feat(write): write xlsx through TabularWriter, with our own streaming zip writer

[vaceslav]

What and why

Closes #71. Part 2 of 5 of #67: TabularWriter.Create(stream, TabularFormat.Xlsx) streams an xlsx workbook — several sheets, typed cells, flat memory — that the library's own import reads back as written and the Open XML SDK's schema validator accepts.

Spec: docs/superpowers/specs/2026-10-03-writing-design.md. Plan: docs/superpowers/plans/2026-10-03-writing-part-2-xlsx.md.

• Our own zip writer (ZipWriter, internal), because ZipArchive does not fit: written to a stream that cannot seek, LibreOffice refuses its output; given one that can, it keeps each whole entry in memory. Small parts are stored with their sizes up front; each sheet streams through DeflateStream with a data descriptor; zip64 past 4 GB; a slicing-by-8 CRC-32 of our own (System.IO.Hashing is a package).
• XlsxSheetWriter: inline strings (no shared-string table held in memory), _x000D_ for CR and _x005F_ before a literal _x, booleans, numbers, dates as serials across Excel's 1900 leap-year bug, four styles (General, date, date-time, and 0 for integers so an id does not show as 1.23457E+11), column widths, r on every row and cell. It refuses what a workbook would not give back: a long a double cannot hold exactly or a decimal past 15 significant digits (write.precision-loss), a date before 1900 (write.date-out-of-range), text over 32,767 characters (write.text-too-long); more than 1,048,576 rows is limit.exceeded.
• Sheet names follow Excel's rules (1–31 characters, none of []:*?/\ or control characters, no apostrophe at either end, not History, unique ignoring case), checked by TabularWriter for every format that names sheets — ods inherits them.
• XlsxWriterOptions.CompressionLevel, Fastest by default for now; part 5 benchmarks it.
• Reader fix: XlsxCursor.TryFromSerial turned the rounded milliseconds into ticks with a double multiply, inexact past about year 2355 — 9999-12-31 23:59:59.999 read 24 µs short. It now multiplies in long. Not benchmarked: it is one constant division and one long multiply per date cell, in place of one double multiply.
• RowText, the row buffer, is shared by the csv and xlsx writers (the csv writer's behaviour is unchanged; every csv test passes untouched). An abandoned writer now releases its open deflate entry.

Performance (final review, Apple arm64, net10, 1,048,576 rows × 8 columns): 1.75 s at Fastest, 1.3 s stored, 3.8 s at Optimal; 425 MB of sheet XML → a 76 MB file; about 3 MB allocated in total, no gen0 collections — memory is flat in the row count. Part 5 will measure buffered row writes (7–9 % measured) and a hardware CRC.

Checked beyond the tests (final review): LibreOffice 25 headless converts a written workbook with every value correct; Info-ZIP, Python's zipfile and streaming bsdtar accept zip64 output; 5M random decimals and 20M random dates agree between writer and reader. Excel itself is checked by hand before the release.

Tests: the Open XML SDK validator (test project only — the library still takes no package) on every workbook; round trips through TabularImporter for every value kind and edge value; a fuzz over three seeds; a full 1,048,576-row sheet read back to its last row and the next row refused; zip structures incl. zip64 at a lowered threshold; CRC against standard vectors and a bitwise reference; the reader fix against the old behaviour.

Checklist

• A test that failed before the change and passes after it (for a fix or a new behaviour).
• dotnet build and dotnet test pass on net8.0 and net10.0 with zero warnings. (2324 tests)
• If the read path changed: measured, and the numbers are in the description. (one arithmetic change in TryFromSerial; reasoned above, not benchmarked)
• If an error code was added or changed: ErrorCodes and the guide's error-code table agree. (no new code; the table names xlsx's causes)
• Public API changes are described, and breaking ones are called out. (additive: TabularFormat.Xlsx accepted by Create, TabularWriterOptions.Xlsx, XlsxWriterOptions)
• No third-party package in src/.