tests: generic conformance checks over every material (Tier 1 of #35)

[petlenz]

Part of #35 — Tier 1 only. Stacked on the #33 fix, which it would otherwise fail.

Five checks, none needing a per-material fixture:

• every class deriving material_base is registered, or on an explicit opt-out list with its reason
• the opt-out list has no stale entries, so an exemption cannot outlive the material it exempts
• every schema declares name
• no schema is empty (the signature of a parameters() that forgot to chain base::parameters())
• a missing required parameter throws

The opt-out list is the point

Three materials are deliberately absent because they are templates registered under concrete aliases. Each is now a decision someone wrote down, instead of being indistinguishable from an oversight — which is exactly how #33 happened.

Verified load-bearing

Deleting one register_type() line fails the check, naming the material and the fix. The tensor_component_stepper exemption was itself found by the check on its first run — it is registered under _rank1/_rank2, which the check could not know.

Tiers 2 and 3 need a type -> minimal JSON fixture table and are not in this PR.

[petlenz]

Critical review of my own PR. One finding: a test in it asserted nothing. Fixed in the branch.

MissingRequiredParametersThrow could not fail

try {
  nm::create_from_json<policy>(ctx, j);
  // Reaching here is fine only if nothing was actually required.
} catch (const std::exception&) {
  // Threw, which is the contract.
}

Both branches empty. The only assertion was checked > 0 — that the loop had run. It passed whether or not any material enforced a required parameter, in the PR whose entire purpose is catching that class of test.

The cause was real: is_required is a check attached with .add<is_required>(), and the controller exposes no way to ask whether a parameter is required. A loop over every registered type genuinely cannot distinguish "threw because something was required" from "succeeded because nothing was". Faced with that, the first version kept the loop and quietly dropped the assertion instead of narrowing the claim.

It now names five types with known-required parameters and asserts each throws on a document containing only name. Not generic, and the comment says so, along with the condition that would make it generic again — if the controller ever exposes required-ness, the list goes and the loop returns.

Verified load-bearing: swapping linear_stress's is_required for a default fails it, naming the material.

What I checked and could not break

• The registration check fails, by name, when any single register_type() line is deleted.
• The stale-exemption check works: an opt-out entry for a material that no longer exists fails.
• The tensor_component_stepper exemption is genuine — registered as _rank1/_rank2 — and was found by the check itself.
• CI green.