feat(spec): GF-T SGD weight update w'=w-eta*g — training loop closed (Refs #1764)

[gHashTag]

Stacked on #1806. Base retargets to master as the stack merges.

gft_sgd_step.t27 — GF-T SGD weight update w' = w − η·g

The final brick of an on-device training step. η = positive learning rate, g = signed gradient (from #1806), w = signed weight. Composes a signed multiply smul (sign = XOR of the signs, magnitude = the verified RNE magnitude multiply) with subtract (sadd + neg).

• Bit-exact to the integer oracle 500/500 (iverilog).
• Accuracy to GF-T16 precision (≤1 ULP; ~0.03 abs only at the largest magnitudes).
• Spot: w=1, g=0.5, η=0.5 → 0.75 exact; g=0 → w unchanged; w=1, g=−1, η=1 → 2.0 (ascent) exact.

---

The full on-device training loop is now expressible spec-first on GF-T:

logits → softmax → prob → NLL loss     (forward, #1804/#1805)
       → gradient  p − y                (backward, #1806)
       → w' = w − η·g                   (update, this PR)

Every stage iverilog-verified bit-exact. Combined with the BitNet×GF-T layers/MLP/classifier, GF-T now spans a complete train + infer stack in synthesizable spec-first hardware.

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