Add dataset split methodology: SplitMetadata infrastructure, benchmark split enforcement, and CLI split commands

[Copilot]

The benchmark workflow had no train/val/test separation, making it trivial to overfit profiles to benchmark images and cherry-pick results. The Anti-DreamBooth set_C split was documented as "holdout metadata" rather than being actively used for held-out evaluation.

Core split infrastructure (src/auralock/core/splits.py)

• SplitType enum (TRAIN, VALIDATION, TEST, DEVELOPMENT)
• SplitMetadata dataclass — tracks image_ids, split_method, deterministic split_hash, split_ratio, random_seed; includes verify_no_leakage() and full JSON serialization
• compute_split_hash() — short SHA-256 fingerprint over sorted image IDs + seed, enabling tamper detection
• create_random_split() — seeded, leakage-free train/val/test partitioning with full metadata
• save_split_manifest() / load_split_manifest() — JSON persistence for reproducible split assignments
• validate_split_manifest() — checks cross-split overlap and hash integrity, returns list of issues
• warn_non_test_split() — emits UserWarning when benchmarking on non-TEST splits

Placed in core/ (not benchmarks/) to avoid the circular import services → benchmarks → lora → services. benchmarks/splits.py is a re-export shim.

Benchmark enforcement (src/auralock/services/protection.py)

• BenchmarkSummary gains split_metadata: SplitMetadata | None = None; serialized in to_report_dict()
• benchmark_file() — when split_metadata provided: raises ValueError if image not in split, warns on non-TEST
• benchmark_directory() — filters images to declared split members, warns on excluded images and non-TEST splits

splits = create_random_split(image_paths, random_seed=42)
save_split_manifest(splits, Path("splits.json"))

# Later: benchmark only the held-out test images
test_split = splits[SplitType.TEST]
summary = service.benchmark_directory(dataset_dir, profiles=("balanced",), split_metadata=test_split)
# UserWarning if split_type != TEST; ValueError if images outside split

CLI split commands (src/auralock/cli.py)

New auralock split sub-command group:

• auralock split create <dir> --output splits.json [--train-ratio 0.7 --val-ratio 0.15 --test-ratio 0.15 --seed 42]
• auralock split validate splits.json — exits 1 with details on leakage or hash mismatch

auralock benchmark gains --split-manifest and --split-type options; prints a non-test-split bias warning to stdout before running.

Anti-DreamBooth set_C note

Corrected the misleading "set_C is preserved as holdout metadata" note to explicitly document set_C as the held-out validation split for measuring out-of-sample protection effectiveness.