fix: revert gpt-4o logprobs to hardcoded confidence heuristic

fix: scale up small images for gpt-4o vision mode and simplify prompt

feat: calculate true confidence using openai logprobs

style: fix flake8 line length error in pipeline_llm.py

style: run black formatter on pipeline_llm.py to fix CI

feat: apply strong deterministic system prompt and settings for GPT-4o vision pipeline

refactor: remove Pipeline D (Finetuned YOLO) to massively improve inference latency and simplify architecture to 3 pipelines

fix: revert pipeline D to dot clustering because finetuned model predicts individual dots, not characters

fix: revert to gpt-4o because gpt-4-turbo lacks image_url support in the current openai package version

style: wrap LLM system prompt strings to fix flake8 E501 line length CI failure

style: apply black formatting to fix CI pipeline failures

feat: switch OpenAI model from gpt-4o to gpt-4-turbo to test hallucination rates

chore: clean up unused imports from pipelines B and D

fix: correctly handle YOLO bounding boxes as characters instead of individual dots in pipelines B and D

fix: normalize ensemble weights and improve GPT-4o system prompt to reduce hallucination

fix: clean all non-ASCII chars, fix CI (lint + tests), secure API keys

feat: four-pipeline ensemble (CV + Roboflow + GPT-4o + finetuned YOLOv8)

fix: tune blob detector for real images, update benchmark results

feat: scaffold BrailleVision app