bug: WriteLn(real) default format differs from FPC

• Type: bug (output formatting / FPC-compat) — Track A
• Status: DONE 2026-06-23
• Found: 2026-06-23, differential probe vs FPC
• Severity: low (formatting parity; loud, not a silent miscompile)
• Relation: sibling of bug-writeln-boolean-format.

Resolution (2026-06-23)

Default scientific float format now matches FPC's Double field width: 17 significant digits (d.<16 digits>) and a 3-digit exponent. The mantissa is extracted MSD-first into a 17-digit integer (each step truncates a value in [0,10), with one guard digit rounding half-up), instead of one value * 10^16 double multiply that overflows the 53-bit mantissa and zeroed the low digits. This also fixed two accuracy bugs the old path had: 1.0e-100 printed as 0.0...E+000, and 1.0e100 as 9.99..E+099.

Touched all four code paths so output is byte-identical across targets:

• x86-64 EmitWriteFloatSci (symtab.inc)
• aarch64 EmitWriteFloatSciA64 (ir_codegen_aarch64.inc)
• arm32 + i386 PXXWriteFloatSci (builtin/builtinheap.pas)

Gate: make test green (golden test/test_float_write.pas updated to 17-digit output); self-host byte-identical. Verified x86-64/aarch64/i386/arm32 all emit identical text under qemu.

Residual (intentionally not chased — separate, deeper issues)

1. Bare-literal parity is unreachable. FPC types a bare real literal as the smallest type that holds it exactly: writeln(1.0) → Single (1.000000000E+00), writeln(3.14159) → Extended (3.14158999999999999993E+0000). pxx has only Double, so it always prints the Double form. Parity holds for Double-typed values, which is the meaningful target. (Per-literal type inference + Single/Extended would be a separate feature.)
2. Last 1–2 digits diverge on hard values (e.g. 1234.5 → 1.2345000000000002E+003, large/small exponents). Double-arithmetic scaling cannot reproduce FPC's correctly-rounded last digit at 17 sig figs; that needs a bignum dtoa (Ryu/Grisu/Dragon). The existing code comment already declared this tolerance ("last 1–2 digits may differ … adequate"). Filed as a possible future feature-bignum-float-dtoa if exact parity is ever needed.

Symptom

writeln(3.14159) — default (no width/precision) float formatting:

fpc:  3.14158999999999999993E+0000
pxx:  3.141590000000000E+000

Two divergences:

1. Mantissa precision — FPC emits ~17 significant digits (full double), pxx 16.
2. Exponent width — FPC pads the exponent to 4 digits (E+0000), pxx 3 (E+000).

The explicit fixed form matches (writeln(1.5:0:2) agrees), so only the default scientific path differs.

Expected

Match FPC's default Str(real) formatting: 17 significant digits and a 4-digit exponent field, for output parity with the FPC-seeded toolchain and any golden-output tests.

Repro

begin writeln(3.14159); end.