Studio RAG trim: remove cross-encoder reranker

The cross-encoder reranker is off by default (enable_rerank=False everywhere)
and adds a second model download plus a candidate-widening pass on every search.
Removing it keeps the core retrieval (parse, chunk, embed, BM25 + dense, RRF,
search_knowledge_base tool) intact while dropping ~375 lines.

- delete core/rag/reranker.py and its test
- drop enable_rerank / reranker_model from the search tool, tools dispatch,
  and the /rag/search route (candidate_k is now just top_k)
- remove the /rag/reranker/precache endpoint and reranker config knobs
- update tool-handler test to the trimmed rag_scope shape

42 RAG tests pass.

tests/python/test_rag_reranker.py

@@ -1,64 +0,0 @@
-"""Reranker tests — skipped if sentence_transformers is unavailable.
-
-These tests load a real CrossEncoder, so they're slow and gated under
-the ``server`` marker so a default ``pytest`` run skips them. Force
-with ``pytest -m server``.
-"""
-
-import sys
-from pathlib import Path
-
-import pytest
-
-REPO_ROOT = Path(__file__).resolve().parents[2]
-STUDIO_BACKEND = REPO_ROOT / "studio" / "backend"
-if str(STUDIO_BACKEND) not in sys.path:
-    sys.path.insert(0, str(STUDIO_BACKEND))
-
-pytest.importorskip("sentence_transformers")
-
-
-def test_rerank_empty_returns_empty():
-    from core.rag.reranker import rerank
-
-    assert rerank("anything", []) == []
-
-
-@pytest.mark.server
-def test_rerank_reorders_by_relevance(monkeypatch):
-    """Hide the relevant chunk at the back of the input and check it bubbles up."""
-    monkeypatch.setenv(
-        "UNSLOTH_RAG_RERANKER_MODEL", "cross-encoder/ms-marco-MiniLM-L-6-v2"
-    )
-    from core.rag.reranker import rerank, unload
-    from core.rag.retrieval import Hit
-
-    pairs = [
-        (Hit("noise1", 0.0), "Cats are small carnivorous mammals."),
-        (Hit("noise2", 0.0), "The Eiffel Tower is in Paris, France."),
-        (Hit("noise3", 0.0), "Python is a programming language."),
-        (
-            Hit("answer", 0.0),
-            "The speed of light in vacuum is approximately 299792458 meters per second.",
-        ),
-    ]
-    try:
-        ranked = rerank("How fast does light travel?", pairs, top_k = 2)
-        assert ranked
-        assert ranked[0].chunk_id == "answer"
-    finally:
-        unload()
-
-
-@pytest.mark.server
-def test_unload_clears_singleton(monkeypatch):
-    monkeypatch.setenv(
-        "UNSLOTH_RAG_RERANKER_MODEL", "cross-encoder/ms-marco-MiniLM-L-6-v2"
-    )
-    from core.rag import reranker
-    from core.rag.retrieval import Hit
-
-    reranker.rerank("q", [(Hit("a", 0.0), "some text")])
-    assert reranker._model is not None
-    reranker.unload()
-    assert reranker._model is None

tests/python/test_rag_tool_handler.py

@@ -196,8 +196,6 @@ def test_execute_tool_dispatches_to_search_knowledge_base():
         top_k = None,
         scope_kb_id = None,
         scope_thread_id = None,
-        enable_rerank = False,
-        reranker_model = None,
         default_top_k = 5,
         min_score = 0.0,
         **kwargs,
@@ -206,7 +204,6 @@ def test_execute_tool_dispatches_to_search_knowledge_base():
         called["top_k"] = top_k
         called["scope_kb_id"] = scope_kb_id
         called["scope_thread_id"] = scope_thread_id
-        called["enable_rerank"] = enable_rerank
         called["default_top_k"] = default_top_k
         called["min_score"] = min_score
         return "stub-result"
@@ -218,7 +215,6 @@ def test_execute_tool_dispatches_to_search_knowledge_base():
             tool_context = {
                 "rag_scope": {
                     "kb_id": "kb-1",
-                    "enable_rerank": True,
                     "default_top_k": 3,
                     "min_score": 0.35,
                 }
@@ -229,7 +225,6 @@ def test_execute_tool_dispatches_to_search_knowledge_base():
     assert called["top_k"] == 7
     assert called["scope_kb_id"] == "kb-1"
     assert called["scope_thread_id"] is None
-    assert called["enable_rerank"] is True
     assert called["default_top_k"] == 3
     assert called["min_score"] == 0.35