[AndreyStartsev]

Summary

Completes the remaining acceptance items for #39:

1. Fixed-timestamp IFC export (--ifc-timestamp / IfcExporter.Timestamp):

   • Adds IfcExporter.Timestamp (DateTimeOffset?) and forwards it to StepWriter.WriteTo so callers can write a deterministic UTC timestamp (yyyy-MM-ddTHH:mm:ss) into the STEP FILE_NAME header.
   • Adds --ifc-timestamp <timestamp> to rvtio (accepting ISO-8601 timestamps and Unix epoch seconds, rejecting non-IFC output paths, duplicate flags, and malformed values with exit code 1).
   • Adds IfcDeterminismTests verifying option parsing, CLI execution, and byte-for-byte identical IFC output across repeated runs for both IfcProfile.Current and IfcProfile.Legacy.

2. CI-runnable synthetic .rvt corruption & fuzz tests with time and memory bounds:

   • Adds CorruptionFuzzTests over synthetic .rvt containers (Global/ElemTable, Global/History, Formats/Latest, Partitions/*), testing container truncations, CFB header/FAT/directory mutations, zeroed byte ranges, uint.MaxValue length overwrites, and inner-stream payload corruptions while enforcing per-iteration and total execution-time bounds, per-iteration allocation caps, and retained GC heap bounds.
   • Extends PartitionFramingFuzzTests with an unconditional [Fact] over synthetic multi-block partition streams enforcing both elapsed-time and thread-allocation bounds.

3. Cross-platform CI (windows-latest, ubuntu-latest, macos-latest):

   • Scales the pole Jacobian degeneracy threshold in RevolvedMesher.NormalAt by the profile tangent norm (1e-7 * Math.Max(puNorm, 1.0)) so 1-ULP trigonometric differences at sphere poles fall through to the pole limit normal on Linux and macOS as on Windows (MeshRevolvedTests).
   • Resolves existing temp-directory symlinks (such as /var -> /private/var on macOS) before testing ExportReport.RealPath on non-existent child path components, and resolves the active dotnet host path via Environment.ProcessPath in CliExitCodeTests.
   • Expands .github/workflows/ci.yml to run full restore, build, test, and pack across windows-latest, ubuntu-latest, and macos-latest.

Verification

• dotnet build RvtIO.slnx -c Release --no-restore (0 warnings, 0 errors)
• dotnet test RvtIO.slnx -c Release --no-build across net8.0 and net10.0 (0 failures)
• dotnet pack RvtIO.slnx -c Release --no-build
• dotnet docfx docfx.json --warningsAsErrors (0 warnings, 0 errors)

Closes #39

[mahach666]

The Windows and Ubuntu jobs passed, but the macOS job fails three test cases on each of net8.0 and net10.0. The three-platform acceptance criterion of #39 is therefore not met yet. Please resolve the platform-specific assertions or numerical behavior while preserving the Windows reference coverage, keep all three OS jobs enabled, and rerun the full matrix. Details and the failing job are in the inline thread.

[mahach666]

[P2] Resolve the macOS test failures before merging the expanded matrix

The approved macOS job fails the same three cases on both target frameworks:

• CompactValueFormatterTests.DegenerateAccuraciesPrintWhatPrints: at accuracy = 0, value = 1234.5678 expects -NaN m² but gets NaN m²; value = -0.0929 expects NaN m² but gets -NaN m².
• UnitDisplayEvaluatorTests.SlopeCloneClearsDimensionsAndUsesNativeNearVerticalSnapIncludingNegative: the near-vertical conversion differs by one bit (4663021752216772328 expected, 4663021752216772329 actual).

These assertions and their formatting/conversion code predate this PR; the new matrix exposes the platform assumptions. Nevertheless, its macOS job remains red and #39 explicitly requires tests to pass on all three operating systems. Resolve these differences without removing the macOS job or weakening the Windows reference checks, then rerun the full matrix. This run did not show a mesher regression or a fuzz timing failure.

[mahach666]

@AndreyStartsev could you fix this?