feat(spec): GF-T log2 primitive + cross-entropy (NLL) loss (Refs #1764)

[gHashTag]

Stacked on #1804 (softmax branch). Base retargets to master as the stack merges.

Deliverable C of "все три параллельно" (A = merge queue, infra-blocked; B = softmax synthesizability check, run separately).

gft_log2.t27 — GF-T log2 primitive (inverse of exp2)

log2(x) for a positive GF-T16 → signed GF-T16. log2(x) = (o−40) + log2(1+m/512); the fractional part is a Q Horner quartic, and the integer+fraction real value is normalized fixed→GF-T via a flat 31-step priority encoder (yosys-synthesizable, no while).

Accuracy ≤0.008 abs vs true log2 (output-quantization limited). iverilog 505/505 bit-exact. Spot checks exact: log2(1)=0, log2(2)=1, log2(0.5)=−1, log2(4)=2.

🐞 Bug found + fixed (broken-ruler class): the log2 poly has negative coefficients, and t27 emits >> as a logical shift on signed regs — it filled 0 for negative intermediates instead of arithmetic floor, so the generated Verilog diverged from the Python model. Fixed with an explicit asr9 helper using only non-negative shifts. (exp2's all-positive poly never triggered it.)

gft_nll.t27 — GF-T cross-entropy / NLL loss

For a one-hot label: given the softmax probability p of the true class, returns the loss −log2(p) (composes gft_log2 + sign flip). iverilog 403/403; p=1→0, p=0.5→1.0, p=0.25→2.0 (exact cross-entropy).

---

This opens the path from inference to training on GF-T: forward softmax → prob, then NLL → loss, with the inverse pair exp2/log2 both verified. Fresh seals (seal --verify → MATCH). No compiler change.

Refs #1764

⚠️ GitHub Actions infra incident ongoing. All verification is local (iverilog + seals).

🤖 Generated with Claude Code