Add source session id to each sin and force smart-quote wrapping

Make sin meta a verbatim user quote, not a paraphrased citation

Switch back to Inference Providers (Qwen3.5-9B)

Bump bulletin max_new_tokens + retry until all 3 sins emit

Swap model to Qwen3.5-9B for faster cold-start on ZeroGPU

Prime CUDA + stream weights to avoid NVML assert on ZeroGPU

Run Qwen3.6-35B-A3B locally on ZeroGPU instead of Inference API

Parse both trace formats: Claude-Code (sessions/<date>/) and pi-sessions (root-level message events with toolCall/toolResult roles)

Initial bulletin app: Qwen3.6-35B-A3B → vintage 1080×1440 retro report