fix(agent): persist the LLM confidence through metadata writes

The post-processor stored the LLM's confidence as `_llm_confidence`, but
that value could never be read back: `BaseModelMetadata.from_dict()`
deliberately excludes underscore-prefixed keys from `_unknown_fields` and
`to_dict()` strips private fields, so it was erased by the next metadata
write and was invisible to `read_metadata()`.  The enrichment evaluation
harness reads this field to score runs, so confidence was always scored
as blank.

Store it as `llm_confidence`, which round-trips as an ordinary unknown
field — the same mechanism `llm_enriched_at` already relies on.  Nothing
else consumed the old name, and the harness still accepts it so sidecars
written by earlier versions keep evaluating.

Covered by a metadata load/save round-trip regression test plus
assertions that the post-processor writes the persisted key and no longer
writes the private one.

py/services/agent/post_processor.py

@@ -273,10 +273,13 @@ class PostProcessor:
         updates["metadata_source"] = "agent:enrich_hf_metadata"
         updates["llm_enriched_at"] = datetime.now(timezone.utc).isoformat()
 
-        # Store LLM confidence in metadata so it's accessible for evaluation
+        # LLM confidence, stored for the enrichment evaluation harness.  The key
+        # must NOT start with an underscore: `BaseModelMetadata.from_dict()`
+        # deliberately drops underscore-prefixed keys so they never round-trip,
+        # which silently erased this field on the next metadata write.
         raw_confidence = (llm_output.get("confidence") or "").strip()
         if raw_confidence:
-            updates["_llm_confidence"] = raw_confidence
+            updates["llm_confidence"] = raw_confidence
 
         # Fallback: use the trigger words the site records for this exact file,
         # then the README's YAML `instance_prompt`, when the LLM returned none.