Raise Q&A max output tokens to 2048

Raise the generation budget from 768 to 2048 tokens in both the
transformers and llama.cpp Q&A adapters. With thinking enabled the
reasoning trace consumes output budget before the answer starts, and a
768-token cap risks truncating the trace, which the strippers now
convert into an empty answer rather than leaked reasoning.

The 32k context window leaves ample room, and the 1B model on the
ZeroGPU a10g stays well inside the 90-second GPU task duration even at
the full budget.

src/vivamais/adapters/models/llama_cpp.py

@@ -365,7 +365,7 @@ QA_VARIANT = ModelVariant(
 
 
 QA_TEMPERATURE = 0.2
-QA_MAX_TOKENS = 768
+QA_MAX_TOKENS = 2048
 QA_N_CTX_DEFAULT = 32768
 
 

src/vivamais/adapters/models/transformers_text.py

@@ -36,7 +36,7 @@ def _zero_gpu_active() -> bool:
 
 
 DEFAULT_QA_MODEL = "marinarosa/MiniCPM5-1B-PTBR-v3"
-QA_MAX_NEW_TOKENS = 768
+QA_MAX_NEW_TOKENS = 2048
 QA_GPU_DURATION_SECONDS = 90
 STREAM_CHUNK_TIMEOUT_SECONDS = 60
 QA_TEMPERATURE = 0.2

tests/unit/adapters/test_llama_cpp.py

@@ -200,7 +200,7 @@ class TestMiniCPM5TextModel:
         model.generate([{"role": "user", "content": "q"}])
         _, kwargs = fake.create_chat_completion.call_args
         assert kwargs["temperature"] == 0.2
-        assert kwargs["max_tokens"] == 768
+        assert kwargs["max_tokens"] == 2048
 
     def test_default_context_window_is_32k(self) -> None:
         from vivamais.adapters.models.llama_cpp import MiniCPM5TextModel