fix: parse YAML numeric literals invariantly [patch]

[matt-edmondson]

Fixes #73

What was wrong

YamlSerializer writes numeric literals through YamlDotNet's scalar emitter, which is always invariant and dot-decimal (3.14). YamlDeserializer read them back with the TryParse overloads that take no culture, so they parsed against Thread.CurrentThread.CurrentCulture.

double.TryParse(string, out double) defaults to NumberStyles.Float | AllowThousands. On a culture where . is the group separator and , the decimal separator — de-DE, fr-FR, es-ES, it-IT, ru-RU, pl-PL, nl-NL and the rest of continental Europe — "3.14" parses as 314. Not an error, not a dropped node: a silently wrong value in the loaded AST, so round-tripping a .coder.yaml file changed the behaviour of the program it described.

The change

NumberStyles + CultureInfo.InvariantCulture at the two numeric parse sites in YamlDeserializer.cs:

• DeserializeLeafNode — Int32
• DeserializeLiteralExpression — Int32 and Double

This matches the pattern the rest of the codebase already uses for source values (LanguageGeneratorBase.cs:125,133, CSharpGenerator.cs:100,106, AstFields.AssignInteger).

bool.TryParse is deliberately left alone. The issue suggested passing a culture there too, but bool.TryParse accepts only "true"/"false" (case- and whitespace-insensitive) and has no culture-sensitive overload — it is already invariant by construction. The remaining TryParse calls in the file are Enum.TryParse, which is likewise culture-independent.

Test

New Coder.Test/Serialization/CultureInvariantDeserializationTests.cs, covering the fraction round trip the issue's acceptance criteria asks for, plus negative fractions, a fraction nested in a function body, and integer/boolean contract guards.

Two notes on how it is built:

• The culture is constructed, not named. The test project runs in globalization-invariant mode, where new CultureInfo("de-DE") throws CultureNotFoundException. The test instead clones InvariantCulture and sets NumberDecimalSeparator = "," / NumberGroupSeparator = ".", which is the exact pairing that causes the corruption. That also keeps the test hermetic — it pins a fixed number format rather than whatever ICU data the host carries.
• The culture never touches the test host's threads. The round trip runs on a LongRunning task, so the mutated culture dies with its dedicated thread instead of being inherited by a pool thread later.

Verification

Reverting YamlDeserializer.cs while keeping the new test file:

failed DeserializeFractionLiteral_UnderDotGroupingCulture_RoundTripsValue
  Assert.AreEqual(3.14, roundTripped.Value)
failed DeserializeNegativeFractionLiteral_UnderDotGroupingCulture_RoundTripsValue
  Assert.AreEqual(-0.125, roundTripped.Value)
failed DeserializeFunctionBody_UnderDotGroupingCulture_PreservesReturnedFraction
  Assert.AreEqual(0.5, value.Value)
  total: 6, failed: 3, succeeded: 3

With the fix restored: 6/6 pass, and the full suite is 867/867 green. dotnet build Coder.sln succeeds with no warnings across both target frameworks.

To be straight about coverage: only the three double cases demonstrably regress today. The Int32 and Boolean tests pass either way — int.TryParse is not currently broken by any culture .NET ships, since it special-cases the ASCII hyphen even where the culture's negative sign differs. They are kept as contract guards for the invariant round trip, not presented as reproductions.

🤖 Generated with Claude Code

https://claude.ai/code/session_01B1XQbfXNGfPRfGr4bRjsT3

---

Generated by Claude Code

[sonarqubecloud]

Quality Gate passed

Issues
0 New issues
0 Accepted issues

Measures
0 Security Hotspots
100.0% Coverage on New Code
0.0% Duplication on New Code

See analysis details on SonarQube Cloud