math: add f32 (binary32) storage — the float sibling of f64

[NiceAndPeter]

Phase 4b (part 2b) — f32 (binary32) storage · the final plan item

A bound carrying f32 holds its value in a binary32 raw (float) — the single-precision sibling of f64, for single-precision FPUs and the flt engine. The natural pairing flt compute + f32 storage keeps the whole path in hardware float with no double round-trip at the I/O boundary (this closes the soft-double cost discussed during review).

Approach: fp_raw unification (not a parallel float path)

fp_raw = f64_raw || f32_raw. Read/store/compare/arithmetic compute in double and narrow to the raw type on store — lossless because the grid is fp-exact. So f32 rides the proven f64 machinery instead of duplicating it.

• policy_flag: new f32 (bundles round_nearest); widest-wins exact > f64 > f32 > direct > indexed.
• grid.hpp: float_exact<G> (24-bit significand, exp ≥ −126); storage_pick returns float{} for an f32 dyadic float-exact grid (static_asserts on direct misuse, like f64).
• generic.hpp: f32_raw/fp_raw; value_raw/index_raw exclude fp_raw; as_double/operator rational/sentinel_raw were already fp-generic.
• core.hpp: store_f64→store_fp (narrows the double snap to the raw type); value-path branches keyed on fp_raw; operator double implicit for f32; f32 dyadic guard.
• arithmetic (add/mul/div): rep propagation gains f32 with demotion — f32 stays only when both operands are f32-only and the result fits float, else widens to f64, else drops to exact; bodies snap-and-narrow via fp_raw.
• cmath: dbl::store/flt::store store straight into any fp-backed Out (f32 results no longer detour through the rational path); fast-path guards exclude fp_raw.
• io/assignment: f64_raw→fp_raw so f32 prints/assigns like f64.

Tests / docs

test_storage_flags.cpp: f32 storage selection, lossless construct/read, arithmetic stays f32 then demotes f32→f64 when a result grid outgrows float, widest-wins, and math output landing in f32. Docs: policies.md (f32 row + widest-wins) and math.md (the flt+f32 pairing + the 24-bit overflow note).

(First commit is the mechanical real_raw→f64_raw internal rename; second adds f32.)

Verification

default 407/407, CORDIC 443/443 (f32 elided → integer under FIXED), FLOAT 398/398, all single-header smokes build.

🤖 Generated with Claude Code