Show raw AI output in Insights per turn

Store the model's literal reply (before the world-state block is stripped)
in each turn's snapshot, and render it in the Insights panel when inspecting
a past AI message (the 🔍 button). Makes it possible to see exactly what the
model emitted — including whether it sent a state delta and with what paths —
which is the fastest way to debug under-reported or misrouted stat changes.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

backend/app/routers/adventures.py

@@ -354,6 +354,9 @@ async def generate_turn(
         return
 
     text = "".join(chunks).strip()
+    # The model's literal reply, kept for the Insights "Raw AI output" view —
+    # this still contains any world-state block before it gets stripped below.
+    raw_output = text
     if not text:
         # If the model streamed reasoning but no story text, it spent its whole
         # budget thinking — say so instead of a mysterious "empty response".
@@ -390,6 +393,8 @@ async def generate_turn(
         adventure.world_state = new_world_state
         snapshot["world_state"] = {"delta": delta, "report": ws_report, "state": new_world_state}
 
+    snapshot["raw_output"] = raw_output
+
     ai_action = models.Action(
         adventure_id=adventure.id,
         index=ai_index,