Fix container type and empty-vector defects in the composite quadratures

[susilehtola]

Three unrelated defects in the same two headers.

SubQuadrature's generator declared its weights in the point container
type and returned it as the weight_container half of the tuple:

point_container points ( ... );
point_container weights( ... );   // wrong type

This compiles today only because every radial quadrature happens to use
std::vector for both. It breaks as soon as the two differ, which
is exactly what mixed precision needs: SubQuadrature<Becke<float,double>>
does not compile before this change. Two of the surrounding assertions
also tested >= 0 on size_t, which is vacuous; drop them and keep the
bounds that say something.

RadialGridPartition::finalize() read partition_idx_.back() without
checking that anything had been added, which is undefined behaviour on an
empty vector. make_pruned_grid() guards against an empty region list one
layer up, but finalize() is public and the class is default-constructible,
so the guard is not structural. Throw instead.

Six adaptor overrides in the two spherical quadrature headers were
written as override {; -- a stray empty statement after the opening
brace. Harmless, but remove them.

Co-Authored-By: Claude Opus 5 (1M context) noreply@anthropic.com
Claude-Session: https://claude.ai/code/session_01FDTFYJMQ76iujDFNHzZyXF

[wavefunction91]

See inline comments.

[wavefunction91]

The same default-constructed partition still hits UB through the iterators: end(), cend() and end() const (lines 141/148/155) compute partition_idx_.end()-1, which on an empty vector steps before begin(). With -D_GLIBCXX_DEBUG on this head, RadialGridPartition<LebedevLaikov<double>>{}.end() aborts with "attempt to retreat a past-the-end iterator". Since this PR is about the empty-vector cases, could those be made safe here too? A REQUIRE_THROWS next to the default-constructed rgp in test/composite_quadratures.cxx:202 would also pin the new finalize() behaviour.

[susilehtola]

You are right, and the iterators were the live path. Fixed in 7ab221b: end(), cend() and end() const present an empty range when partition_idx_ is empty instead of computing end()-1. I reproduced your abort first — -D_GLIBCXX_DEBUG on master gives "attempt to retreat a past-the-end iterator" from RadialGridPartition<LebedevLaikov<double>>{}.end(), and the patched version reports an empty range.

The new "Unfinalized Partition" section pins REQUIRE_THROWS( rgp.finalize(r) ) together with begin() == end() for all three overloads.

[wavefunction91]

Nothing in the suite instantiates SubQuadrature, so the mixed-precision case this fixes isn't covered, and it would regress silently. Could we add a small case, e.g. SubQuadrature<Becke<float,double>> over a slice of a Becke<float,double>, checking that the points and (double) weights match the parent's slice?

[susilehtola]

Added as a "Sub Quadratures" case over Becke<float,double>, checking the slice against the parent exactly, plus static_asserts on both container types.

One correction to the premise, in the good direction: it would not have regressed silently. With master's sub_quadrature.hpp the instantiation does not compile at all — make_tuple yields tuple<vector<float>,vector<float>>, which will not convert to tuple<vector<float>,vector<double>>. So the case was simply absent rather than passing for the wrong reason.

[wavefunction91]

This comment describes the asserts that were removed rather than the code that remains. Could we drop it?

[susilehtola]

Dropped.