FEAT: Add /api/analyze-video endpoint with dedicated Gemini quota isolation

FIX: geometry_scatter — also fix rings→ring_count on secondary sphere (line 115)

FIX: import asyncio in latex_agent (run_in_executor fix was missing import)

FIX: geometry_scatter — rings→ring_count (Blender 3.4) + fix grid surface duplicate

FIX: Prevent job queue starvation and event loop blocking

FIX: Pin moviepy<2.0 — v2 removed moviepy.editor module

FIX: Robust response.text extraction + improve error visibility

FIX: Switch thumbnail.py from Cycles to WORKBENCH to prevent 300s timeout

FIX: Repair 3 consistently-failing Blender scripts

FIX: Stop Gemini→Claude fallthrough when LLM_PROVIDER=gemini

FIX: Add .dockerignore to exclude 319MB .venv from build context

REVERT: Restore llm_client.py to last known-working state (1f88f40)

FIX: Restore generate_text Gemini call to working form (google_genai.types)

BUILD: Force fresh Docker build after SDK migration

FIX: Correct google-genai SDK import path, remove Claude fallback