Fix #439 and #447: t2s rank gates throw; no real claims for a negative radicand

[petlenz]

Closes #439.
Closes #447.

#439 — t2s invariants reject a wrong-rank operand. trace, norm and det gated rank with assert, so an NDEBUG build constructed a malformed node that reached the evaluator. They now throw invalid_expression_error through the same validate_rank helper the tensor side uses (#438), and dot's hand-written check moved onto it too.

Side benefit: the rules added in #427 rely on trace's rank precondition — try_trace_identity assumes a rank-2 identity, try_trace_outer_product assumes rank-1 operands. A real throw makes those preconditions hold in release builds as well; verified that trace(I4) and trace(otimes(X,X)) (rank 4) both throw.

#447 — no real-domain claims for a negative radicand. sqrt claimed nonnegative + real for any argument, and pow claimed real whenever base and exponent were real, though both are NaN on a provably negative operand. Now suppressed when the base is provably negative and the exponent is provably fractional — including a negated fractional exponent (pow(n, -1/2)) and a double spelling (0.5), both verified by probe.

Deliberate scope: an unknown-sign operand keeps the claims, since unassumed symbols are real by default in this propagator. Tightening that would mean sqrt(x) on an unassumed x claims nothing, a broader policy change than this issue.

Negative controls:

• t2s trace/norm/det gate rank with assert only — throw invalid_expression_error (follow-up to #360) #439 in Release (the whole point, since the assert vanishes under NDEBUG): without the fix trace(C), trace(u) and norm(C) on wrong-rank operands "throw nothing"; with it, T2sShapeValidation.InvariantsRejectNonRank2 passes.
• Scalar propagator claims nonnegative/real for sqrt and pow of a negative argument (value is NaN) #447 in Debug: AssumptionFixture.NoDomainClaimsForNegativeRadicand fails without the fix.

Full suite 2407/2407 (gcc-14 Debug) plus a Release build; clang-format clean.

Authoring note: the agent that wrote these changes was interrupted by a rate limit mid-review. Its in-progress probe (the negated-exponent hole it suspected, and the #427 preconditions) was completed — both came back clean — the probe file removed, and the verification above run against the finished branch.