- Correct NeMo native word-CTM manifest handling
- Close real GPU smoke, five-row, and full-36 evidence
- Record the final three-aligner selection and final 36-row annotation
- Preserve historical Stage 5 evidence and bounded architecture
- Preserve the configured NeMo virtualenv interpreter path instead of resolving through uv symlinks
- Accept pinned NeMo 3.0.0 source builds with PEP 440 local-version suffixes
- Provide a workspace-local Numba cache for isolated NFA execution
- Add symlink, runtime-version, cache, and real-integration regression coverage
chore(repo): finalize licensing and documentation
- Add the standard Apache License 2.0 and VertoX-AI copyright headers
- Preserve deliberately empty package markers and the standalone worker shebang policy
- Rewrite the README as a concise Hugging Face repository landing page
- Add project status, modular architecture, and detailed workflow documentation
- Add ASCII diagrams for application, benchmark, and backend execution flows
- Normalize public branding to VertoX-AI and add repository-specific citation metadata
- Add concise upstream acknowledgments and official project links
- Add deterministic license-header and empty-package-marker validation
- Preserve all Stage 0-5 reports, benchmark evidence, runtime pins, and implementation behavior
- Add isolated WhisperX reference-text alignment with typed configuration and provider boundaries
- Align authoritative transcripts directly without using Whisper ASR transcription as ground truth
- Parse native word timings strictly with explicit missing-timestamp and alignment failure handling
- Pin the WhisperX runtime and English WAV2VEC2 alignment model with reproducible provenance
- Preserve heavyweight model imports inside the isolated worker and keep neutral imports optional
- Preserve programming and systemic worker failures at the whole-process boundary
- Add deterministic unit, architecture, coverage, and opt-in real integration tests
- Record the real five-row evaluation with 5/5 successes and 0.0154 s first-word-end MAE
- Add typed MFA configuration, corpus materialization, subprocess execution, and TextGrid parsing
- Preserve authoritative transcripts with reversible MFA-specific normalization and lexical mapping
- Add strict handling for OOVs, silence intervals, malformed output, and timestamp validation
- Keep MFA isolated in its dedicated Conda environment and behind the shared backend protocol
- Add deterministic unit, architecture, coverage, and opt-in real integration tests
- Record the real five-row MFA evaluation with 5/5 successes and 0.0162 s first-word-end MAE
- Document runtime, model, dictionary, G2P, provenance, and Stage 4 readiness
- Add typed NeMo configuration, reference-text manifest generation, and isolated NFA execution
- Parse native word CTMs into backend-neutral alignments with strict transcript reconciliation
- Enforce authoritative reference-text mode and pinned NeMo script/runtime provenance
- Preserve optional native alignment artifacts and expose them through benchmark reports
- Add lazy library and CLI composition without importing NeMo or Torch in the parent process
- Add deterministic unit, architecture, coverage, and opt-in real integration tests
- Document reproducible NeMo setup and explicitly record the blocked five-row empirical run
