fix(runtime): grammar bench measures OFF only to avoid ZeroGPU abort

Running grammar-on + grammar-off back-to-back inside the single demo.load
warmup handler stacked a 3rd GPU acquisition after the warmup gen, which
ZeroGPU aborted ("GPU task aborted"). grammar-on throughput is already the
normal [zerogpu-bench] number every real freeform call prints (~17-22 tok/s),
so measure only the unconstrained grammar-off run here — keeping the handler
to 2 GPU acquisitions (warmup + bench), the range the log shows succeeds.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

src/buddy_fusion/runtime.py

@@ -475,34 +475,38 @@ def warmup() -> None:
 
 
 def _grammar_bench(cfg: "LlamaRuntimeConfig", model_path: str) -> None:
-    """One-off diagnostic: run the same realistic freeform prompt twice — once
-    with the GBNF grammar enforced, once unconstrained — to isolate how much
-    grammar-constrained sampling costs vs raw 12B decode. Gemma-4-12B measured
-    ~20 tok/s on ZeroGPU *with* the schema; if the grammar is the throttle the
-    unconstrained run should jump well past it. Gated by BUDDY_GRAMMAR_BENCH=1;
-    read the two labelled lines off the Space log, then unset the env. Both runs
-    use the same n_ctx=4096 / freeform GPU wrapper as the real calls so the
-    numbers transfer directly."""
+    """One-off diagnostic: run the freeform prompt UNCONSTRAINED (no GBNF grammar)
+    to isolate how much grammar-constrained sampling costs vs raw 12B decode.
+
+    Only the grammar-OFF run lives here: the grammar-ON throughput is already the
+    normal [zerogpu-bench] number every real freeform call prints (~17-22 tok/s),
+    so there is no need to re-measure it. Running BOTH inside this one demo.load
+    handler stacked a 3rd GPU acquisition after the warmup gen and ZeroGPU aborted
+    it ("GPU task aborted"); keeping the handler to 2 acquisitions (warmup + this)
+    is in the range the log shows succeeds. Uses the same n_ctx=4096 / freeform
+    GPU wrapper as the real calls so the number transfers directly. Gated by
+    BUDDY_GRAMMAR_BENCH=1; read [grammar-bench-off] off the log, compare against
+    the [zerogpu-bench] (grammar-on) lines, then unset the env."""
     import json
     from .exemplars import pick_exemplar
-    from .prompts import build_freeform_messages, BOX_SCHEMA
+    from .prompts import build_freeform_messages
 
     prompt = "a small round dragon with two curved horns and a long tail"
     ex = pick_exemplar(prompt)
-    common = dict(
-        model_path=model_path,
-        messages=build_freeform_messages(prompt, json.dumps(ex)),
-        max_tokens=int(os.getenv("BUDDY_GRAMMAR_BENCH_TOKENS", "400")),
-        temperature=0.8,
-        top_p=0.9,
-        n_ctx=max(cfg.n_ctx, 4096),
-        n_threads=cfg.n_threads,
-        n_gpu_layers=cfg.n_gpu_layers,
-        flash_attn=cfg.flash_attn,
-    )
     try:
-        _run_freeform_completion(schema=BOX_SCHEMA, label="grammar-bench-on", **common)
-        _run_freeform_completion(schema=None, label="grammar-bench-off", **common)
+        _run_freeform_completion(
+            model_path=model_path,
+            messages=build_freeform_messages(prompt, json.dumps(ex)),
+            schema=None,
+            max_tokens=int(os.getenv("BUDDY_GRAMMAR_BENCH_TOKENS", "400")),
+            temperature=0.8,
+            top_p=0.9,
+            n_ctx=max(cfg.n_ctx, 4096),
+            n_threads=cfg.n_threads,
+            n_gpu_layers=cfg.n_gpu_layers,
+            flash_attn=cfg.flash_attn,
+            label="grammar-bench-off",
+        )
     except Exception as exc:  # noqa: BLE001 - diagnostic only, never break warmup
         print(f"[grammar-bench] failed: {exc}", flush=True)