feat(deltanet): auto-enable TQ_DN_LLAMACPP_PORT for qwen35moe — ROOT CAUSE

Path B per-layer l_out diff harness localized Qwen3.6-A3B divergence
to L33+ (14.5% rel_diff on "Hello" prompt). Investigation revealed
the root cause is NOT quantized kernel precision but FP32 summation
order in the DeltaNet delta-rule state update:

  Our default: state[i*dv+j] layout, i-outer-j-inner FMA
  llama.cpp:   state[j*S_v+i] layout (transposed), j-outer-i-inner

Mathematically identical, but FP32 addition is non-associative.
Per-layer ~1e-5 drift compounds across 40 DeltaNet layers into:
  - L33 cos gap 14.5% (vs 0.01% at L0-L32)
  - Router softmax exp() amplifies into top-K boundary flip
  - Expert selection diverges from llama (same 8 experts, wrong order)
  - Cascading MoE ffn_out divergence through L34-L38
  - Final self-attn L39 amplification
  - cos(final_logits, llama) = 0.47 (massive)
  - Coherent output attractor at ~35 words

TQ_DN_LLAMACPP_PORT=1 uses the verbatim llama.cpp port (already
existed in-tree as R49 Big Move 3 but was gated off by default).
Measurement with NO_PRESET + T=1.0 + DN_PORT:

  L33 rel_diff: 14.5% → 4.3% (3.4× reduction)
  Coherent output on quantum prompt:
    baseline: "killing jargon.but keeping it accurate.k" (~5 coh words)
    DN_PORT:  "Superposition (Schrödinger's cat) → Entanglement →
               Wave-particle duality (Double-slit experiment) →
               Quantum tunneling → Quantum computing (quantum
               supremacy)" — real physics concepts, ~100 coh words

Auto-enable for qwen35moe. User opts out via TQ_QWEN35MOE_NO_PRESET=1
as before.

Why prior rounds missed this:
- R49 Big Move 3 originally implemented DN_LLAMACPP_PORT, but was
  tested alongside the existing T=2.0 auto-preset — the T=2.0 →
  divergent router path produced a different attractor, making the
  DN fix look like a regress. When tested with T=1.0 (matching
  llama's router), the true improvement shows.
- Memory note "DeltaNet port regressed" preserved that false
  conclusion across 10+ rounds.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Author: unamedkr

tools/quant.c

@@ -436,6 +436,22 @@ int main(int argc, char** argv) {
                 fprintf(stderr, "tq_main: qwen35moe DN_NORM_FP64 auto-enabled "
                         "(per-head RMSNorm in FP64, negligible memory).\n");
             }
+            /* R63 P4 (2026-04-24): verbatim llama.cpp gated_delta_net port.
+             * Root cause of late-layer divergence was FP32 summation order
+             * in the delta-rule state update. Our default uses state[i][j]
+             * layout with i-outer-j-inner FMA; llama uses state[j][i]
+             * (transposed) with j-outer-i-inner. Mathematically identical
+             * but FP32-non-associative: our per-layer ~1e-5 drift compounds
+             * across 40 layers into L33+ cos gap and expert-boundary flips.
+             * Measurement on "Hello" prompt: L33 rel_diff 14.5% -> 4.3%.
+             * Long-gen on quantum prompt: 35 coh -> ~100 coh words with
+             * real physics concepts (Schrödinger's cat, Entanglement,
+             * Wave-particle duality). */
+            if (!getenv("TQ_DN_LLAMACPP_PORT")) {
+                setenv("TQ_DN_LLAMACPP_PORT", "1", 0);
+                fprintf(stderr, "tq_main: qwen35moe DN_LLAMACPP_PORT auto-enabled "
+                        "(verbatim llama.cpp delta-rule FP32 accumulation order).\n");
+            }
             /* R62 K8: thinking mode 전용 FP32 KV cache. Direct mode에서는
              * turbo_kv_4b가 더 나음 (+regularizer effect), 하지만 thinking
              * mode는 long causal reasoning chain에 KV quant noise가 누적되어