[HyTruongSon]

The original three primary checkpoints remain missing. This PR recovers the available journal artifacts, retrains only the required primary ensemble under fresh identities, and adds reproducibility/release tooling while preserving every manuscript and historical result file.

• Recovered 132 unique best states (150 study references), plus 48 resume files. Trained three primary replacements once with original source/validation splits, scalers, architecture and stopping rules. All predeclared external ensemble agreement criteria pass; originals remain explicitly missing and the original doctor rejects replacements.
• Exact molecular/geometry replay: 22,500 runs and 3.375 million primary per-gene values pass rtol=atol=1e-10. Actual recovered-weight GCN/local inference on all 22 Bassiouni sections reproduces all 792 residual-route reference rows. Reports retain seed variation, rounded targets, initial runtime sensitivity and partial verification of unselected trials.
• Published research-artifacts-v1: 99 public best heads (3 replacements +96 recovered ResNet18-based), 48 additional resume files, configs/scalers, numerical exports, hashes and instructions. Fifteen assets total 706,647,760 bytes, with all public URLs checked without credentials. The tag is tested commit 44918ef on this unmerged branch; the later commit adds publication/clean-checkout receipts and their auditable driver, without changing packaged training/inference code.
• Clean GitHub checkout: downloaded and verified all 14 initial draft assets; restored every indexed member; safely loaded all 147 public PT files; ran all 96 recovered public heads on a real 554-location section and the replacements on eight real image patches using a freshly downloaded official encoder. Fresh/cached features match exactly. The existing pinned environment on the same machine was reused; independent hardware/environment replication is not claimed.
• Metadata-only Phikon rights-review draft: 36 verified Phikon heads and their derivative archives remain local. No such binary was uploaded. Compatibility/permission under the custom Owkin terms and HEST obligations needs clarification; exact sources and checksums are documented.

Validation: 168 tests pass (142 original +26 new), along with all 29 publication targets/71 regression checks. All 83,136 frozen scientific/history files and original scientific Git trees are byte-identical. No manuscript, supplement, figure, table, bibliography, PDF or reported number was changed or rebuilt in place. No paid computing service was used.

Review entry points: docs/RELEASES.md; reproduction/REPRODUCTION_REPORT.md, INVENTORY.csv, RESULTS_COMPARISON.csv, RELEASE_ASSETS.json, RELEASE_VERIFICATION.json, PUBLIC_AVAILABILITY.json and PRESERVATION_REPORT.json.

Leave this draft PR unmerged for author review. Main remains 67738be.