Motivation

• Make the npm launcher robust in real-world installs by reliably locating Python, capturing child output, and only bootstrapping deps when truly missing.
• Enable large-scale, interruption-safe audio corpus generation with parallelism, sharding, and resume/checkpoint semantics to support very large --variants-per-input runs.

Description

• Hardened npm/verbx.js to search Python executables ($PYTHON, python3, python, and py on Windows), capture stdout/stderr for accurate ModuleNotFoundError detection, and print clearer bootstrap diagnostics when invoking -m pip install --user.
• Extended verbx batch corpus-generate in src/verbx/cli.py with --jobs, --checkpoint-file, --resume, --num-shards, and --shard-index, streaming JSONL manifest writes, shard-aware planning, and per-run resume/failure accounting.
• Refactored variant rendering into _generate_corpus_variant with deterministic per-variant RNG seeding so parallel generation is reproducible, and used ThreadPoolExecutor to parallelize per-input variant generation.
• Updated documentation (README.md, docs/AI_AUGMENTATION.md) to document launcher fallback/bootstrap behavior and the new corpus-generation flags, and added unit tests covering shard dry-run output, checkpoint/resume, and npm-launcher behavior.

Testing

• Ran static checks with python -m ruff check src/verbx/cli.py tests/test_cli.py tests/test_npm_launcher.py which passed.
• Executed python -m pytest tests/test_cli.py -k "corpus_generate" -q which produced 4 passed, 104 deselected.
• Executed python -m pytest tests/test_npm_launcher.py -q which produced 2 passed.

Codex Task