v0.105.0

Changed

• Serializer pattern emission refactored to classification-based dispatch (REFACTOR-SERIALIZER-PATTERN-001):
  • Replaced three interleaved methods (_serialize_pattern, _serialize_text_element, _pattern_needs_separate_line) with a unified continuation line classification model
  • New _LineKind enum (EMPTY, WHITESPACE_ONLY, SYNTAX_LEADING, NORMAL) classifies each continuation line exactly once; a single match/case dispatch point selects the correct wrapping strategy
  • _classify_line() pure function and _escape_text() (brace-only escaping) extracted as module-level functions, enabling independent testing and reuse
  • _emit_classified_line() static method serves as the single dispatch point for all continuation line ambiguity classes
  • All match/case dispatches now use assert_never() exhaustiveness guards (_serialize_entry, _serialize_expression, _emit_classified_line) to catch missing cases at type-check time when new AST variants are added
  • _serialize_expression uses full structural destructuring in patterns (StringLiteral(value=value), TermReference(id=Identifier(name=name), ...)) instead of body-level attribute access
  • No public API changes; serialize() and FluentSerializer.serialize() semantics preserved
  • Location: syntax/serializer.py

Fixed

• Serializer roundtrip idempotence failure on SYNTAX_LEADING continuation lines with content whitespace (BUG-SERIALIZER-WS-SYNTAX-002):

  • Same bug class as BUG-SERIALIZER-WS-SYNTAX-001 (fixed in 0.104.0): syntactically ambiguous whitespace in serializer output; residual case missed by the original fix and inherited by the 0.105.0 classification refactor
  • When a continuation line had content spaces preceding a syntax character (., *, [), e.g., "dS7aQ\n .h?Q", the SYNTAX_LEADING emission branch wrapped the syntax character in a StringLiteral placeable ({ "." }) but emitted the preceding content spaces as raw text
  • The FTL parser computes common indent as the minimum indent across all continuation lines; structural indent (4 spaces) plus raw content spaces (e.g., 6 spaces = 10 total) became the common indent, and all 10 spaces were stripped during common-indent removal, losing the 6 content spaces
  • Re-serializing the parsed AST produced only the 4-space structural indent before the placeable, breaking idempotence: S(AST) != S(P(S(AST)))
  • Fix: _emit_classified_line SYNTAX_LEADING branch now wraps leading content spaces in a StringLiteral placeable ({ " " }{ "." }...), matching the treatment already used by WHITESPACE_ONLY and leading-whitespace-after-=; the parser sees the spaces as expression content (opaque to indent stripping), preserving them through roundtrip
  • Invariant enforced: all content whitespace preceding the first non-whitespace character on a continuation line must be placeable-wrapped; raw spaces are indistinguishable from structural indent
  • Location: syntax/serializer.py _emit_classified_line()

• Diagnostic.__str__ returns raw dataclass repr instead of human-readable message (BUG-DIAGNOSTIC-STR-001):

  • Diagnostic (frozen dataclass) had no __str__ method; str(diagnostic) returned the full repr() with all fields including None values, e.g., Diagnostic(code=<DiagnosticCode.VARIABLE_NOT_PROVIDED: 1005>, message="...", span=None, ...)
  • All FrozenFluentError construction sites use str(diag) as the message parameter, so error.message and str(error) both returned the raw repr
  • Users calling str(error) in logging, error display, or exception handlers saw opaque dataclass output instead of the human-readable message
  • Fix: Added __str__ to Diagnostic returning self.message; repr() unchanged for debugging
  • Location: diagnostics/codes.py Diagnostic.__str__()