feat(spec): GF-T reciprocal + complete GF-T softmax (Refs #1764)

[gHashTag]

Stacked on #1803 (needs its exp2 primitive). Base will be retargeted to master once #1803 merges.

gft_recip.t27 — GF-T reciprocal 1/x (the last softmax primitive)

Closed form, no polynomial: 1/(1+m/512) = 512/(512+m), written in GF-T form gives 512+m' = round(524288/(512+m)), off' = 79−o (+renorm). One rounded integer divide → EXACT (0 ULP) vs true 1/x (30000 samples). iverilog 506/506 bit-exact; spot checks exact (1/1=1, 1/2=0.5, 1/0.5=2, 1/−1=−1, 1/4=0.25).

gft_softmax4.t27 — the complete GF-T softmax

Softmax over 4 signed GF-T16 logits, base-2, max-stabilized:

p_i = 2^(l_i − M) / Σ_j 2^(l_j − M),   M = max_j l_j

Returns p_sel. Composes all the now-verified GF-T primitives: max (gt), subtract (sadd+neg), exp2 (≤1 ULP), sum (sadd), reciprocal (exact), multiply (RNE).

• Bit-exact to the integer oracle 2000/2000 (iverilog).
• Accuracy ≤ 0.0017 absolute probability vs true float softmax. Sum-of-probs ∈ [0.9982, 1.0018]; uniform logits → 0.25 exact; [2,1,1,1] → 0.40.

Fresh seals for GftRecip + GftSoftmax4 (seal --verify → MATCH). No compiler change.

This completes the GF-T classifier head: activations → MLP → logits → {argmax → class} OR {softmax → calibrated probabilities}.

Refs #1764

⚠️ GitHub Actions infra incident still ongoing (jobs stuck queued). All verification above is local (iverilog + seals).

🤖 Generated with Claude Code

[gHashTag]

✅ Synthesizability confirmed (yosys). read_verilog → hierarchy → proc → flatten → opt → techmap → opt → stat completes cleanly on GftSoftmax4:

• 132,981 cells (42324 \$_AND_, 36994 \$_MUX_, 24918 \$_OR_, 14724 \$_XOR_, 14017 \$_NOT_), 49817 wires. ~54s, 2.3 GB peak.
• No Function can only be called with constant arguments error — the nested runtime-arg composition (exp2(sadd(l_i, neg(mx))), recip, magmul) synthesizes fine (combinational, no latches). This retires the earlier open question about whether nested GF-T primitive calls are synthesizable.

Area is large (deep GF-T arith trees × 4 logits + the exp2 poly + recip divide), as expected for a full softmax — it is a candidate for pipelining/sharing before a board run, but it is synthesizable today.