SqlVariant::Get<T>() is noexcept but calls std::get — a type mismatch terminates the process

[christianparpart]

Problem

SqlVariant::Get<T>() (src/Lightweight/DataBinder/SqlVariant.hpp) is declared noexcept, and its fallback path is std::get<T>(value). std::get throws std::bad_variant_access when the variant holds a different alternative, and a throw crossing a noexcept boundary calls std::terminate. So asking for the wrong type aborts the process instead of reporting an error.

auto v = SqlVariant { 42 };        // holds int
auto d = v.Get<double>();          // std::terminate — not an exception, not a wrong value

This is reachable from ordinary code: which alternative a column fills depends on the driver's reported SQL type, so a caller can be right about the column and still wrong about the alternative. #602 and #586 are both examples of a driver filling an alternative the caller did not expect.

Scope

Get<T>() converts for the cases where a column's alternative recently moved (a DECIMAL column fills SqlDynamicNumeric, a binary column fills SqlBinary), so those specific conversions are safe. The general mismatch is not. ValueOr<T>() was already made total in #606 — it returns the caller's default rather than reaching for an alternative the variant does not hold.

Why it is filed rather than fixed

Fixing it is a contract decision, not a mechanical change, and there are three defensible answers:

1. Drop noexcept and let std::bad_variant_access propagate. Matches TryGetNumeric()/TryGetBinary(), which already throw. Changes the exception profile of a hot accessor.
2. Keep noexcept and return T {} on mismatch. Total, but silently turns a caller bug into a zero — the failure mode this library generally refuses.
3. Keep Get<T>() for the checked case and steer callers to TryGet*/ValueOr<T>.

(1) looks right, but it is an API change that deserves its own PR rather than riding along with unrelated work.

Context

Found while working on #606, which is also where ValueOr<T>() was fixed. Get<T>() was deliberately left alone there to keep that PR to one concern.

[Yaraslaut]

Resolved on master by #627 (commit 8b26daf, "fix: stop promising noexcept from accessors that run queries"), which took option 1 from this issue: SqlVariant::Get<T>() is no longer noexcept, and its documentation now states that it throws std::bad_variant_access when the variant holds neither T nor one of the alternatives it converts from. A mismatched Get<T>() therefore raises an exception the caller can handle instead of terminating the process; the TryGet* accessors and ValueOr<T>() remain the non-throwing probes.