Make the Q&A system prompt accent-correct PT-BR

Update vision GGUF variant test to v1 fine-tune defaults

Update QA default-model test to v4 fine-tune

Isolate uploads per browser session (no shared/persisted data)

Move text-gen files to finetune/text/; drop finetune tests

Merge branch 'vivamais-vision-v1' into vision-into-main

Consolidate vision finetune files under finetune/vision/

Split Stage-2 eval into recall-on-populated vs over-fill

Fix generic PT-BR refusal scoring

Wire v3 train, merge, GGUF export, and runtime selection

Add accent-sensitive vision eval harness

Add LLM-as-judge scoring for teacher labels

Add real-corpus builder and stage-2 contract coercion

Accent-correct synthetic pools, loud fonts, photo augmentation

Re-accent extraction glossary and prompts (shared constant)

Purge legacy v1 finetune/eval; re-ground on current domain

Add shared PT-BR conversation eval suite

Expand Viva Mais text eval suite

Cap image size for the llama.cpp vision slicer

Prefer the torch transcriber over CTranslate2 on the GPU backend

Shrink QA context to fit the T4 alongside other models

Update smoke tests for English-default UI chrome

Bake the live Space's production models as defaults

Drop transformers-v46 backend; route cuda runtime to llama.cpp

Show honest counts on Resumo and translate presenter chrome

Answer Q&A in the selected UI language (EN default)

Add Viva Mais travel vocabulary SFT bucket

Align MiniCPM5 v1 QA grounding behavior

Expose a persisted-work counter on every upload status tick

Render Q&A answers as Markdown in the chat dock

Add Nemotron Embed VL adapter and wire it per backend

Answer dashboard questions with hybrid retrieved context

Index the Q&A corpus as a final pipeline stage

Add deterministic MockEmbedder for offline retrieval tests

Store corpus chunks with embeddings in both repositories

Add Embedder port, corpus chunker, and hybrid retrieval logic

Show the thinking trace greyed out under a Pensando label

Stream Q&A phases through the dashboard interactor

Expose phased streaming on the TextModel port

Tell the Q&A model the asker is the travel agent

Recognize the agent by alias and rename her at ingest

Raise Q&A max output tokens to 2048

Enable MiniCPM5 thinking for Q&A by default

Support prompt-embedded think tags in reasoning strippers

Strip generated MiniCPM thinking prefixes

Make MiniCPM5 SFT mix weights configurable

Skip unsafe public SFT rows

Prepare redacted MiniCPM5 text SFT data

Add Viva Mais QA eval gate

Switch Q&A text model to MiniCPM5-1B-PTBR-v3