Fix #348: discriminate symbols across domains in expression identity

[petlenz]

Closes #348.

A scalar symbol x and a tensor symbol x share id() (both index 0 of their own domain's type list) and hash (symbol_base::update_hash_value covers only the name). expression::operator== / operator< therefore fell through to the same-type downcast in visitable_impl, casting a tensor to a scalar.

Reproduced at 05f3c3a, worse than the issue reported:

Build	Behavior
Debug	assert(dynamic_cast<Derived const *>(&rhs) != nullptr) → abort, core dumped
Release (NDEBUG)	SIGSEGV, core dumped (exit 139) — the wrong-type read walks a std::string out of a tensor's layout

Fix

• expression::operator== / operator< (src/numsim_cas/core/expression.cpp): compare the dynamic type once id() and hash tie, before the downcast. operator< orders unequal types with type_info::before. Within a domain, equal id() already implies equal type, so this changes nothing there.

  RTTI is safe here: the library builds as a static archive and its typeinfo carries ordinary mangled names, so libstdc++ compares them with strcmp — deterministic, and correct even across a DSO boundary. There is no -fvisibility=hidden, and Windows sets WINDOWS_EXPORT_ALL_SYMBOLS.

• symbol_base::operator== compares m_name; operator< uses it as the tiebreak (include/numsim_cas/core/symbol_base.h). Two names whose per-char hashes collide no longer merge.

• relation::operator< (include/numsim_cas/core/assumptions.h) orders by the operands themselves instead of their hashes, so colliding relations stay distinct in relation::set.

Not changed: the hash discriminator

The issue also proposes mixing a domain tag into symbol_base::update_hash_value. I implemented it and measured: 43 existing tests fail, all print-order flips (y*pow(x,2)*z*X vs pow(x,2)*y*z*X) — same factors, different hash-driven order. Symbol hashes drive operator<, which drives n-ary print order, and many tests pin that output.

It is also not needed for correctness: the hash is a fast-reject and an ordering key, not an identity. With the type discrimination above, a hash tie between a scalar and a tensor symbol is a benign collision — std::map orders through operator<, which is now total and UB-free. I left the hash alone rather than rewrite 43 expected-output strings; changing print order deserves its own decision.

Tests

• SymbolIdentity.SameNameAcrossDomainsIsDistinct — cross-domain == is false and the order is total.
• SymbolIdentity.EvaluatorKeepsBothDomainBindings — evaluator_base keys one std::map<expression_holder<expression>, std::any>, so both domains share it; binding a scalar x and a tensor x must keep both values.
• AssumptionFixture.RelationSetKeepsDistinctOperands — relation<> has no users anywhere, so its operator< was never instantiated; this test compiles the changed code and pins that distinct relations stay distinct. It cannot demonstrate the collision fix (no way to force a hash collision through the public API).

Negative control (revert expression.cpp only, keep the tests):

	Debug	Release
SameNameAcrossDomainsIsDistinct	abort in equals_same_type	SIGSEGV
EvaluatorKeepsBothDomainBindings	abort in less_than_same_type	SIGSEGV

Both pass with the fix.

Suite: Debug 2349/2349, Release 2349/2349 (gcc-14, -Werror minus the pedantic / deprecated-declarations warnings already on main). clang-format-18 clean.

Not fixed (reported)

• n_ary_tree::like_term_of and tensor_scalar_mul::like_term_of use the same "id + hash tie ⇒ downcast" pattern. Not reachable cross-domain today — I probed scalar_add vs tensor_add: ids (4 vs 1) and hashes both differ, and like_term_of is only called from per-domain n-ary merging. Left alone because n_ary_tree.h is being edited by Fix #369: drop noexcept from allocating paths #436 and there is no repro to verify against.
• A tensor symbol's hash and equality ignore dim/rank, so A(3,2) and A(2,2) are the same key. Same name-only-identity family, tracked separately in Tensor symbols with the same name but different dim/rank are one key — evaluator returns the wrong tensor #443 — closing Scalar and tensor symbols with the same name collide (name-only hash, per-domain ids): assert abort in debug, UB with NDEBUG #348 does not cover it.

Overlap

Touches symbol_base.h (#431 adds the hash reset in update_hash_value) and the comparison operators (#436 removes their noexcept). Hunks are kept minimal and disjoint from both.