[GabrielMarquezMatte]

What does this change?

Everything on develop since v6.0.0, as the candidate for 7.0.0. It started as 6.1 (own byte
sources, managed inflate, native column buffers) and became a major release once parallel reading
replaced CsvParallel and the public API was reworked so that new settings can be added without
another break. The full v6 → v7 port is in docs/guide/migrating-to-v7.md.

The C ABI keeps every signature and XL_ABI_VERSION stays 6; the native behavior changes are listed below.

Parallel sheet reading (SheetParallel, breaking)

• SheetParallel replaces CsvParallel: one static class with ParseAsync, ForEachAsync and
  AggregateAsync over any IExcelSheet. You open the workbook and pass the sheet, and the library
  parallelises as far as the format allows. The contracts lost their Csv prefix (IRowRecord,
  IRowAccumulator, RowAggregation, RowAction).
• XLSX and XLSB now read on several threads. One thread inflates and cuts row-aligned 256 KB chunks,
  workers parse them in place, and an ordered merge yields rows in sheet order. When an XLSX chunk ends
  mid-markup, that part is read again sequentially, so the results always match a sequential read.
  XLS, table views and small sheets stay sequential.
• Sheet stage on the 65K corpus (Ryzen 7 5700X, JIT harness minimum): XLSX 17.6 ms, XLSB 13.3 ms.
  NativeAOT xl_parse_typed, parallel vs sequential: XLSX 20.4 vs 53.4 ms, XLSB 15.2 vs 27.9 ms.
• Native: degree_of_parallelism = 0 on xl_parse_typed and the Arrow exports now parallelises XLSX and
  XLSB as well as CSV. The bindings need no code change, but peak memory is higher because partitions are
  held until they merge.

Excel tables

• Read: IExcelWorkbook.Tables and TryGetTable on every format, and ExcelTable.AsSheet() to read a
  table's rows through the normal sheet API, with column indexes relative to the table. Name binding on a
  headerless table is rejected with InvalidOperationException.
• Write: BeginTable/EndTable on the XLSX and XLSB sheet writers (ITableSheetWriter<TRow>),
  ExcelTableOptions, and WriteTableAsync to write a record layout as a table.

Open a workbook from your own bytes

• Core: any byte source can be read through a seekable view that fetches it in blocks, each block once.
• C ABI: xl_open_source (random access through read_at), xl_open_stream (forward-only) and
  xl_set_source_error. xl_open_options gains source and buffering fields and still accepts the 6.0
  layout through struct_size.
• Bindings: Python takes a Source or any readable stream, C++ an xl::Source or xl::InputStream,
  and Rust a Source or any Read.
• Schema inference and replay work on a CSV stream that cannot seek. The Python Arrow stream releases the
  GIL during native callbacks.

Managed inflate decoder

DeflateStream is replaced in ZipIndex by InflateStream, a managed, resumable port of libdeflate's
design (the MIT notice is in THIRD-PARTY-NOTICES.md, which now ships in the package). Inflate alone is
about 1.7x faster on XLSX and 1.4x on XLSB. It has a canary after the output window, exact truncation
detection, and an inflate fuzz target that compares its output with DeflateStream.

Typed parse builds columns in native memory

ColumnBuilder writes into a native block that grows by doubling and hands it over without a copy.
Peak memory in to_record_batch on the 65K fixture fell from 26.0 to 18.1 MiB (whole sheet) and from
16.7 to 10.7 MiB (streamed). Typed parsing is 3–5% faster.

Public API changes (breaking)

v6	v7
CsvParallel and its contracts	SheetParallel, IRowRecord, IRowAccumulator, RowAggregation, RowAction
ExcelParserConfig	ExcelParserOptions
Converters and cell readers take (bool isDate1904, IFormatProvider provider)	They take in ExcelRowContext
InferSchema(sheet, headerRow, sampleSize, parseText) returning an array	InferSchema(sheet, SchemaInferenceOptions) returning IReadOnlyList<ExcelColumnSchema>
ExcelDataReader(sheet, headerRow)	SheetDataReader(sheet, SheetDataReaderOptions) (the old name collided with Sylvan.Data.Excel and the ExcelDataReader package)
ToArrowRecordBatch(schema, headerRow), WriteRecordBatch(batch, sheetName, writeHeader)	ArrowReadOptions, ArrowWriteOptions
XlsWorkbookWriter.Create(stream, leaveOpen, date1904)	XlsWriterOptions; XlsxWriterOptions also gains Date1904
Excel.FromCsv*(…, CsvReaderOptions)	Excel.FromCsv*(…, ExcelReaderOptions), with the dialect in .Csv
CsvReaderOptions.MaxCellBytes, .InternStrings	On ExcelReaderOptions only
Cell.TryGetDateTime(out DateTime) (silently assumed 1900)	Removed; pass isDate1904
CellType.Formula	Removed; a formula cell reports its cached result's type

Behavior changes worth a release-note line

• A ZIP entry that is cut short or corrupt throws InvalidDataException. Before, a sheet could silently
  end early (one damaged workbook stopped at 298 of 30,000 rows). Both the managed inflate and the new
  declared-size check on streamed entries catch it. CRC32 is still not verified.
• XLSX text now decodes OOXML _xHHHH_ escapes (Excel writes _x000D_ for CR). Before, text written by
  this library's own writer did not read back unchanged.
• Formula cells are typed by their cached result in every format. Before, an XLS formula with a date
  style lost its Date type. XL_CELL_FORMULA stays defined but is no longer produced.
• Excel.Open with ExcelFileFormat.Csv now applies MaxCellBytes and InternStrings.
• The CLI convert command no longer shifts dates by four years when converting a 1904 workbook.
• xl_write_options.date1904 now applies to XLSX.

Other fixes

• XLSX rows whose opening tag straddled an async buffer refill were dropped.
• The generator sets required members in the factory it emits.
• Table writer fixes: column names are escaped like header cells, a header-only table keeps its reserved
  row out of the range, and stray AutoFilter and sort state are no longer written to XLSB.
• HexVal accepted bytes below '0' and 'a' as hex digits, which also affected &#x…; entities.
• The sheet pipeline surfaces the source stream's own cancellation and reports a parse failure at the
  right sheet row after recovery.

Also

• The static PGO profile was regenerated after each signature change.
• Project icon on the NuGet packages and the binding READMEs.
• Dependabot bumps.

Checklist

• dotnet build ExcelReader.slnx --configuration Release builds clean (warnings are errors)
• dotnet test tests/ExcelReader.Tests/ExcelReader.Tests.csproj --configuration Release passes (full suite)
• Public API changes recorded in PublicAPI.Shipped.txt/PublicAPI.Unshipped.txt
• Tests added/updated for every behavior change
• Migration guide (docs/guide/migrating-to-v7.md)
• One focused change: no, this is the release PR from develop

Test plan

• Affected suites pass with 0 failures: Reader 1,767, Writer 508, Native 395, Parser 359, Arrow 27,
  CLI 26.
• Sheet pipeline: parity tests against the sequential read for every format, plus a fuzz target comparing
  the two.
• Inflate: rejection-path unit tests, differential tests against DeflateStream, and fuzz check
  (156,780 cases, 0 failures).
• Corrupt-entry and _xHHHH_ tests were written failing first; the escape decoder passed 25,186 fuzz
  cases.

Before tagging 7.0.0:

• BenchmarkDotNet rerun on an idle Ryzen 7 5700X. The parallel tables in benchmarks.md were taken
  under load and are marked pending.
• Promote PublicAPI.Unshipped.txt into Shipped.
• Python, Rust and C++ suites against the republished NativeAOT library.
• A few CI fuzz cycles on the pipeline and inflate targets.