[TimelordUK]

Fixes #497. Partly fixes #498 (doubles only; decimal/money as text is still open, see below).

#497

• d_variant wrote the first row's base type into the column definition, so every later row of a sql_variant column was read as that type. It now dispatches each row on its own type, and stores a NULL from its first SQLGetData (a second read returns SQL_NO_DATA, which lost the row).
• The 0xC0000005: a failed cell left a hole that fromQueryResult dereferenced. A missing cell is now null, and a failed cell ends the fetch: the partial row is dropped and the query ends in an error. When the driver gives no diagnostic, the error is ours (HY000 cannot read column N), so a failed read is never silent.
• bigint max read as a double came back as INT64_MIN; the range check now uses bounds that are exact as doubles.

#498 part 1: under numeric strings a double is String(n)'s text ("1e-10", "1.5e+300"), not %f ("0.000000").

Still open (owner's call): decimal/money/smallmoney are still read as a double, so digits beyond a double are lost; prepared statements bind numeric as a double. Reading SQL_DECIMAL as text fixes it, but the driver's text keeps the scale ("5.00") and has no leading zero (".0000001").

Tests

• Corpus scripts issue-497, issue-497-ns (agreed: every base type first, followed by all the others, with numeric strings off and on) and issue-498 (known-bug, pins what is still open). Before the fix, issue-497 was PROCESS FAILED 3221225477.
• Unit tests in params-complete-refactored.test.js.
• The replay now keeps whole rows for scripts (column values replay), and the recorder prints native bigints.
• Full suite on Windows / ODBC 18: 1190 passing, 1 pending.

🤖 Generated with Claude Code