Show the anisotropy the elastic averages throw away, and draw the Fermi surface

[Starlitnightly]

mv.pl.elastic — what four numbers were hiding

mv.prop.elastic computes the whole 6×6 stiffness tensor and stores it, then reports four Voigt-Reuss-Hill averages. Those are what a screen ranks on, and they are exactly the part of the tensor that survives averaging the direction dependence away. Computing the tensor and reading only the averages throws away the reason for computing it.

Copper, from its measured constants:

direction	Young's modulus	published
[100]	66.7 GPa	66.7
[110]	130.3 GPa	130.3
[111]	191.1 GPa	191.1
isotropic average	~120 GPa	—

A factor of 2.9 with direction, reported as one number. An isotropic solid returns exactly 1.000.

The subtle part is the Voigt factors. A compliance matrix carries a half on every shear index and a quarter on a shear–shear pair; a stiffness matrix carries none. Inverting the 6×6 and expanding without them gives a tensor that looks reasonable and reports an anisotropy that is not there. There is a test that deliberately drops the factors and asserts the result does go wrong, so that check cannot quietly stop checking.

mv.pl.fermi_surface — draw the stored mesh, not a recomputation

mv.elec.fermi_surface now keeps each sheet's vertices and faces. The interpolation takes minutes, and a plotting function that repeated it on every call would not be an interface worth having; keep_mesh=False is there for a screen that wants only the numbers.

A sphere wider than the Brillouin zone comes out with flat faces where it crosses the boundary — physics, not a rendering artefact.

Executing the tutorials found what the unit tests could not

mv.pl.elastic resolved a row against obs_names only, so row="Cu" reached int() and failed with a message about base 10 — a matverse dataset puts the formula in obs["name"] and leaves obs_names as integers.

The tests had named their own rows and always took the branch that matched. Fixed to follow mv.pl.structure's convention, with a regression test that says why it should have existed. This is the case for executing tutorials rather than rendering stored output.

Plotting coverage

Every plotting function is now called in a tutorial. The two that appeared only once — distribution and spacegroups — have a second use each, placed where they answer something rather than to make up a number: a histogram beside a screening threshold, and a space-group distribution over a search library.

Verification

981 pass on py3.12, 861 on py3.10. Contract-verified rate stays 165/165. All 16 notebooks execute. Docs build.

(Three TestBatchedEngine failures appeared in one run from GPU contention by other jobs on this node — nvidia-smi showed the card free afterwards and they pass on rerun. Not a regression.)

🤖 Generated with Claude Code

https://claude.ai/code/session_016wh5NzNE9Wn49nuUm7zg2a