lib/math: re-fold F64_EXPBIAS to const f64info/f32info struct (#40)

Restore the Hare-faithful floatinfo struct defs (ref/hare/math/floats.ha
:117,126), removing the Drew-flattened-(b) #129-era bypass (F64_EXPBIAS:int
scalar). #149 (db7523e) now lowers cross-module &def to LEAQ, so the struct
exports are addressable via the stof consumer pattern (&math.f64info ->
f: *floatinfo -> f.expbias). First real consumer of A.2 struct-composite
static-init (DATA byte-validated: f64info/f32info 40B each).

ADD export def f64info/f32info: floatinfo (hex masks value-identical to
Hare's (1<<52)-1 etc.; A.2 helper folds bare literals only, documented
at-site). DELETE export def F64_EXPBIAS (zero consumers). KEEP NAN_BITS/
INF_BITS u64 sentinels (ruling b: def NAN=0.0/0.0 blocked by #147) +
F64_EXPONENT_BIAS:u64 (bit-ops alias, Hare keeps both).

Test 952: pointer-param row reads &math.f64info + &math.f32info through
*math.floatinfo, all 5 fields each (pointer-param not direct field-read,
dodges #150). Make test 184/184 incl 990-997 byte-id + combined_ww_fresh.
lib/math not compiler-imported -> no regen.

Unblocks fold-4 stof.ha.

lib/math/floats.ww

@@ -3,8 +3,9 @@
 // ref/hare/math/floats.ha (fold-1: classify/sign/bits; fold-2a:
 // issubnormalf64/normalizef64/frexpf64; strconv-foundation fold-1a:
 // F32 bit-layout + f32bits/f32frombits + floatinfo struct type;
-// fold-1b: NAN_BITS/INF_BITS sentinels + F64_EXPBIAS (flattened from
-// f64info per #129)). frexpf64's zero guard `n == 0f64` rides the #103
+// fold-1b: NAN_BITS/INF_BITS sentinels; γ-cleanup: f64info/f32info
+// instances re-folded once #149 lowered &math.f64info). frexpf64's
+// zero guard `n == 0f64` rides the #103
 // fix (no-decimal f64 literal now materialized into XMM) and its
 // (f64, i64) tuple return rides the #105 fix (tuple f64-word read).
 // The ldexp/modfrac/nextafter family stays deferred (need f64 DIVIDE).
@@ -78,15 +79,6 @@ def F64_EXP_REMOVAL_MASK: u64 = ~(F64_EXPONENT_MASK << F64_MANTISSA_BITS);
 // ref/hare/math/floats.ha:84
 def F64_EXP_ZERO: u64 = (F64_EXPONENT_BIAS - 1) << F64_MANTISSA_BITS;
 
-// flattened from f64info per #129 (cgen module-level struct-init gap);
-// cite ref/hare/math/floats.ha:117. Re-fold to const f64info: floatinfo
-// when #129 closes (γ-cleanup pattern per amalloc-drop precedent).
-// The existing F64_EXPONENT_BIAS:u64 above is the bit-ops alias; this
-// int-typed F64_EXPBIAS is the floatinfo.expbias-field counterpart and
-// drops the cast at the construction site (see ref/hare/strconv/stof.ha
-// :248,288 for the consumer pattern).
-export def F64_EXPBIAS: int = 1023;
-
 // F32 bit-structure constants. ref/hare/math/floats.ha:27,30,33 declare
 // these as untyped int; ww has no untyped def, so they ride u32 (matching
 // the u32 bit container, the same way the F64 family rides u64 — see
@@ -134,13 +126,6 @@ def F32_EXP_ZERO: u32 = (F32_EXPONENT_BIAS - 1u32) << F32_MANTISSA_BITS;
 // expbias field stays `int` — that keeps the fold-4 stof port byte-for-byte
 // against ref/hare/strconv/stof.ha:248,288 (`let e: int = 0` arithmetic
 // against `f.expbias` of the same type, no cast at use site).
-//
-// The companion `f64info` / `f32info` module-scope instances Hare exports
-// (ref/hare/math/floats.ha:117,126) are DEFERRED — ww's cgen doesn't
-// lower module-level struct lets-with-initializer (the symbol comes out
-// undefined at link time). Callers in the fold-4 stof port construct a
-// stack-local `floatinfo { ... }` and pass `&info` until that gap closes.
-// Repro logged with team lead.
 export type floatinfo = struct {
 	// Bits in significand.
 	mantbits: u64,
@@ -154,6 +139,31 @@ export type floatinfo = struct {
 	expmask: u64,
 };
 
+// floatinfo instances for the f64 / f32 types, consumed by the
+// width-generic strconv helpers via &math.f64info (cross-module
+// address-of, lowered since #149). ref/hare/math/floats.ha:117,126.
+// Hare spells the masks (1 << 52) - 1 / (1 << 23) - 1; the #129 A.2
+// struct-composite static-init path folds only bare-literal field
+// initializers, not const-fold expressions, so the value-identical hex
+// literals are used here (0xFFFFFFFFFFFFF == (1<<52)-1, 0x7FFFFF ==
+// (1<<23)-1 — same hex-literal style as the NAN_BITS/INF_BITS sentinels
+// below). expbias rides `int` (the field type) with no suffix.
+export def f64info: floatinfo = floatinfo {
+	mantbits = 52u64,
+	expbits = 11u64,
+	expbias = 1023,
+	mantmask = 0xFFFFFFFFFFFFFu64,
+	expmask = 0x7FFu64,
+};
+
+export def f32info: floatinfo = floatinfo {
+	mantbits = 23u64,
+	expbits = 8u64,
+	expbias = 127,
+	mantmask = 0x7FFFFFu64,
+	expmask = 0xFFu64,
+};
+
 // IEEE-754 quiet-NaN and positive-Infinity f64 bit sentinels.
 // ref/hare/math/floats.ha:137,141. Hare exports `def NAN = 0.0/0.0;` and
 // `def INF = 1.0/0.0;` (untyped float def-fold); ww's cgen doesn't lower