docs: README/ARCHITECTURE — list get_policy_facts tool + note #47 UIN dedup

The brain has 4 function-calling tools; the §4.2 ASCII fallback and
ARCHITECTURE.md's one-screen summary listed only 3 (get_policy_facts
was missing — the §4.3 Mermaid diagram already had it). Also documents
the #47 UIN net-new dedup in §6.1 Uploaded-PDF defence. Architecture
itself unchanged this session — the bug fixes (#41/#43/#44/#45/#47)
were internal logic/data corrections, not architecture changes.

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

ARCHITECTURE.md

@@ -20,7 +20,8 @@ authority.
   `/api/admin/*`.
 - **Brain** — one LLM call per turn: Google Gemini
   (`gemini-2.5-flash`) + function-calling tools
-  (`save_profile_field`, `retrieve_policies`, `mark_recommendation`) in
+  (`save_profile_field`, `retrieve_policies`, `get_policy_facts`,
+  `mark_recommendation`) in
   `backend/single_brain.py` / `backend/brain_tools.py`. A single call owns
   the whole turn: fact-find, retrieval, QA, and recommendation. On a
   transient Gemini error / cold-start 503 → small `backend/nim_fallback.py`

README.md

@@ -375,6 +375,7 @@ flowchart LR
             │  tools (backend/brain_tools.py):              │
             │   • retrieve_policies(query, filters, top_k)  │
             │   • save_profile_field(...)                   │
+            │   • get_policy_facts(policy_ids)              │
             │   • mark_recommendation(...)                  │
             │  └ on Gemini failure / first-turn 503 →       │
             │    backend/nim_fallback.py  (NIM chain,       │
@@ -498,6 +499,11 @@ before the file is ever embedded or shown to the model:
 7. **Page-count ceiling** — >200 pages is an abuse/bundle vector.
 8. **Hash dedupe + reject-cache** — identical re-uploads short-circuit.
 
+Beyond identical-file dedup, a **UIN net-new check** runs on every upload:
+if the PDF's IRDAI UIN already belongs to a catalogued policy, the upload
+is recognised as *not* net-new and the caller is pointed at the existing
+marketplace card instead of a duplicate being indexed.
+
 Accepted uploads are embedded into a **separate, per-session quarantine**
 Chroma collection (never the shared corpus), scoped by `session_id` so one
 user's document is invisible to another, and auto-purged after a 24-hour idle