small GPUs at once.

Context

Hit while training on a 2 GB Pascal GPU. Validation previously used batch_size=len(X_val_fp_labels), which could request multiple GB of feature windows in one step.

Test plan

• python -m openwakeword.train --help shows flags defaulting off
• Training config without piper_sample_generator_path gets past import when not generating clips
• Validation step on a low-VRAM GPU completes without allocating the full validation tensor set in one batch

Made with Cursor

[zain1806481]

Week submission index: https://github.com/zain1806481/upstream-week-2026-09-14