feat: spirit engine — seance WS, entity minting/Codex, Piper TTS voices, wire telemetry

- WS /ws/session: modes, summon, anomaly fragments, streaming direct contact,
  passive wire-ghost ambient loop, per-user rate limits, event transcript
- entities: anomaly-signature fingerprinting, LLM minting + procedural
  fallback, Codex matching with contact counts and sightings
- tts: 8 local Piper voices (EN/ES), per-entity voice profiles, numpy
  effects chain (pitch/rate/bitcrush/echo/static)
- llm: streaming client, submit_stream in bounded queue, SpiritService
  with offline fallbacks for every channel
- routes: public /api/codex, /api/codex/{id}, /api/stats; /audio static mount
- models: Entity, EntitySighting, Event, ContactSession(entity_id, language)

backend/app/llm/queue.py

@@ -1,5 +1,5 @@
 import asyncio
-from collections.abc import Awaitable, Callable
+from collections.abc import AsyncIterator, Awaitable, Callable
 from typing import TypeVar
 
 T = TypeVar("T")
@@ -44,3 +44,27 @@ class LLMQueue:
                 self._semaphore.release()
             else:
                 self._waiting -= 1
+
+    async def submit_stream(
+        self, gen_factory: Callable[[], AsyncIterator[T]]
+    ) -> AsyncIterator[T]:
+        """Like submit(), but for async generators (token streams). The
+        concurrency slot is held for the stream's whole lifetime, since the
+        Ollama box stays busy until the last token."""
+        async with self._lock:
+            if self._waiting >= self._max_queue_depth:
+                raise QueueFullError("too many seekers right now")
+            self._waiting += 1
+
+        acquired = False
+        try:
+            await self._semaphore.acquire()
+            acquired = True
+            self._waiting -= 1
+            async for item in gen_factory():
+                yield item
+        finally:
+            if acquired:
+                self._semaphore.release()
+            else:
+                self._waiting -= 1