feat(spec): GF-T softmax+cross-entropy gradient — the backward pass (Refs #1764)

[gHashTag]

Stacked on #1805. Base retargets to master as the stack merges.

gft_softmax_grad4.t27 — the softmax + cross-entropy gradient

For a one-hot target t, the classic closed form:

∂L/∂l_i = p_i − y_i

So the entire backward pass is the softmax forward p_i minus the one-hot label — grad = (i==t) ? sadd(p, −1.0) : p. Composes the verified softmax (max, sadd+neg, exp2, recip, RNE mul).

• Bit-exact to the integer oracle 1600/1600 (iverilog).
• Gradient accuracy ≤0.0018 abs vs true p_i − y_i.
• Spot: uniform logits, target 0 → grad₀ = 0.25 − 1 = −0.75 exact, grad₁ = 0.25 exact.

---

Forward + backward now complete on GF-T: logits → softmax → prob → NLL loss, and → gradient p−y. With the exp2/log2 inverse pair (#1805), the per-sample training step — forward loss and the backprop signal into the logits — is fully expressible spec-first and iverilog-verified. Foundation for on-device training, not just inference.

Fresh seal (seal --verify → MATCH). No compiler change.

Refs #1764

⚠️ GitHub Actions runner availability still degraded (jobs queue). Verified locally (iverilog + seals).

🤖 Generated with Claude Code