fix: switch model back to gemini-2.5-flash per user request

fix: initialize raw variable to avoid UnboundLocalError during 429 quota exceptions, and switch to gemini-2.0-flash for higher daily limit

fix: use strict=False in json.loads to gracefully handle literal newlines

chore: add raw output to fallback rationale for debugging

fix: prompt formatting to avoid literal newlines in json strings

feat: migrate to google-genai SDK for reliable JSON response

fix: robust JSON extraction for gemini-2.5-flash responses

feat: upgrade to gemini-2.5-flash model

fix: move system_instruction to model constructor (Gemini SDK compat)

feat: Vera bot - FastAPI + Gemini 2.0 Flash, 4-context composer, auto-reply detection