fix(config): default agent model to gemma4:e2b

mistral-nemo-large holds ~9.2 GB of the 11 GB card it shares with Speaches,
which starves Whisper and breaks voice transcription. gemma4:e2b holds
1.9 GB and is faster. The deployed stack already overrides this via
OLLAMA_AGENT_MODEL; this aligns the default so a deployment without that
override does not reintroduce the contention.

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

webber-api/.env.example

@@ -14,7 +14,7 @@ CORS_ORIGINS=["http://localhost:3000","http://localhost:8080"]
 
 # LLM - Ollama (tower-of-joy)
 OLLAMA_URL=http://192.168.86.149:11434
-OLLAMA_AGENT_MODEL=mistral-nemo-large:latest
+OLLAMA_AGENT_MODEL=gemma4:e2b
 OLLAMA_EMBED_MODEL=nomic-embed-text:latest
 
 # Auth - Tatlock integration (optional)

webber-api/CHANGELOG.md

@@ -7,6 +7,9 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
+### Changed
+- Default `OLLAMA_AGENT_MODEL` is now `gemma4:e2b` instead of `mistral-nemo-large:latest`, so a deployment without an explicit override no longer exhausts shared GPU memory
+
 ## [1.0.1] - 2026-07-19
 
 ### Changed

webber-api/src/shared/config.py

@@ -64,7 +64,7 @@ class Settings(BaseSettings):
 
     # LLM - Ollama (always hot in VRAM on tower-of-joy)
     ollama_url: str = "http://192.168.86.149:11434"
-    ollama_agent_model: str = "mistral-nemo-large:latest"
+    ollama_agent_model: str = "gemma4:e2b"
     ollama_embed_model: str = "nomic-embed-text:latest"
 
     # Auth - Tatlock integration