Implement Semantic Consolidator V0

- add deterministic canonicalization support for extraction items
- add facts-only semantic consolidation using local Ollama
- preserve source evidence and validate complete fact coverage
- add conservative merge rules and non-LLM tests
- record the first validated real-life consolidation benchmark
- document current scope, limitations and next evaluation step

prompts/consolidate_facts.md

@@ -0,0 +1,52 @@
+You consolidate fact items from Meeting Lab.
+
+Return only valid JSON.
+
+Semantic equivalence is stricter than topical similarity.
+
+Merge fact items only when they express the same core factual proposition.
+Related facts are not enough. Facts from the same topic are not enough.
+
+Preserve distinctions in:
+
+- actor
+- scope
+- timing
+- condition
+- certainty
+- responsibility
+- current state versus future intention
+
+Do not merge:
+
+- broader and narrower statements
+- cause and effect
+- process and responsibility
+- fact and interpretation
+- related statements from the same topic
+- statements that differ in actor, scope, timing, condition, or certainty
+
+Use only the provided fact items.
+Do not invent information.
+Do not rewrite evidence.
+Do not add source item IDs that were not provided.
+Every provided fact item ID must appear exactly once.
+When uncertain, keep items separate.
+False negatives are preferable to false-positive merges.
+
+Return JSON with this exact top-level shape:
+
+{
+  "groups": [
+    {
+      "canonical_text": "One factual statement preserving the shared meaning",
+      "source_item_ids": ["fact_0001"],
+      "merge_reason": "Why these items are equivalent, or why this singleton remains separate"
+    }
+  ]
+}
+
+For singleton groups, use the original fact text as canonical_text.
+For merged groups, canonical_text must contain only information shared by all
+source facts in the group.
+