Prepare exact per-segment quaternion differences for one physical plan.
Unchanged endpoint principal frames share only their unscaled rotation
vector and midpoint frame. Actual time intervals, point chords, masses,
moments, spin products and model-order sums are evaluated on every call;
unchanged midpoint intrinsic tensors share the physical owner's exact
moment/quaternion preparation before those current spin products.
States which endpoint prerequisites are retained and which current physical values remain evaluated.
The shortest-arc rotation vector of Q1 Q0 inverse and the half-way slerp depend only on the two endpoint rotations; their reuse leaves the original angular-rate division and orbital-plus-spin formula unchanged.
A bounded numeric quaternion reader is lent to the single shared interval reduction, while the raw entry computes fresh differences. Reentrant reads use that same uncached reduction so foreign getters cannot overwrite the active outer rate scratch.
Exact copied quaternion coordinates are the only reuse key; current velocities, inertia and momentum answers are never replayed or clamped.
O(S) current orbital/spin work for S segments. Unchanged endpoint frames omit quaternion inverse/product, logarithm and slerp; unchanged midpoint moments/rotation also omit intrinsic axis rotations. Changed dependencies do their original computations. Retained storage is O(S).
Each ordered index captures all eight endpoint coordinates once, then uses that same private tuple for Object.is, inverse/log and slerp. A row becomes reusable only after both results complete. Mutation/signed zero remain visible; current mass, points and interval seconds remain outside this rotation-only key.
The caller owns one closure, one active-call guard and at most one numeric interval per current part index; shrinking truncates old entries. Finally releases the guard after success or throw, with no depth pool or retained pose/model/result; the mass token drops this reader on release.
Prepare exact per-segment quaternion differences for one physical plan.
Unchanged endpoint principal frames share only their unscaled rotation vector and midpoint frame. Actual time intervals, point chords, masses, moments, spin products and model-order sums are evaluated on every call; unchanged midpoint intrinsic tensors share the physical owner's exact moment/quaternion preparation before those current spin products.
States which endpoint prerequisites are retained and which current physical values remain evaluated. The shortest-arc rotation vector of Q1 Q0 inverse and the half-way slerp depend only on the two endpoint rotations; their reuse leaves the original angular-rate division and orbital-plus-spin formula unchanged. A bounded numeric quaternion reader is lent to the single shared interval reduction, while the raw entry computes fresh differences. Reentrant reads use that same uncached reduction so foreign getters cannot overwrite the active outer rate scratch. Exact copied quaternion coordinates are the only reuse key; current velocities, inertia and momentum answers are never replayed or clamped. O(S) current orbital/spin work for S segments. Unchanged endpoint frames omit quaternion inverse/product, logarithm and slerp; unchanged midpoint moments/rotation also omit intrinsic axis rotations. Changed dependencies do their original computations. Retained storage is O(S). Each ordered index captures all eight endpoint coordinates once, then uses that same private tuple for Object.is, inverse/log and slerp. A row becomes reusable only after both results complete. Mutation/signed zero remain visible; current mass, points and interval seconds remain outside this rotation-only key. The caller owns one closure, one active-call guard and at most one numeric interval per current part index; shrinking truncates old entries. Finally releases the guard after success or throw, with no depth pool or retained pose/model/result; the mass token drops this reader on release.