Prepare exact rotated intrinsic tensors for one assembly operation.
Each reader snapshots only the last inertia and principal quaternion at
each part index. Mass, centre and reference point remain current inputs;
the parallel-axis term and part-order accumulation run on every call.
Release the reader with its physical planning operation.
States cached intrinsic inputs, current shift inputs, exact numeric keys and operation ownership.
R diag(I) R transpose is independent of both part centre and reference point; caching its nine scalar entries preserves the original axis-product association before the unchanged parallel-axis addition.
One last rotated tensor per ordered part feeds the same nine-entry accumulation. An active-call guard delegates reentrant requests to the existing direct component kernel so foreign getters cannot overwrite outer cached rows or spin scratch.
Every changed inertia or quaternion coordinate recomputes its actual tensor, and every mass/centre/reference shift remains evaluated. No normalized-input shortcut, symmetric-entry copy, time key or tolerance is introduced.
O(P) comparisons, shifts and ordered sums for P parts, with three axis rotations and nine row products only at changed intrinsic frames. Each call returns three fresh rows; retained scalar snapshots are O(P).
Each moment/quaternion component is captured once before key comparison; that same private tuple drives every product. Object.is retains signed zero and caller edits, and a failed changed-row computation stays ineligible for reuse. Mass and translation do not enter an intrinsic tensor's key.
The caller owns one active-call guard and assembly/spin slots per current part index, each with a seven-scalar key and nine entries; either reader truncates both populations on shrink. Finally restores the guard after success or throw. No part, result or depth pool is retained, and the mass owner drops this reader on release.
Prepare exact rotated intrinsic tensors for one assembly operation.
Each reader snapshots only the last inertia and principal quaternion at each part index. Mass, centre and reference point remain current inputs; the parallel-axis term and part-order accumulation run on every call. Release the reader with its physical planning operation.
States cached intrinsic inputs, current shift inputs, exact numeric keys and operation ownership. R diag(I) R transpose is independent of both part centre and reference point; caching its nine scalar entries preserves the original axis-product association before the unchanged parallel-axis addition. One last rotated tensor per ordered part feeds the same nine-entry accumulation. An active-call guard delegates reentrant requests to the existing direct component kernel so foreign getters cannot overwrite outer cached rows or spin scratch. Every changed inertia or quaternion coordinate recomputes its actual tensor, and every mass/centre/reference shift remains evaluated. No normalized-input shortcut, symmetric-entry copy, time key or tolerance is introduced. O(P) comparisons, shifts and ordered sums for P parts, with three axis rotations and nine row products only at changed intrinsic frames. Each call returns three fresh rows; retained scalar snapshots are O(P). Each moment/quaternion component is captured once before key comparison; that same private tuple drives every product. Object.is retains signed zero and caller edits, and a failed changed-row computation stays ineligible for reuse. Mass and translation do not enter an intrinsic tensor's key. The caller owns one active-call guard and assembly/spin slots per current part index, each with a seven-scalar key and nine entries; either reader truncates both populations on shrink. Finally restores the guard after success or throw. No part, result or depth pool is retained, and the mass owner drops this reader on release.