Audit non-crypto modules' L1 tests for self-consistency-only gaps

#87 found that the ntt module's L1 "known-answer" tests were checked against
a naive reference implementation written for this project, not an
independent external source. Crypto modules now require external,
independently-citable references (CONTRIBUTING.md's tightened L1
definition), but the same self-consistency pattern may exist in the
non-crypto modules (linalg, stats, integrate, differentiate, interpolate,
poly, ode, signal, fft, autodiff, compressed_sensing, sketch).

Lower priority than the crypto work: for general numerical modules an
internally-written reference is still valid per CONTRIBUTING.md, so this is
about checking whether the analytical-truth/closed-form claims actually
hold, not about requiring external validation everywhere. Phase 2 of the
citation/validation rigor work, after #87 and #88 land.

Start by reading CONTRIBUTING.md's tightened L1 definition and reviewing the L1 known-answer tests across linalg, stats, integrate, differentiate, interpolate, poly, ode, signal, fft, autodiff, compressed_sensing, and sketch. Compare each analytical or closed-form claim with its test reference, then document which modules have self-consistency-only gaps and what follow-up validation is needed after #87 and #88.