Correctness fixes found in the pre-release review

[vaceslav]

From the pre-release review (seven independent reviewers). Each item was confirmed with a reproducer; the first three were re-checked by hand.

Wrong or silently corrupted values

• Mis-grouped numbers import as other numbers. Under a German plan 19.99 imports as 1999 and 1.5 as 15 (ValueReading.cs parses with NumberStyles.Number / AllowThousands), while the profiler's HasWellFormedGroups refuses them and the precheck blocks them. The import must apply the same well-formed-grouping rule as the profiler.
• ImportPolicy.AllOrNothing is ignored by the run. Only MappingPrecheck reads TargetSchema.Policy; extractor and importer import partially. Either enforce it (refuse the run / stop at the first failing row) or move the setting to where it is honoured.
• Precheck is blind to native xlsx numbers mapped to a Date/Boolean/Integer field (MappingPrecheck.cs returns early for natively typed columns): it passes, then every row fails — or, under de-DE, native 1.5 goes through the text fallback and becomes 15.
• Zoned ISO dates are shifted to the host's time zone. …Z / +05:00 text is parsed with default styles in XlsxCursor (t="d") and DateReading, so one file yields different dates on different servers. Keep wall-clock (or RoundtripKind + documented rule).
• HeaderRowIndex counts present <row> elements while every reported row number uses the sheet's r; with sparse rows a header on spreadsheet row 3 skips the header and the first data row. Decide one meaning (spreadsheet row) and use it everywhere.

Failure modes

• Malformed XML throws XmlException from every XmlReader path (workbook, styles, relationships, shared strings — straight out of the constructor). Not documented; a server catching the documented types crashes on a hostile upload. Wrap into the library's exception (see the API issue).
• A truncated file reads as a shorter, plausible table. When the stream ends mid-row/cell the scanner returns a clean end, and the cursor emits a corrupted final row without any signal. Detect an unterminated document at EOF and fault.
• A poisoned xlsx cursor stays poisoned after MoveToSheet reopens a sheet from the start: it returns true, then ReadRow throws "cannot continue". Clear the fault when a sheet is reopened, or make MoveToSheet refuse.
• .xls / .xlsb / other binary files are sent to the csv cursor (TabularFile.Detect calls every non-zip a csv; an xlsb zip fails with "workbook.xml missing"). Decided: reject both with a clear "format not supported" error; also refuse obvious binaries instead of profiling them as text.
• AnalysisOptions.Cultures = [] throws IndexOutOfRangeException inside the profiler — validate options.
• Negative / zero r row numbers are accepted.

Needs a decision

• Invariant globalization. Under DOTNET_SYSTEM_GLOBALIZATION_INVARIANT=1 the default cultures (de-DE, en-US) do not exist and 168 of 394 tests fail. Degrade gracefully (skip unavailable cultures, report it) and document it.
• German-leaning defaults (Cultures = ["", "de-DE", "en-US"], ; first in delimiter candidates) — keep, but state them in the README/guide.

[vaceslav]

All items done (406f6f3 … b18cd0b), each with tests written to fail first:

• Mis-grouped numbers: the import applies the profiler's grouping rule (19.99 under de-DE is a type mismatch, not 1999).
• AllOrNothing: the run stops at the first failing row; All() returns no items; the failing batch carries none.
• Precheck judges natively typed xlsx columns against the field.
• Zoned ISO timestamps keep the clock time as written, on every server (tested under three time zones).
• HeaderRowIndex names a spreadsheet row, the numbering every report uses.
• Malformed XML → InvalidDataException (inner XmlException); truncated worksheets fault instead of shortening; reading stops at </sheetData>.
• A successful MoveToSheet clears the previous fault.
• .xls/encrypted (OLE2), binary files, .xlsb and .ods are refused by name; UTF-16 without BOM is recognised.
• AnalysisOptions validated up front; unknown culture names refused; invariant globalization degrades (named cultures left out, mapping.unknown-culture from validator and precheck) and is documented, with the default cultures, in the guide.
• Row numbers only ever increase (1…1,048,576, after the previous row).

Regression checks: the ~1,100-workbook corpus from #8 fails on exactly the same 44 files as before (encrypted, bombs, empty/corrupt zips); 41 real csv files still read. Performance on the large fixtures, incl. the 5M-row csv: row counts identical, readers within ±4 %, full csv analysis 7 % faster with 2–3× less allocation (the grouping check no longer allocates).