Hydration round-trips every fetched value through std::wstring (lossy + slow reads)

[jkalias]

Problem

Database::Fetch builds a FetchQueryResults whose row_values is a std::vector<std::vector<std::wstring>> (include/fetch_query_results.h). FetchRecordsQuery::GetResults stringifies every column into a wstring via GetColumnValue, and FetchRecordsQuery::Hydrate then parses each string back into the typed member. This text round-trip on the read path is both a performance cost and the root cause of two concrete data-corruption bugs.

Concrete defects caused by the text round-trip

1. 64-bit integers truncated to 32 bits (consolidated from #15)
GetColumnValue reads SQLITE_INTEGER with the 32-bit sqlite3_column_int (return std::to_wstring(sqlite3_column_int(stmt_, col));), and FetchMaxIdQuery::GetMaxId does the same. Any stored int64_t above INT32_MAX — including AUTOINCREMENT ids once they grow large — is silently truncated/wrapped on read, breaking subsequent Fetch(id) / Update / Delete. Fix: use sqlite3_column_int64.

2. REAL values lose precision (consolidated from #16)
GetColumnValue formats SQLITE_FLOAT via std::to_wstring(double), which uses %f with a fixed 6 decimal places, so e.g. 0.123456789 round-trips back as 0.123457 and large/scientific magnitudes are mangled. Writes are fine (sqlite3_bind_double stores the true value); the loss is only on the read path. Fix: read sqlite3_column_double directly, or format with full round-trip precision (%.17g).

Performance / memory

• Every value incurs an allocation + parse on top of the SQLite copy; significant for wide rows / large result sets.
• The entire result set is materialized as strings before hydration (no streaming), so large queries hold the whole set twice (string + typed form).

Suggested direction

Hydrate directly from the prepared statement using the typed column accessors (sqlite3_column_int64, sqlite3_column_double, sqlite3_column_text/_bytes) into the destination members, skipping the intermediate wstring. This removes the per-value allocation/parse, fixes both data-corruption bugs above at the source, and opens the door to row-by-row streaming.

Acceptance criteria

• Integer columns round-trip values greater than INT32_MAX exactly (regression test).
• Double columns round-trip high-precision values exactly (regression test).
• Hydration reads typed columns directly rather than via std::wstring.

Related (tracked separately)

• StringUtilities::ToInt/ToDouble silently return 0 on parse failure #19 — StringUtilities::ToInt/ToDouble silently return 0 on parse failure. This is an independent error-handling defect in those helpers; moving off the text path (per this redesign) reduces reliance on them but does not by itself fix them, so it remains its own issue.

---

Consolidates #15 (int truncation) and #16 (REAL precision), which were symptoms of this single root cause and are closed as duplicates of this issue.