stillwater-sc / universal / 37856086365

86%

master: 84%

push

github

fix(cfloat): reject assign(string) fields of the wrong width (#1648)

cfloat::assign and bfloat16::assign checked the total bit count and the
exponent width of a "0b<sign>.<exponent>.<fraction>" string, but not the
sign width. A misplaced first '.' was accepted silently: bfloat16
"0b01.00000000.000000" became 2.0, and cfloat<8,2> "0b01.10.0000" became
NaN. Both parsers now measure all three field widths before writing any
bits and reject anything but 1.es.fbits, leaving the value at zero with
a diagnostic that names the widths found and needed.

bfloat16's to_binary printed positive values with a "0x0." prefix
instead of "0b0.", so its output could not be assigned back. The prefix
is fixed. bfloat16's diagnostics also named cfloat<>.

Tests: cfloat_assign and bfloat16_assign check that every encoding round
trips through to_binary and assign (exhaustive at 8 and 16 bits,
sampled at 32). They check that each malformed field width is rejected
with the total bit count right, and that ' digit separators work. Run
against the old headers, the field-width and round-trip checks fail. All
122 cfloat and bfloat16 tests pass.

Fixes #1647

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

22 of 22 new or added lines in 3 files covered. (100.0%)

47138 of 54596 relevant lines covered (86.34%)

9996588.05 hits per line