Add deterministic geography: precomputed nearest-office distances injected per turn

The 2026-07-05 Weston call (CA3e88ce) showed the model guessing geography: it
recommended Boca Raton for a Weston caller (real answer: Pembroke Pines, ~9 mi),
invented distances ('Weston to Hialeah is twenty miles'), and confirmed a
nonexistent Miami Lakes office when the caller insisted.

- practice.py: office lat/lon + ~70-place South Florida gazetteer + haversine;
  geography_block() renders a full distance table (closest first) from the
  caller's area to every office. 'Westin' aliased to Weston (STT spelling).
- callstate.py: extractor gains caller_area; CallStateGroomer appends the
  GEOGRAPHY block to the system message — zero-lag via a synchronous gazetteer
  scan of user turns until the extractor catches up.
- bot.py: DISTANCES prompt rule (answer only from GEOGRAPHY, never estimate),
  only-these-8-offices guard, caller-area cities added to Whisper hotwords
  ('Weston' was heard as 'Western Florida').

Verified: staged replay of the failed exchanges on activeblue-avc:latest —
correct office, correct distances, Miami Lakes denied (5/5 + 3/3 runs).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

callstate.py

@@ -28,6 +28,8 @@ from loguru import logger
 from pipecat.frames.frames import BotStoppedSpeakingFrame, Frame, LLMContextFrame
 from pipecat.processors.frame_processor import FrameDirection, FrameProcessor
 
+import practice
+
 # Short, in-call variant of the post-call extractor (extract.py): only what's needed to
 # build the checklist, temperature 0, capped output. Runs on the local Ollama model.
 _STATE_INSTRUCTIONS = (
@@ -54,6 +56,9 @@ _STATE_INSTRUCTIONS = (
     '"billing question"). Only \'an appointment\' with no visit reason is NOT a reason — '
     "use null then.\n"
     '  "location": string or null — the office/city the caller wants\n'
+    '  "caller_area": string or null — the city/area where the CALLER themselves is located '
+    "or wants an office near (\"I'm in Weston\", \"what's closest to Davie\") — NOT an office "
+    "they merely asked about. Corrections win here too.\n"
     '  "patient_name": string or null — ONLY a name the caller explicitly gave as their own '
     "(\"my name is…\", \"this is…\", or answering a name question); NEVER a name guessed "
     "from other words, a garbled transcription, or a name only the assistant used\n"
@@ -252,12 +257,34 @@ class CallStateGroomer(FrameProcessor):
         )
         self._task.add_done_callback(self._extract_done)
 
+    def _geo_block(self, messages) -> str:
+        """Precomputed office-distance table for the caller's area (see
+        practice.geography_block). Anchor = the LLM-extracted caller_area when we have
+        one (it understands context and corrections); before that (extraction lags one
+        turn) a synchronous gazetteer scan of the user's own words — first place that
+        isn't one of our office cities, since office names in caller speech are usually
+        about our offices, not where they live."""
+        state = self._state or {}
+        for key in ("caller_area", "location"):
+            v = state.get(key)
+            if isinstance(v, str) and v.strip():
+                anchor = practice.match_place(v)
+                if anchor:
+                    return practice.geography_block(anchor)
+        hits = []
+        for m in messages:
+            if m.get("role") == "user" and isinstance(m.get("content"), str):
+                hits += practice.places_in_text(m["content"])
+        non_office = [h for h in hits if h not in practice.OFFICE_PLACE_KEYS]
+        anchor = (non_office or hits or [None])[0]
+        return practice.geography_block(anchor) if anchor else ""
+
     def _groom_context(self):
         messages = merge_consecutive_user_messages(list(self._context.messages))
-        block = build_state_block(self._state)
+        blocks = [b for b in (build_state_block(self._state), self._geo_block(messages)) if b]
         for i, m in enumerate(messages):
             if m.get("role") == "system":
-                content = self._base_system + ("\n\n" + block if block else "")
+                content = "\n\n".join([self._base_system, *blocks])
                 if m.get("content") != content:
                     messages[i] = {**m, "content": content}
                 break