Reusing unchanged packet prefixes

Complete-cadence prediction bounds currently repeat much of the same packet history. Reusing a verified prefix can reduce that cost without narrowing a physical bound or changing a model. The accepted design is in the prefix plan.

The first component gives the reference, domain-derivative and point-derivative source kernels typed copy-and-retarget entry points. New translation units include the existing kernels; no measurement or derivative formula is copied. Each state records whether every processed packet was a constant, complete packet. A fork must reject any history that touched a clipped fraction, even if a later query would put that history before its endpoint.

The copied state includes the body rotation, all 27 parameter columns, force, mount transformation, elapsed time and derivative intervals. Copying it must not change its parent or another copy. Its layout, binary/compiler identity, input hashes, physical domain, reference binding and packet allowance remain attached. Retargeting changes only the query expression, through C++ fields of the declared type. Python must never guess byte offsets into a buffer.

Both represented packet tags and exact binary-duration arithmetic constrain the fork. The last packet must end strictly before the earliest supported new query. Equality remains a knot case and is not eligible. The original packet allowance is preserved; reused work is not counted as newly calculated work or a fresh start.

Small tests compare independent cold source streams with copied streams for nonuniform periods, nonzero clock shifts, exact endpoints and adjacent values. They check all output values and derivative columns, failure behavior and parent isolation. A separately frozen finite source comparison follows only after those tests pass. No aircraft files, fits or protected confirmation outcomes are needed.

This source-only component does not qualify complete native-prefix reuse. The geographic position/velocity, relative attitudes, accumulated force integrals, common-clock state and conditioning proof still need a complete-state implementation and independent cold-run checks before it can replace an existing native bound. V01 remains unqualified and the 53-stretch findings remain unchanged.

The complete-state short controls have now passed separately. Their long-history follow-up remains necessary.