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/

Correct v1 eval with fixed metric + write v1 model card

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

Record v1 synthetic-only held-out eval: base vs v1

Record v3 vision eval results: base vs merged-v3

Fix generic PT-BR refusal scoring

Add Rio 3.1 teacher eval harness

Add accent-sensitive vision eval harness

Add LLM-as-judge scoring for teacher labels

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

Add shared PT-BR conversation eval suite

Expand Viva Mais text eval suite

Record rejected MiniCPM5 candidate v2 eval

Record MiniCPM5 candidate v1 evals

Align MiniCPM5 v1 QA grounding behavior

Record MiniCPM5 thinking eval comparison

Strip generated MiniCPM thinking prefixes

Record rejected MiniCPM5 candidate two eval

Record rejected MiniCPM5 candidate eval

Record MiniCPM5 base QA eval

Add Viva Mais QA eval gate

Fix lint and format drift in modal and evals scripts

Add combined Stage 1 and Stage 2 v2 eval comparison

Record Stage 2 vision payload eval results (v2 val)

Harden Stage 2 coercion scoring for partial model JSON

Add Modal Stage 2 vision payload eval pipeline

Fix v2 merge export paths and Stage 1 Modal evals (T13)

Fix lint and format debt left by the T12 eval pipeline

Add DSPy-aligned v2 vision fine-tune pipeline (T12)

Apply ruff format across the repository

Add Modal eval and merge harness for the tuned vision model