Invalid expression_holder is rejected by get()/operators but silently accepted by evaluators (returns 0), diff and to_string

[petlenz]

An invalid (default-constructed / moved-from) expression_holder is rejected by some entry points and silently accepted by others:

Entry point	Behaviour on an invalid holder
holder.get(), *holder, holder->	throws "expression_holder: access to invalid (null) expression"
invalid + x (operators)	throws (same message)
to_string(invalid)	no throw, returns a string
scalar_evaluator::apply(invalid)	no throw, returns 0 (scalar_evaluator.h:29-36)
diff(invalid, x)	no throw, returns a valid holder

So ev.apply(std::move(e)) after an accidental move, or evaluating a holder that a failed rebuild left empty, returns 0.0 with no signal — a silent wrong value, the class the rest of this week's work has been removing. diff producing a valid result from nothing is the same problem one step later.

Severity: low–medium. Needs a programming error to reach, but the failure is silent and the inconsistency means callers cannot rely on either behaviour.

Fix: pick one contract and apply it everywhere. Given get() already throws, the consistent choice is for apply, diff, substitute, to_string/to_latex and the limit/contains entry points to call throw_if_invalid() (or get()) up front. The evaluators' is_valid() branches that return ValueType{0} should become throws; the differentiation visitors' internal "invalid = zero accumulator" convention is separate and can stay internal.

Test: one parameterised test that feeds a default holder to each public entry point and expects invalid_expression_error (or whichever type is chosen).

Signed-off-by: petlenz peterlenz89.pl@gmail.com