Hold the lattice to the temporal types MEOS admits

[estebanzimanyi]

temporal_type() admits 20 user-facing types; the lattice's leaves model 14.
Six have no leaf class — T_TH3INDEX, T_TQUADBIN, T_TS2CELL,
T_TPOSECHAIN, T_TPCPOINT, T_TPCPATCH — so the object model cannot express
them and a binding projected from it reaches those values only through the flat
C surface.

The gate already here could not see it. Its assertions hold each node to its own
predicate, which passes whether or not a type has a class at all: the root's
membership comes from temporal_type() and already names T_TPOSECHAIN, while
nothing asked which types a LEAF claims. Coverage is a different question from
membership, and only the second was being asked.

_UNMODELLED records the six as a RATCHET rather than a permission. The
assertion compares against it exactly, so a newly uncovered type fails the
suite, and giving one of these a class fails it too until its entry goes — the
gap can neither grow nor close unnoticed. Both directions are exercised:
removing an entry fails, and a leaf that stops claiming its type fails.

A type claimed by two leaves fails as well, since two classes for one type is an
ambiguity the projection cannot resolve.

What the six need is not uniform. T_TPOSECHAIN sits beside TPose, and the
point-cloud pair beside the other spatial leaves. The three cell-index families
are one question rather than three: H3, quadbin and S2 are the same kind of
thing, a temporal value over a discrete global grid, and whether they share an
abstract parent the way TPoint groups TGeomPoint and TGeogPoint is a
taxonomy decision for the model's owner. The comment says so where the next
reader will meet it.