feat: memory v2, prompt styles, Dream/GAZE integration, Wyoming TTS fix

SQLite + sqlite-vec replaces JSON memory files with semantic search,
follow-up injection, privacy levels, and lifecycle management.

Six prompt styles (quick/standard/creative/roleplayer/game-master/storyteller)
with per-style Claude model tiering (Haiku/Sonnet/Opus), temperature control,
and section stripping. Characters can set default style and per-style overrides.

Dream character import and GAZE character linking in the dashboard editor
with auto-populated fields, cover image resolution, and preset assignment.

Bridge: session isolation (conversation_id / 12h satellite buckets),
model routing refactor, PUT/DELETE support, memory REST endpoints.

Dashboard: mobile-responsive sidebar, retry button, style picker in chat,
follow-up banner, memory lifecycle/privacy UI, cloud model options in editor.

Wyoming TTS: upgraded to v1.8.0 for HA 1.7.2 compatibility.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

homeai-agent/prompt-styles/quick.json

@@ -0,0 +1,13 @@
+{
+  "id": "quick",
+  "name": "Quick",
+  "group": "cloud",
+  "model": "anthropic/claude-haiku-4-5-20251001",
+  "description": "Brief responses for commands and quick questions",
+  "thinking": "off",
+  "params": {
+    "temperature": 0.15
+  },
+  "instruction": "RESPONSE RULES — STRICT:\n- Respond as briefly as possible. For smart home commands, confirm with 1-3 words (\"Done.\", \"Lights on.\", \"Playing jazz.\").\n- For factual questions, give the shortest correct answer. One sentence max.\n- No small talk, no elaboration, no follow-up questions unless the request is genuinely ambiguous.\n- Never describe your actions, emotions, or thought process.\n- Never add flair, personality, or creative embellishments — be a reliable, predictable tool.\n- If a tool call is needed, execute it and report the result. Nothing else.",
+  "strip_sections": ["background", "appearance", "dialogue_style"]
+}