feat(llm): LLM_INSTRUCT_MODEL overrides the instruct model per run

Lets eval-tags --live compare judge models (14b vs 32b on the rig)
without editing stack.toml.

src/llm/config.py

@@ -4,6 +4,7 @@ Env contract:
     LLM_OLLAMA_HOSTS     comma-separated Ollama base URLs, each optionally
                          annotated with declared VRAM: ``http://rig:11434@24``
                          (beats [llm].ollama)
+    LLM_INSTRUCT_MODEL   beats [llm].instruct_model (judge/model comparisons)
     LLM_PG_HOST          beats [llm].pg_host (containers set this to "postgres")
     LLM_DB_PASSWORD      required for pg_url(); lives in .env, never stack.toml
     LLM_BUILD_ANN_INDEX  beats [llm].build_ann_index ("1"/"true"/"yes" = True)
@@ -121,7 +122,8 @@ def load() -> LlmConfig:
         ollama_hosts=tuple(url for url, _ in pairs),
         host_vram={url: vram for url, vram in pairs},
         embed_model=str(section.embed_model),
-        instruct_model=str(section.instruct_model),
+        instruct_model=os.environ.get("LLM_INSTRUCT_MODEL")
+        or str(section.instruct_model),
         instruct_model_large=str(_opt(section, "instruct_model_large", "")),
         large_min_vram_gb=float(_opt(section, "large_min_vram_gb", 20.0)),
         chat_num_ctx=int(_opt(section, "chat_num_ctx", 8192)),

tests/llm/test_config.py

@@ -143,3 +143,13 @@ def test_iom_companion_max_defaults_to_four():
         pg_user="u",
     )
     assert cfg.iom_companion_max == 4
+
+
+def test_instruct_model_env_override(monkeypatch):
+    # Judge-model comparisons (14b vs 32b) without editing stack.toml.
+    from llm.config import load
+
+    monkeypatch.setenv("LLM_INSTRUCT_MODEL", "qwen2.5:32b")
+    assert load().instruct_model == "qwen2.5:32b"
+    monkeypatch.delenv("LLM_INSTRUCT_MODEL")
+    assert load().instruct_model == "qwen2.5:14b"