Original direct-reference-v0 result. Read record/REPORT.md and MANUAL_REVIEW.md. launch-source.zip preserves the prospective implementation, design, image provenance, neutral checks and reproduction commands. The comparison measures stock native against a source-matched Mode50 reconstruction on that exact stock base. Historical Mode50 image identity is unestablished; credit for the method belongs to its published authors. Three repetitions on ten development and forty broader identities are reported separately; all identities have prior research exposure. Missing quality remains unknown, incomplete cost is a lower bound, and necessary-condition failures remain separate from formal decisions. Complete physical expense includes construction, retries and grading once. Public numerical projections retain saved values. No task, grader, bootstrap or selector is rerun for this archive. Eight prospectively chosen deliveries supply fixed interpretation; the record does not claim causal attribution from this sample. Source fixtures establish operation, not capability. Private task sources, delivered raw content, histories and provider payloads are omitted. The prelaunch recovery exposure is disclosed in support/notes/direct-reference-prelaunch-recovery-scope.md. Saved tables can be reproduced from this ZIP alone using support/analysis/direct_reference_tables.py --archive with Python's standard library. Experiment reproduction additionally needs external unchanged benchmark/baseline assets, pinned images and metered Luna access. See support/notes/direct-reference-reproduction.md.