Improve VLM client lifecycle, retries, metrics, and cache keys

- Adds abstract aclose() and shared load_prompt() helper to BaseVLMClient.

- OllamaClient now logs latency, tokens, and prompt hash per call.

- Retries cover HTTP 429 and 5xx in addition to network/timeout errors.

- Cache key now includes response_format and max_tokens.

- Warns when JSON schema generation falls back to plain 'json' format.

- Adds FakeVLMClient.aclose() and cache unit tests.

tests/test_cache.py

@@ -0,0 +1,44 @@
+"""Tests for the LLM response cache."""
+
+from __future__ import annotations
+
+import pytest
+
+from odoo_ocr.config import Settings
+from odoo_ocr.utils import cache
+
+
+@pytest.fixture
+def cache_settings(tmp_path) -> Settings:
+    return Settings(cache={"enabled": True, "dir": tmp_path / "llm_cache"})
+
+
+def test_cache_roundtrip(cache_settings: Settings) -> None:
+    response = {"content": "hello", "model": "m", "prompt_tokens": 1}
+    cache.set_cached("sys", "user", [b"img"], "model", 0.1, 4096, response, cache_settings)
+    assert cache.get_cached("sys", "user", [b"img"], "model", 0.1, 4096, cache_settings) == response
+
+
+def test_cache_miss(cache_settings: Settings) -> None:
+    assert cache.get_cached("a", "b", [], "m", 0.1, 4096, cache_settings) is None
+
+
+def test_cache_disabled(tmp_path) -> None:
+    settings = Settings(cache={"enabled": False, "dir": tmp_path / "llm_cache"})
+    cache.set_cached("sys", "user", [], "m", 0.1, 4096, {"content": "x"}, settings)
+    assert cache.get_cached("sys", "user", [], "m", 0.1, 4096, settings) is None
+    assert not (tmp_path / "llm_cache").exists()
+
+
+def test_cache_key_includes_temperature_format_and_max_tokens(cache_settings: Settings) -> None:
+    response = {"content": "x"}
+    cache.set_cached("sys", "user", [], "m", 0.1, 4096, response, cache_settings, format_key="text")
+    # Different temperature must not hit the same entry.
+    assert cache.get_cached("sys", "user", [], "m", 0.2, 4096, cache_settings) is None
+    # Different response_format must not hit the same entry.
+    assert (
+        cache.get_cached("sys", "user", [], "m", 0.1, 4096, cache_settings, format_key="ExtractedInvoice")
+        is None
+    )
+    # Different max_tokens must not hit the same entry.
+    assert cache.get_cached("sys", "user", [], "m", 0.1, 2048, cache_settings) is None