[Tarzelf]

The next absolute-versus-reference-relative control experiment needs exact prior servo references. Existing model features round those references to float32. This adds a separately versioned, training-only export with float64 prior references and feasible next-reference labels, bound to the unchanged parent cache, protocol, source and control authority.

A standalone numerical preflight uses a common head scale and identical initial network weights. All 64 synthetic input vectors at the reset reference have bitwise-equal decoded outputs before projection; later-state differences and endpoint-loss units are explicit. It does not train or execute a policy.

Validation:

• 13 exact-reference synthetic tests; 34 combined cache/export regression tests.
• 15 head-preflight tests and one frozen 128-case numerical run.
• Independent peer review of both additions.
• CI adds the new checks to Linux, macOS and Windows with Python 3.12/3.13. Source-byte contracts retain LF endings.

This is stacked on the reviewed cache branch. Frozen v0.1 runtime, models, released wheel, installer and qualification remain unchanged: the learned candidate is still 5/64 and unqualified. The proposed 164-attempt learning screen remains unrun. No real recordings were processed by these new tools in this initial source change.