API-COMPLETENESS-T27 · GF4/GF8/GF12/GF20/GF24 имеют только extract_*, нет encode/decode/arithmetic

[gHashTag]

API-COMPLETENESS-T27 · GF4/GF8/GF12/GF20/GF24 имеют только extract_*, нет from_f32/to_f32/arithmetic

File: ffi/src/lib.rs
Severity: P1 — partial paperware (formats declared but not usable)

Current API surface

Format	extract_sign	extract_exp	extract_mant	from_f32	to_f32	arithmetic
GF4	✅	✅	✅	❌	❌	❌
GF8	✅	✅	✅	❌	❌	❌
GF12	✅	✅	✅	❌	❌	❌
GF16	✅	✅	✅	✅	✅	✅
GF20	✅	✅	✅	❌	❌	❌
GF24	✅	✅	✅	❌	❌	❌
GF32	✅	✅	✅	✅*	✅*	✅

* GF32 is IEEE-fp32 layout (see BUG-003-T27) — not whitepaper Lucas-18.

Whitepaper claim

whitepaper §1.2:

Golden Float family: GF4, GF8, GF12, GF16, GF20, GF24, GF32

But only GF16 + (broken) GF32 can encode/decode/compute. The other 5 formats are introspection-only.

Required for honest publication

Either:

• (a) Implement encode/decode/arithmetic for GF4/8/12/20/24 — full family
• (b) Retract from whitepaper as «GF16 + GF32 only; smaller/larger formats reserved for future»

Decomposition (if path a)

Format	Layout	Effort
GF4	1:1:2 (sign+exp+mant)	30 мин
GF8	1:3:4 (whitepaper)	30 мин
GF12	1:4:7	30 мин
GF20	1:7:12	30 мин
GF24	1:9:14	30 мин

Plus arithmetic (add/sub/mul/div) per format: × 5 formats × 4 ops = 20 functions.

Acceptance

For each missing format:

• gfN_from_f32(x: f32) -> uN — encode via integer bit manipulation
• gfN_to_f32(value: uN) -> f32 — decode
• gfN_add/sub/mul/div(a, b) — arithmetic via integer ops
• Round-trip tests for 0, ±0, 1.0, π, NaN, Inf
• ULP @ |x|=1 verified against theoretical 2^-mant_bits

Cross-references

• BENCH-010 (zig-golden-float#23) currently tests only fp32, fp16, bf16, gf16, ternary — extending to GF4/8/12/20/24 requires this issue resolved
• Whitepaper §1.2 family table claim [SEED-2] Ring-2: Type declarations — enum/struct gen to Zig #5 depends

phi^2 + phi^-2 = 3 · TRINITY · NEVER STOP