Epic: diagnostic outputs - parity, cost and verification (nd / LLt / MIR)

[neuromechanist]

Four issues about quantities AMICA reports but never fits on. None of them can change a decomposition, which is why they sit apart from the parity-gate work in #230.

Children

• numpy_impl ndtmpsum uses current-iteration gm; Fortran uses the previous iteration's #219 — numpy_impl uses the current iteration's gm for ndtmpsum; Fortran uses the previous one. Parity deviation, visible only under share_comps multi-model. Includes checking AMICATorchNG for the same ordering.
• Document that min_nd is unreachable on small datasets in all three implementations #218 — document that min_nd=1e-7 is unreachable on small datasets in all three implementations (Fortran's own floor is ~1e-5). Documentation, not a fix: the default is Fortran-faithful.
• Test mir_history_ against the keep_best rollback and save/load #161 — mir_history_ survives the keep_best rollback and is empty after save/load. Both documented, neither tested.
• Stash per-sample logV during the E-step instead of recomputing for LLt #157 — stash per-sample logV during the E-step instead of recomputing full-dataset LLt at write time. Carries a real design decision: whether to keep Fortran's one-M-step-stale LLt or the current self-consistent one.

Order

#161 before or with #157: they share the keep_best rollback and the same test recipe for forcing it, and #157's acceptance criteria need exactly those assertions in place first.

#219 and #218 are independent of both and of each other, beyond both concerning nd.