The f32 coda of Drew's strconv 5-fold plan — f64tos shipped in fold-5a
(0e66073); this completes the plan. Re-lands the f32-exclusive Ryū path
that fold-5a removed under the dead-code rule (it was #143-blocked):
pow5fac32/pow5multiple32/pow2multiple32, mulshift32, mulpow5inv_divpow2/
mulpow5_divpow2, decf32, f32todecf32 (ftos_ryu.ha), and the f32tos driver
(ftos.ha:448). Plus F32_POW5_*_BITCOUNT in ftos_data.ww.
The f32 path REUSES the shared u64 core (mulshiftall64/u128mul/u128rshift/
log*) and f64computeinvpow5/f64computepow5 — and thus the f64 SPLIT2
tables — exactly as ftos_ryu.ha does; there are no separate f32 tables.
Unblocked by #143 (aff7725): f32tos calls math.f32bits(n), passing an f32
arg, which now spills MOVSS (4B) in both stages. Verified: f32tos's arg
push/pop is MOVSS, w6c vs w6c_ww 0-diff on the strconv-embedding
combined.ww (w6c/wwdump/smoke regenerated).
Dodges (cgen bugs still deferred, each cited at-site): decf32.exponent:i64
sidesteps the #169 narrow-second-field struct-return unpack (byte-id
gate-confirmed, not an ABI guarantee); #168 div/mod local-bind on the two
`%10` sites; *decimal pointer field reads (#170); [32]u8 buffer reuses
f64tos's byte-id-clean band over Hare's [14] (#43). mulshift32's U32_MAX
bound inlines the literal — ww's types.U32_MAX is package-private (#172).
Test: ftostest.ww gains f32 vectors — the tcs G/void rows (shared f32/f64
shortest), the f32-exclusive tcsf32 extremes (1e-45 / 1.1754944e-38 /
3.4028235e38, full 24-bit mantissa), specials, a negative-normal, and
33554432 (the sole e2>=0/q<=9 runtime cover). make test 187/187 incl
908_ftos_run + 990-997 byte-id + combined_ww_fresh.