Preserve exact servo references and match initial action heads

.gitattributes

@@ -6,3 +6,27 @@ src/dvidia_training/piper_imitation_viewer.bundle.js linguist-generated=true -di
 src/dvidia_training/mujoco.wasm binary
 # Preserve exact NVIDIA metadata bytes, including upstream trailing blank lines.
 benchmarks/nvidia-material-appearance-v01-20261025/central-metadata/*.usda -text -whitespace
+# Source-byte contracts must retain the same reviewed bytes on every OS.
+src/dvidia_training/piper_task.py text eol=lf
+src/dvidia_training/piper_imitation.py text eol=lf
+src/dvidia_training/camera_action_model.py text eol=lf
+src/dvidia_training/camera_action_export.py text eol=lf
+src/dvidia_training/piper_action_capture.py text eol=lf
+src/dvidia_training/camera_action_capture.py text eol=lf
+src/dvidia_training/robot_profiles.lock.json text eol=lf
+src/dvidia_training/camera_action_cache.py text eol=lf
+src/dvidia_training/camera_action_reference_export.py text eol=lf
+src/dvidia_training/camera_action_study.py text eol=lf
+src/dvidia_training/piper_rgbd_tracker.py text eol=lf
+src/dvidia_training/action_demonstration.py text eol=lf
+src/dvidia_training/reference_task.py text eol=lf
+src/dvidia_training/reference_sim.py text eol=lf
+src/dvidia_training/piper_imitation_score.py text eol=lf
+src/dvidia_training/piper_motor_primitive.py text eol=lf
+src/dvidia_training/piper_feedback_env.py text eol=lf
+src/dvidia_training/piper_feedback_score.py text eol=lf
+src/dvidia_training/robot_profiles.py text eol=lf
+src/dvidia_training/episode_tape.py text eol=lf
+src/dvidia_training/embodiment_calibration.py text eol=lf
+benchmarks/camera-action-v02-preflight/*.py text eol=lf
+benchmarks/camera-action-v02-head-preflight/*.py text eol=lf

.github/workflows/ci.yml

@@ -80,6 +80,10 @@ jobs:
           dvidia-training-doctor --json
       - name: Verify pinned learning caches
         run: python -m unittest tests.test_camera_action_cache tests.test_camera_cache_cli tests.test_camera_action_export tests.test_camera_action_model -v
+      - name: Verify exact-reference and initial-head contracts without physics
+        run: |
+          python -B -m unittest tests.test_camera_action_reference_export -v
+          python -B -m unittest discover -s benchmarks/camera-action-v02-head-preflight -p test_preflight.py -v
       - name: Test footage, visual training and loopback studio
         run: python -m unittest tests.test_footage_data tests.test_footage_train tests.test_footage_pipeline tests.test_footage_benchmark tests.test_training_studio tests.test_doctor tests.test_coverage tests.test_episode_tape tests.test_reaction tests.test_reaction_metrics tests.test_reaction_convergence tests.test_reaction_benchmark tests.test_motor_memory tests.test_motor_metrics tests.test_motor_benchmark tests.test_observations tests.test_observation_models tests.test_motion tests.test_motion_benchmark tests.test_transition_model tests.test_transition_pilot tests.test_transition_cli tests.test_embodiment_calibration tests.test_embodiment_cli tests.test_reference_sim -v
 

benchmarks/camera-action-reference-export-v01/README.md

@@ -0,0 +1,54 @@
+# Exact reference export: full training metadata validation
+
+One bounded run validated the 96,000 rows from 64 teacher demonstrations and
+32 mixed-control corrections. It built exact float64 prior-reference and
+feasible-label artifacts without reading raw images, depth, teacher-query
+bodies, native traces, held-out recordings or trained models. No policy was
+fitted or executed. The original candidate remains unqualified.
+
+- [Exact result receipt](full-training-results.json)
+- [Portable reproduction helper](validate_full_training_references.py)
+- [Executed versus portable script provenance](script-provenance.json)
+- [Contract, metrics and remaining integration work](../../docs/camera-action-reference-export-v01.md)
+
+Reproduction requires the pinned source checkout at or containing `a9340b7`,
+the unchanged v0.1 study metadata, and both independently verified parent caches
+from the [full training cache validation](../camera-action-export-cache-v01/).
+The reference helper reuses their strict source/data pins; it cannot create
+trusted inputs from arbitrary metadata. Parent-cache creation performs the
+separate raw verification step. Preserve those source bytes after verification.
+
+Use an existing NumPy environment and a fresh output outside the study/cache.
+The measured run used Python 3.12.9 and NumPy 2.5.3 on arm64 macOS, one worker,
+with at least 30 GiB free space retained. Example using absolute paths:
+
+```sh
+cd /absolute/work-directory-outside-source
+PYTHONPATH=/absolute/source/src OPENBLAS_NUM_THREADS=1 OMP_NUM_THREADS=1 \
+VECLIB_MAXIMUM_THREADS=1 /absolute/environment/bin/python -B \
+  /absolute/source/benchmarks/camera-action-reference-export-v01/validate_full_training_references.py \
+  --source-root /absolute/source \
+  --study-root /absolute/sealed-study \
+  --parent-cache-root /absolute/parent-cache-root \
+  --output /absolute/fresh-reference-validation \
+  --python-env /absolute/environment \
+  --execution-cwd /absolute/work-directory-outside-source
+```
+
+The parent root must contain `demonstrations-cache` and `corrections-cache` with
+the exact pinned manifests. The script builds each derived artifact once and
+loads each three times. It refuses existing output and retains interrupted
+prefixes. It makes no network request. The portable variant changes path
+bindings and normalizes platform-specific RSS reporting only; its numerical
+and validation operations are unchanged. RSS is null where unavailable.
+The portable variant has not run the real benchmark again.
+
+The receipt is the exact original result, SHA-256
+`68908bb9111ca92c4d672c25aa58c55417b4197f60a50d89bc96cd4e46bd799e`.
+Original script SHA-256:
+`3eb816e1145a9e3ad24bda0468ef9312e9cae9f08325948ead8c7d41981b7e45`.
+The measured run verifies normal reference endpoints after twenty updates,
+not a new replay of each original native microstep. Mixed feasible labels are
+counterfactual normal endpoints toward teacher suggestions, not proof that
+those suggestions controlled the arm. These are data-preparation measurements;
+there is no inference, training-speed or placement-success comparison.