Immutable data manifests

Tested parser and normalized sample inventory.

Implemented a hash-before-deserialization NTU HRNet importer and canonical S/C/P/R/A parser. Normalized 113,945 unique usable samples with frame count, tensor shapes, pose-track count, and nonempty-track count; malformed identifiers, duplicate aliases, label disagreement, invalid body metadata, missing-listed overlap, and non-finite arrays are rejected.

Seven deterministic, checksummed manifest files.

Generated seven immutable JSONL manifests from the locked class and exemplar policy: 76,013 development-train, 18,988 development-validation, 95,001 final-train, 18,944 official-novel, 20 anchor, 18,884 primary-query, and 18,924 exact-official-query records. The audit records every SHA-256 digest and generation input.

Passing leakage and duplication test suite.

Automated checks prove class disjointness, one anchor per novel class, source and manifest identity uniqueness, exact exclusion of 40 synchronized camera mates, stable ordering, and complete 113,945-sample source coverage. An independent regeneration was byte-identical to the published release bundle.

Human-readable manifest audit report.

Published a source-safe manifest audit with per-class counts, all manifest digests, body-metadata coverage, invariant results, and one explicit explanation for each of the 40 synchronized-view exclusions. The required result-blind amendment is explicitly scoped to both the 114,480-versus-113,945 count and aggregate-aware source verification.