verbose live UI + tool-level SSE events + Groq default + regression tests

api/prompts/compare_periods.pair.txt

@@ -0,0 +1,16 @@
+Task: build two parallel Postgres SELECT queries that compute the same metric across two different time windows, so an analyst can compare them.
+
+Output: a JSON object with exactly this shape, in one fenced ```json block:
+{
+  "a": {"label": "<short period label, e.g. '1995'>", "sql": "<SELECT …>"},
+  "b": {"label": "<short period label, e.g. '1996'>", "sql": "<SELECT …>"}
+}
+
+Constraints:
+- The two queries must return result sets with the same columns and shape — otherwise the diff is meaningless.
+- Dates are stored as YYMMDD integers; convert with TO_DATE(LPAD(date::text, 6, '0'), 'YYMMDD') when filtering by date.
+- Quote `"order"` if you reference it.
+- search_path is `financial`; don't qualify table names with the schema.
+- Read-only.
+
+If either period is given as "(infer)", pick a sensible window from the question and reflect it in the label.