fix: comprehensive repo audit — 15 issues fixed

fix: self-audit — 7 issues found and corrected

refactor: Sprint 3b — remove dead code, non-functional provider, unused fields

perf: Sprint 2a — caching, parallel pipeline, async I/O, tests

fix: Sprint 1b — null deref, supports_vision, empty response logging, test

fix: robust JSON parser for VLM responses (extra data, trailing commas, bad escapes)

refactor: use API-reported supports_vision instead of hardcoded model names

fix: recognize mistral-small-latest as multimodal (vision-capable)

feat(pipeline): IIIF-native Sprint 3 — in-memory image fetch for AI analysis

fix(backend): Sprint Fix 2 — data integrity, error handling, robustness

fix(sprint-f3): conformité spec — IIIF langue, ai_raw.json, prompts génériques

fix(sprint-f2): fiabilité — error handling, async I/O, pagination, cache, prompts

fix(sprint-f0): fondations — lazy imports, schémas conformes, pydantic-settings

fix(docker): synchroniser le Dockerfile racine avec infra/Dockerfile

fix(providers): désactiver vertex_api_key + ajouter OCR Mistral dédié

fix(vertex): corriger le routing API pour VERTEX_API_KEY — ajouter vertexai=True

fix(mistral): refactoring structurel complet du provider Mistral

fix(mistralai): detect v0.x package + force pip --upgrade in Dockerfile

fix: 3 production bugs — mistralai import, page ID collision, provider logging

feat: multi-provider dynamique — ajout Mistral/Pixtral, détection auto des providers

feat(sprint2-session-c): premier appel Google AI → master.json valide

Sprint 2 Session A — Couche providers Google AI (3 modes d'auth)

Sprint 1 — Setup initial : arborescence, schémas Pydantic, profils JSON, prompts, tests