Vision backends: any OpenAI-compatible endpoint, including local models

The Anthropic key was the last hard gate for other users. extract's
VisionFn seam gains a second factory speaking the chat-completions
format — OpenAI, OpenRouter, or a local runtime (Ollama, LM Studio,
llama.cpp, vLLM) via config.toml: vision_provider, vision_base_url,
and vision_key_env ("" = keyless local endpoint, no Authorization
header sent). Anthropic stays the default. load_config rejects unknown
providers loudly, the pipeline page's credentials warning follows the
configured provider (a keyless local endpoint warns about nothing),
and config.toml + README document the local-model trade honestly:
weaker spine reading means a longer proofread pass, which the shaky-
read workflow absorbs.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016jXZFSTZQKzAC8fqpWSz9g

tests/test_extract.py

@@ -499,3 +499,85 @@ def test_systemic_failures_abort_and_exit_nonzero(tmp_path):
     with pytest.raises(typer.Exit):
         run_extract(cfg, vision=broken_vision)
     assert len(calls) == 3  # aborted after 3 identical failures
+
+
+# -- pluggable vision backends -------------------------------------------
+
+
+def test_openai_compatible_vision_speaks_the_format(monkeypatch):
+    import httpx
+
+    from bggpipe.extract import VISION_PROMPT, openai_compatible_vision
+
+    captured = {}
+
+    def fake_post(url, **kwargs):
+        captured.update(kwargs, url=url)
+        request = httpx.Request("POST", url)
+        return httpx.Response(
+            200,
+            request=request,
+            json={"choices": [{"message": {"content": '{"titles": []}'}}]},
+        )
+
+    monkeypatch.setattr(httpx, "post", fake_post)
+    monkeypatch.setenv("MY_VISION_KEY", "sk-test")
+    vision = openai_compatible_vision(
+        "llava", "http://localhost:11434/v1/", "MY_VISION_KEY"
+    )
+    assert vision("QUJD", "image/jpeg") == '{"titles": []}'
+    assert captured["url"] == "http://localhost:11434/v1/chat/completions"
+    assert captured["headers"]["Authorization"] == "Bearer sk-test"
+    body = captured["json"]
+    assert body["model"] == "llava"
+    image, prompt = body["messages"][0]["content"]
+    assert image["image_url"]["url"] == "data:image/jpeg;base64,QUJD"
+    assert prompt["text"] == VISION_PROMPT
+
+
+def test_openai_compatible_vision_local_needs_no_key(monkeypatch):
+    import httpx
+
+    from bggpipe.extract import openai_compatible_vision
+
+    captured = {}
+
+    def fake_post(url, **kwargs):
+        captured.update(kwargs)
+        return httpx.Response(
+            200,
+            request=httpx.Request("POST", url),
+            json={"choices": [{"message": {"content": "[]"}}]},
+        )
+
+    monkeypatch.setattr(httpx, "post", fake_post)
+    vision = openai_compatible_vision("llava", "http://localhost:11434/v1", "")
+    assert vision("QUJD", "image/jpeg") == "[]"
+    assert "Authorization" not in captured["headers"]
+
+
+def test_vision_for_dispatches_by_provider(monkeypatch, tmp_path):
+    import bggpipe.extract as ex
+
+    monkeypatch.setattr(ex, "default_vision", lambda model: f"anthropic:{model}")
+    cfg = Config(model="claude-sonnet-5")
+    assert ex.vision_for(cfg) == "anthropic:claude-sonnet-5"
+    with pytest.raises(ValueError, match="vision_base_url"):
+        ex.vision_for(Config(vision_provider="openai-compatible"))
+    fn = ex.vision_for(
+        Config(
+            vision_provider="openai-compatible",
+            vision_base_url="http://localhost:11434/v1",
+            model="llava",
+        )
+    )
+    assert callable(fn)
+
+
+def test_config_rejects_unknown_vision_provider(tmp_path):
+    from bggpipe.config import load_config
+
+    p = tmp_path / "config.toml"
+    p.write_text('vision_provider = "gpt-magic"\n')
+    with pytest.raises(ValueError, match="vision_provider"):
+        load_config(p)