[rogeriojorge]

Hou-Li-type exponential Hermite filter that skips n <= 2 (stacked on #60).

What

• filter_spectrum(Nn, Nm, Np, order=36, keep=2) returns the exposed shape s(n) = (n/(N-1))**order for n > keep, and exactly 0 for n <= keep, summed over the velocity axes. Shape (Np, Nm, Nn), like hypercollision_spectrum.
• exponential_filter(Ck, strength, s) applies exp(-strength*s) between steps. Hou & Li use strength = order = 36. With zero strength the state is returned unchanged, bitwise.
• RHS form: pass s as "collision_matrix" with nu = rate. Over a time dt this is the filter with strength = rate*dt, and it needs no driver change.
• With keep = 2, every moment of total degree <= 2 (density, momentum, full pressure tensor) is untouched for any basis centre and width.

Tests (local only) in tests/test_filter.py:

• the spectrum values, shape and zero block;
• bad order and keep values are refused;
• zero strength is a bitwise identity;
• low moments change by less than 1e-14 relative in a shifted, rescaled 3V basis while the tail is damped;
• the RHS form equals -rate*s*C exactly.

Full suite at 7b94668: python -m pytest -q -n 2 tests gave 191 passed (macOS, Python 3.12, JAX 0.11.2, diffrax 0.7.2). GitHub CI does not run on PRs that do not target main.

Dependencies: #56–#60. Merge after #60.

Open question: default order 36 (Hou-Li) vs a gentler order for RHS use. The physics study using it, with strength and order varied, is in Dark-SPECTRAX branch lane-a.