0070 - Shared string-to-number conversion

Status

Accepted.

Context

ECMAScript StringToNumber (ECMA-262 §7.1.4.1.1) was implemented twice. Runtime primitive coercion lived in TGocciaStringLiteralValue.ToNumberLiteral (Goccia.Values.Primitives), and compile-time constant folding had its own StringToCompileTimeNumber (Goccia.Compiler.ConstantValue). Each kept a partial copy of the StringNumericLiteral grammar, so the interpreter and the bytecode constant optimizer (ADR 0040) could fold the same string source to different values. A +Infinity divergence had already been fixed in one path but not the other, and both copies shared latent gaps: binary/octal prefixes returned NaN, large hex wrapped through a fixed-width StrToInt, and TryStrToFloat over-accepted inputs such as "Inf".

The compiler must reuse the parser without instantiating a TGocciaValue, so the shared implementation cannot return a runtime value object.

Decision

Introduce source/shared/NumericText.pas with a single side-effect-free StringToNumber(const AText: string): Double, and route both paths through it: ToNumberLiteral wraps the result with TGocciaNumberLiteralValue.Create, and TryCompileTimeValueToNumber calls it directly. The helper implements the StringNumericLiteral grammar in one place — ECMAScript whitespace trimming, empty → +0, signed Infinity, NonDecimalIntegerLiteral (0b/0o/0x, no sign, digits accumulated into a Double so large magnitudes round instead of wrapping), a grammar-gated StrDecimalLiteral (overflow → ±Infinity, signed zero preserved), and NaN for everything else.

The helper lives in source/shared/ rather than in Goccia.Values.Primitives or TextSemantics:

• It returns a raw Double and depends only on text, so it stays free of the GocciaScript value types — the compiler reuses it with no allocation.
• A dedicated unit gives textual numeric conversion a clear home with room to absorb related parsing (parseInt, parseFloat) later, rather than growing TextSemantics (UTF/whitespace semantics) or Goccia.Values.Primitives (value classes) with an unrelated concern.

No bytecode format or opcode changes; the compiler optimizer design from ADR 0040 is unchanged.

Consequences

Runtime Number(...)/unary + and bytecode constant folding now share one parser and cannot drift. Fixing the parser once also closed the shared conformance gaps: binary/octal literals, full-magnitude hex, rejection of "Inf" and bare punctuation, and sign rejection on non-decimal literals are now correct in both execution modes. Future string↔number conversions should extend NumericText instead of adding a second parser.