Test suite gaps found in the pre-release review

[vaceslav]

From the pre-release review. Coverage today: 91.3 % lines, 83.8 % branches (Analysis 98.6/93, Csv 94/92, Extraction 96/91, Import 91/86, Mapping 87/82, Xlsx 86.5/78, Abstractions 81/58).

Before the release

• The two net8-failing cancellation tests (StopsInsideARowFullOfElementsItWantsNothingFrom, StopsInsideAValueAssembledFromCountlessEmptyRuns) arm on compressed package bytes, while the sheet deflates to a few hundred bytes — net8 and net10 compress the fixture differently, so the runtime's zlib decides the outcome, not the reader. Use incompressible content and assert on the cursor being poisoned. Unblocks Decide which target frameworks to support #4.
• Typed workbook cells never go through extraction/import successfully: number → Integer/Decimal, serial → Date, t="b" → Boolean, t="e" error cells (ValueReading native paths uncovered). Every importer test is csv today.
• Public API contracts: leaveOpen on every entry point, double Dispose, use after dispose (ObjectDisposedException), non-seekable streams into TabularFile/CsvCursor, cursor disposal when Import(Stream…) fails, TranslatedField argument validation, InChunks(0).
• Guide promises with no test: MaxMetadataChars bound, "rows are views", AllOrNothing at import time (Correctness fixes found in the pre-release review #14), header row beyond the sheet's end, precheck Undetermined for an unprofiled culture and for exhausted uniqueness budget. One date-precheck test passes without reaching the code its name claims — fix its setup.
• Invariant globalization — see Correctness fixes found in the pre-release review #14; add a CI leg or a test that runs the core flow with InvariantGlobalization.
• Weakest classes worth a test each: SheetScanner CDATA/comments, XlsxCursor.LooksLikeDateFormat (5 of 28 conditions covered), MappedValue/RawCell equality, Windows1252Encoding 0x80–0x9F range.

After the release

• Rename tests grouped by review round (ReviewFixesTests, SecondRoundFixesTests) into behaviour-named classes; dedupe copied helpers.
• Property-based / fuzz tests: csv write→read round trip with random quoting and line endings; random xlsx generation through XlsxPackage.
• A Stryker.NET mutation pass — expected survivors in date-format detection, ImportField fluent methods, the Windows-1252 table.
• Flakiness candidates: allocation-ratio tests without warm-up, tests that walk the repository (obj/ layout), deadline-based tests.
• Producer diversity in golden fixtures (LibreOffice, Google Sheets, Numbers, generator libraries) — the differential run in Mine the open-source readers' test suites for edge cases we do not cover #8 is the source.

[vaceslav]

All items in Before the release are done (commits 7acb0dc..644436b). Totals: 588 tests on net8.0 and 588 on net10.0, plus 5 in TriasDev.Tabular.InvariantGlobalizationTests, whose whole host runs with InvariantGlobalization.

Cancellation fixtures, net8. The two cancellation tests that failed on net8 now use incompressible content. A mutation check (cancellation checks removed from the read loops) makes both fail on both runtimes.

Typed workbook cells. The import tests now cover number → Integer/Decimal, a date-styled serial number, t="d", t="b", t="s", t="e", a fraction and 1E+30 into Integer, and 1E+300 into Decimal. The plan is under de-DE, to prove that typed cells are not re-read through the culture.

API contracts. Covered by tests: double Dispose, use after Dispose, a run read only once, forward-only streams, and stream ownership. Fixed along the way:

• InChunks(0) was only checked on first enumeration; it now throws at the call;
• CsvCursor.MoveToSheet now throws after Dispose;
• ImportField.* now refuses empty names.

Guide promises. Now pinned by tests:

• the MaxMetadataChars bound;
• a header row past the end of the sheet;
• an Undetermined finding for a culture the profile was not measured under;
• an Undetermined finding when the uniqueness budget runs out;
• rows are views: an allocation test where one allocation per row fails it;
• the date precheck test now really judges the rebuilt date.

Weakest classes.

• Custom date formats: 20 real format codes.
• Built-in format id edges: 13/14, 22/23, 26/27, 36/37, 44/45, 47/48, 49/50, 58/59.
• SheetScanner comments, CDATA and processing instructions. This found a real bug: a comment inside <v> cut the value short (4<!-- -->2 → 4). It is fixed with a one-character check on the hot path; xlsx read time on 1M rows is unchanged.
• Windows1252Encoding, checked against the platform's code page 1252 on all 256 bytes.
• RawCell / MappedValue equality. MappedValue now compares numbers by value: 1.0 == 1.00.
• The regex timeout path (lookaround fallback plus exponential backtracking), with a mutation check.

The After the release items stay open.

[vaceslav]

Status:

1. Behaviour-named test classes, deduplicated helpers: done in test: behaviour-named classes instead of review rounds; shared stream fixtures #39.
2. Seeded fuzz for csv, xlsx and ods, round trip and mangled input: done in test: seeded fuzz for every format; two reader fixes it found #40, widened in fix: read what real xlsx writers write; producer fixtures from seven writers #42. It found two reader bugs, both fixed in test: seeded fuzz for every format; two reader fixes it found #40.
3. Mutation testing: Stryker.NET 5.0 does not work with this suite yet. Its Microsoft.Testing.Platform runner (preview) gets no answer to its run-tests call from the xunit v4 test server, so every surviving mutant shows up as a timeout, and every area scores a false 100 %. The VSTest runner is too slow to use: the first test run alone took 26 minutes, and coverage capture failed. The one survivor verified by hand, the unpinned null guards, is covered in fix: a valid value cannot fail a pattern on a slow first match; null-argument contract tests #43. Worth retrying when the MTP runner leaves preview.
4. Flaky tests: 30 full-suite runs in a row turned up one failure. It was a real defect, a valid value failing its pattern on a slow first match, fixed in fix: a valid value cannot fail a pattern on a slow first match; null-argument contract tests #43. The quote-storm ratio test now warms up first (fix: read what real xlsx writers write; producer fixtures from seven writers #42). CI history: 68 runs, no re-runs.
5. Producer diversity: fixtures from ClosedXML, the Open XML SDK, EPPlus 4.5, NPOI 2.7, MiniExcel, Sylvan and LibreOffice (xlsx, ods, csv) are in fix: read what real xlsx writers write; producer fixtures from seven writers #42. They found three reader bugs, all fixed there: _xHHHH_ escapes, booleans written as true, literal CR line ends. Google Sheets and Numbers were skipped, since neither can be scripted.