Research 0713: CHUG HDR CLI Bootstrap Sweep¶

Date: 2026-05-22

Scope¶

Apply the accepted AI script helper pattern to the CHUG HDR entrypoints:

ai/scripts/chug_extract_features.py
ai/scripts/train_chug_hdr_mos_head.py

The goal is to keep the HDR path aligned with the rest of the AI tooling while the long K150K feature refresh continues separately.

Findings¶

chug_extract_features.py still derived SCRIPT_PATH, REPO_ROOT, and ai/src manually, then captured sys.argv[1:] directly before building the run-provenance sidecar.

train_chug_hdr_mos_head.py manually inserted the AI scripts directory so it could delegate to train_konvid_mos_head.py. It also parsed forwarded arguments from the original argv object while separately constructing a raw argv list for the delegated --run-argv-json payload.

Change¶

Both scripts now use:

bootstrap_ai_script(...) for path setup
collect_cli_argv(...) for the exact CLI vector recorded in manifests
make_argument_parser(...) for consistent AI script parser formatting

The CHUG MOS wrapper uses include_ai_scripts=True and forwards the same collected argv that it records in the delegated manifest payload.

Alternatives Considered¶

Option	Result	Rationale
Leave CHUG alone until the current K150K run completes	Rejected	These scripts are independent from the running K150K extractor and are HDR-priority backlog.
Inline CHUG training instead of delegating to the KonViD MOS trainer	Rejected	That would be a behavior change and a larger training-policy decision; this PR only removes boilerplate drift.
Add a CHUG-specific parser helper	Rejected	The generic AI helper covers the needed path setup and provenance capture.

Validation¶

Focused tests:

.venv/bin/python -m pytest ai/tests/test_chug.py ai/tests/test_train_konvid_mos_head.py -q

Static checks:

.venv/bin/ruff check ai/scripts/chug_extract_features.py ai/scripts/train_chug_hdr_mos_head.py
.venv/bin/black --check ai/scripts/chug_extract_features.py ai/scripts/train_chug_hdr_mos_head.py

The end-to-end ai/tests/test_chug_extract_features_smoke.py run was attempted locally, but the host build/tools/vmaf process stopped making progress after writing the temporary JSON output. That smoke covers the libvmaf binary path rather than this bootstrap-only diff, so it is left out of this PR's required validation.

Rebase Notes¶

No rebase-sensitive invariant changes. The wrapper and extractor retain their existing output schemas and only adopt the shared AI helper pattern.