Release acceptance run

scripts/release/acceptance_run.sh proves an installed stack works end to end: health, self-hosted API docs, a throwaway project, ingest of public COCO images, clustering, VLM labeling, human confirmation, test holdout, YOLO export, a short training run, promote, inference with the promoted model, delete, metrics and log-noise checks, and the installer lifecycle hooks. It needs only bash, curl and jq.

Run it against an isolated install (own compose project and install directory), never a stack you care about, because it creates, trains and deletes a project:

python3 scripts/datasets/fetch_coco_subset.py --out data/samples/coco_va
scripts/release/acceptance_run.sh \
  --base-url http://localhost:4703 --project-name op-acc \
  --install-dir ~/op-acc --models-dir ~/op-acc/models \
  --project acc-run --images-dir data/samples/coco_va --report acceptance.json

Useful flags: --only a,b and --skip a,b (see --list for phase names), --train-gpu, --vlm-switch-id <id>, --upgrade-version vX.Y.Z, --strict. The JSON report records, per phase, pass/fail/skip, seconds and evidence; the exit status is non-zero when a required phase fails. A cleanup trap always removes the throwaway project and any model the run promoted.