fix(retrieval): structural prose⇄card 1:1 gate (KI-PROSE-CARD-1to1)

User requirement: "can't have a situation where it suggests something in
chat but won't generate a card; it should only suggest if it's worth
suggesting." Residual after KI-QUALITY-SEED: the LLM occasionally padded
prose with a sub-floor C-grade policy that the citation builder then gated
out (3 prose / 2 cards).

Fix: for intent=recommendation, apply the SAME fitness floor the citation
builder uses (single_brain._recommendation_fit, lazy import) to the
retrieved set BEFORE it reaches the LLM. The LLM therefore only ever sees
cardable policies → it cannot name one it won't card → prose ⇄ cards are
1:1 by construction. If nothing clears the floor the set is empty and the
LLM honestly says "no strong match" rather than padding (also the
requirement). qa / follow-up intents untouched (they cite supporting
source chunks regardless of recommendation grade).

Verified end-to-end: name captured turn 0 → fact-find → quality-seeded
retrieval → recommends ManipalCigna Sarvah Param (A/86), SBI Super Health
(A/82), ICICI Health AdvantEdge (A/80) with real UINs → every named
policy cited (CITATIONS 3-4, strict 1:1). Full pytest gate green (rc=0,
0 failures); scorecard-parity green. Not pushed (deploy-altogether).

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

backend/brain_tools.py

@@ -989,6 +989,24 @@ async def retrieve_policies(
         # session cache too so prose ⇄ cited cards are gated on the SAME set.
         filtered = filtered[: max((int(top_k) if top_k else 8), 12)]
 
+        # KI-PROSE-CARD-1to1 (2026-05-17) — user requirement: "can't have a
+        # situation where it suggests something in chat but won't generate a
+        # card." For a RECOMMENDATION turn, apply the SAME fitness floor the
+        # citation builder uses (_recommendation_fit) HERE, so the LLM only
+        # ever SEES cardable policies → prose ⇄ cards are 1:1 by
+        # construction (it cannot name a policy it won't card because it
+        # never receives one). If nothing clears the floor we return none —
+        # the LLM then honestly says "no strong match" rather than padding
+        # with a sub-floor policy (also the requirement). qa / follow-up
+        # intents are untouched: they legitimately cite supporting source
+        # chunks regardless of recommendation grade.
+        if (intent or "").lower() == "recommendation" and filtered:
+            try:
+                from backend.single_brain import _recommendation_fit
+                filtered = [c for c in filtered if _recommendation_fit(c)[0]]
+            except Exception as e:  # noqa: BLE001 — never break retrieval
+                _log.warning("rec-fit 1:1 gate failed: %s", e)
+
     # X7 — cache slug→insurer lookups on session so a subsequent
     # mark_recommendation call (same turn) can stamp the right insurer on the
     # shown_policies event. Each retrieve_policies call MERGES into the cache