fix(llm): judge theme tags on instruct_model only

The judge stepped up to instruct_model_large on a host that serves it,
and the run recorded that model in every item's tag state. Pulling
qwen2.5:32b on the rig would have re-tagged all 315k comments with a
model the golden gate never scored. Chat keeps the large tier.

src/cli/llm.py

@@ -248,7 +248,7 @@ def _tagging_runtime(cfg: Any) -> dict[str, Any]:
     in each item's state."""
     from conf.connect import bib
     from llm.index import _engine
-    from llm.pool import HostPool, PoolEmbeddings, pick_model
+    from llm.pool import HostPool, PoolEmbeddings
     from llm.tagger import card_vectors, make_judge, pg_chunks
     from llm.vocab import load as load_vocab
 
@@ -257,8 +257,7 @@ def _tagging_runtime(cfg: Any) -> dict[str, Any]:
     pool.check(cfg.embed_model)
     cards = card_vectors(vocab, PoolEmbeddings(pool, cfg.embed_model).embed_documents)
     judge = make_judge(cfg, pool)
-    with pool.acquire_generation() as host:
-        model = pick_model(cfg, pool, host)
+    model = cfg.instruct_model  # what make_judge sends; recorded in tag state
     return {
         "store": bib(),
         "vocab": vocab,

src/llm/tagger.py

@@ -322,10 +322,11 @@ def card_vectors(vocab: Vocab, embed: Embed) -> dict[str, list[float]]:
 def make_judge(
     cfg: Any, pool: Any, *, post: Callable[..., dict] | None = None
 ) -> Judge:
-    """A yes/no judge on the largest live host, ``temperature 0`` — the
-    same route ``llm.classify.closed_vocab_classifier`` takes."""
+    """A yes/no judge on the largest live host, ``temperature 0``, always
+    on ``cfg.instruct_model`` — the model the golden gate scored. Unlike
+    chat it never steps up to ``instruct_model_large``: tag state records
+    the model, so a bigger host appearing must not re-tag the corpus."""
     from llm.classify import _default_post
-    from llm.pool import pick_model
 
     send = post or _default_post
     pool.check(cfg.instruct_model)
@@ -338,7 +339,7 @@ def make_judge(
         # live host, the same discipline embed_texts follows (#796).
         while True:
             with pool.acquire_generation() as host:
-                model = pick_model(cfg, pool, host)
+                model = cfg.instruct_model
                 try:
                     data = send(
                         f"{host}/api/chat",

tests/llm/test_tagger.py

@@ -200,6 +200,30 @@ class TestJudge:
         )
         assert sent["json"]["model"] == "m"
 
+    def test_judge_stays_on_the_gated_model_when_the_big_host_has_the_large_one(self):
+        # The golden gate scored instruct_model; tag state records the
+        # model, so judging with the large tier would re-tag the corpus.
+        from llm.pool import HostPool
+
+        pool = HostPool(["http://rig"], vram_gb={"http://rig": 24})
+        pool.check = lambda model: ["http://rig"]  # type: ignore[method-assign]
+        pool._models = {"http://rig": {"m", "big"}}
+        sent = {}
+
+        def post(url, json=None):
+            sent["model"] = json["model"]
+            return {"message": {"content": "yes"}}
+
+        class Cfg:
+            instruct_model = "m"
+            instruct_model_large = "big"
+            large_min_vram_gb = 20.0
+            chat_num_ctx = 4096
+
+        judge = make_judge(Cfg(), pool, post=post)
+        assert judge(VOCAB.get("drugs"), "x") is True
+        assert sent["model"] == "m"
+
     def test_judge_drops_a_dying_host_and_retries_on_the_survivor(self):
         import httpx