swap-stack: drop concluded A/B build variants (ik, upstream-master)

FINDINGS answered their questions: ik_llama -23% decode / +13% prefill,
upstream master parity with turboquant on K-quant decode. Binaries stay
in the image, macros kept for future experiments.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

swap-stack/config.yaml

@@ -148,37 +148,10 @@ models:
       --cont-batching --parallel 1
     ttl: 300
 
-  # ── ik_llama.cpp experimental variants ──────────────────────────────────────
-
-  "qwen36-35b-q2-ik":
-    name: "Qwen3.6-35B Q2_K_XL (ik_llama)"
-    description: "ik build: 18.0 tg / 42 pp — loses decode to main build (23.0), wins prefill. -ser 6,1 active. Experimental only"
-    env:
-      - "LD_LIBRARY_PATH=/app-ik"
-    cmd: |
-      ${server-ik} ${q8-kv}
-      --model /models/Qwen3.6-35B-A3B-UD-Q2_K_XL.gguf
-      --n-gpu-layers 99 --n-cpu-moe 37 --jinja -ser 6,1
-      --ctx-size 16384
-      --batch-size 1024 --ubatch-size 512
-      --parallel 1
-    ttl: 300
-
-  "gpt-oss-20b-upstream":
-    name: "gpt-oss-20b (upstream master)"
-    description: "Upstream Jul-2026 build comparison"
-    env:
-      - "LD_LIBRARY_PATH=/app-upstream"
-    cmd: |
-      ${server-upstream} ${q8-kv}
-      --model /models/gpt-oss-20b-mxfp4.gguf
-      --n-gpu-layers 99 --n-cpu-moe 21 --jinja
-      --ctx-size 32768
-      --batch-size 1024 --ubatch-size 512
-      --cont-batching --parallel 1
-    ttl: 300
-
   # ── Dense 9B (RAM-bandwidth-bound, ~4.4 t/s) ───────────────────────────────
+  # (ik_llama / upstream-master A/B variants removed 2026-07-10 — questions
+  #  answered in FINDINGS: ik −23% decode, upstream parity. Macros kept for
+  #  future experiments; binaries still in the image.)
 
   "ornith-9b":
     name: "Ornith-1.0-9B Q8_0"