fix(consolidation): make the exact-dup guard case-sensitive

_norm_obs_text dropped a CREATE whose text matched a shown observation only after
lower-casing. But the guard's premise is that an exact-text match loses no info on
drop — case-folding breaks that (a create differing only in case, e.g. an acronym
like "TLS" vs "tls", carries different information). Collapse whitespace only;
keep case. The guard body stays — it's a zero-cost deterministic backstop for the
verbatim-re-create mode, complementary to interleave recall (which surfaces the twin;
the guard catches the LLM emitting an identical CREATE anyway).

Author: nicoloboschi

hindsight-api-slim/hindsight_api/engine/consolidation/consolidator.py

@@ -54,8 +54,14 @@
 
 
 def _norm_obs_text(text: str) -> str:
-    """Whitespace/case-normalised observation text for exact-duplicate matching."""
-    return " ".join((text or "").split()).strip().lower()
+    """Whitespace-normalised observation text for exact-duplicate matching.
+
+    Collapses runs of whitespace only; case is preserved. The reconciliation guard
+    drops a CREATE on the premise that an exact-text match loses no information — but
+    case-folding would also drop a create differing only in case (e.g. "TLS" vs "tls"),
+    which *does* lose information, so we match case-sensitively.
+    """
+    return " ".join((text or "").split()).strip()
 
 
 def _duplicate_create_target(

hindsight-api-slim/tests/test_consolidation_dedup.py

@@ -20,19 +20,27 @@ def _shown(*observations: _FakeObs) -> dict[str, _FakeObs]:
     return {_norm_obs_text(o.text): o for o in observations}
 
 
-def test_norm_obs_text_collapses_whitespace_and_case() -> None:
-    assert _norm_obs_text("  The  User  likes BASIL.\n") == "the user likes basil."
+def test_norm_obs_text_collapses_whitespace_preserves_case() -> None:
+    # Whitespace (incl. newlines) collapses; case is preserved.
+    assert _norm_obs_text("  The  User  likes BASIL.\n") == "The User likes BASIL."
     assert _norm_obs_text(None) == ""
 
 
 def test_create_matching_shown_observation_is_duplicate() -> None:
     shown = _shown(_FakeObs(id="11111111-aaaa", text="User waters the herbs early in the morning."))
-    # Same text with different whitespace/case still matches.
-    target = _duplicate_create_target("user waters the herbs early in the morning.", shown, set())
+    # Same text with only-whitespace differences still matches.
+    target = _duplicate_create_target("User waters the   herbs early in the morning.", shown, set())
     assert target is not None
     assert target.startswith("shown observation 11111111")
 
 
+def test_create_differing_only_in_case_is_not_duplicate() -> None:
+    # Case-folding would lose information (e.g. acronyms), so a case-only difference
+    # is treated as novel rather than silently dropped.
+    shown = _shown(_FakeObs(id="22222222-bbbb", text="The user prefers TLS."))
+    assert _duplicate_create_target("The user prefers tls.", shown, set()) is None
+
+
 def test_create_matching_inresponse_update_is_duplicate() -> None:
     update_texts = {_norm_obs_text("Mint is kept in its own separate bed.")}
     target = _duplicate_create_target("Mint is kept in its own separate bed.", {}, update_texts)