[犀牛鸟-D1]：冻结基础模型xLatent Mixture（贾宗翰）

[xizaoge-shuai]

Summary

• Add deterministic offline multi-layer DINOv3 feature caching.
• Train only LatentMixture-P3/P4/P5 and the Detect head without loading the teacher during training.
• Add cache validation, GPU/CPU residency, trainability auditing, student-only checkpoint checks, experiments, ablations, and reproducibility evidence.

The complete experimental report is available at docs/d1/D1_FINAL_EXPERIMENT_REPORT.md.

Test evidence

• D1 regression suite: 36 passed, 1 non-blocking warning.
• COCO128 warm training cost reduction: 78.75%.
• VisDrone500 cold cost reduction: 67.07%.

Ablations

• Stream vs CPU preload vs GPU preload.
• Auxiliary loss weights: 0, 0.01, 0.05, 0.1, and 0.2.
• DINOv3 ViT-S/16 vs ViT-L/16 with three training seeds.

Limitations

• Experiments use deterministic COCO128 and VisDrone500 subsets.
• GPU residency trades additional VRAM for training throughput.
• The multi-seed results do not establish statistically significant gains for aux=0.01 or ViT-L.

[isLinXu]

Thanks for the review. You're right that this PR carried too many files that don't belong in the merge. I audited all 27 changed files and will trim the branch so that only the D1 core lands.

Files to be removed (internal / process artifacts, not part of the feature)

These are local working artifacts from the admission-gate and experiment workflow. They contain machine-specific paths and one-off evidence that should not be merged into the repo:

• docs/d1/evidence/ — all 14 files (run JSONs, SHA256 pin lists, commit/diff snapshots; e.g. p0-coco128-s0.json contains a local run path /data/users/zhjia/...)
• docs/d1/824-admission-status.md — internal admission-gate status doc
• docs/d1/p0-status.md — internal P0 status doc
• docs/superpowers/plans/2026-08-25-d1-824-feature-cache.md — personal agent-workspace plan
• docs/superpowers/specs/2026-08-25-d1-offline-dinov3-cache-design.md — personal agent-workspace spec

Files that stay (the D1 feature core)

• ultralytics/data/foundation_cache.py — deterministic cache contracts (schema, identity, manifest validation)
• ultralytics/data/preloaded_feature_cache.py — preloaded batched loader (GPU/CPU residency)
• scripts/d1_build_feature_cache.py — offline DINOv3 multi-level cache builder
• scripts/d1_train_cached_detector.py — cached-feature LatentMixture + Detect trainer
• tests/test_d1_foundation_cache.py, tests/test_d1_feature_cache_builder.py, tests/test_d1_preloaded_feature_cache.py, tests/test_d1_cached_detector.py
• docs/d1/D1_FINAL_EXPERIMENT_REPORT.md — consolidated experiment report (single doc replaces the scattered JSON evidence)

Plan

I'll rebase the branch to drop the files above and force-push. The experiment numbers quoted in the report stay unchanged. After the trimmed branch is up, please take another look.

[xizaoge-shuai]

Thanks for the review. I rebuilt the branch on the current upstream main and trimmed the PR to the D1 feature core only. Internal admission/evidence artifacts and personal workspace documents have been removed; the consolidated experiment report remains. The focused D1 tests pass. Please take another look when convenient.

[xizaoge-shuai]

您好，我想问一下我的提交是否齐全，有无缺漏