C float literal in the subnormal range parses to 0.0

• Track: C (cfront literal parsing; possibly shared with crtl strtod).
• Found 2026-07-19 during the crtl exact-printf differential sweep (the one residual diff in 4000 random doubles + specials vs glibc).

Repro

double d = 4.9406564584124654e-324;   /* DBL_TRUE_MIN */
/* pxx: bits 0000000000000000 (0.0); gcc: 0000000000000001 */

Min NORMAL (2.2250738585072014e-308) parses exactly right — only the subnormal range collapses to zero. Compile-time conversion, not printf (printf digits are exact since b376). Likely the literal-parse scale steps underflow to 0 before the mantissa is applied. Related family: [[bug-crtl-strtod-precision-cjson-floats]] (runtime strtod may share the same path — check both). Severity low: JS/real code rarely spells subnormal literals, and runtime arithmetic that PRODUCES subnormals is unaffected.

2026-07-22: FIXED (flush-to-zero); 1-ulp residual folded into the strtod ticket

StrToDoubleBits's bexp < -1022 -> 0 early exit replaced with proper denormalization (shift the 53-bit significand right with round-to-nearest-even, sticky from the extraction remainder; carry lands on min normal). DBL_TRUE_MIN, mid subnormals, and the min-normal boundary are bit-exact vs gcc; a 400-literal random differential shows 388/400 exact, 12 within 1 ulp near the top of the subnormal range — the same accumulated round-to-odd scaling error class the normal range already carries for extreme exponents (the ~300-step decimal scaling costs ~2^-52 relative). Correct rounding for those needs a big-int scaling pass — tracked with the runtime twin [[bug-crtl-strtod-precision-cjson-floats]]. Regression test test/csubnormal_literal.c in make test. Benefits Pascal literals too (shared lexer routine).

Log

• 2026-07-22 — resolved, commit 8fc9508d.