data: +baidu/Unlimited-OCR na RTX 6000 (4º modelo OCR: 425ms/pág, faixa DeepSeek)

data: re-consolidar _ALL do raw completo (OCR agora cobre B200/L40S/Spark; serving +10 linhas B200/Spark)

fix(data): ocr_pages_ALL normalizado (schema único, modelos reais, +DeepSeek-OCR/-2 vs NuMarkdown na RTX)

data: reconsolidar *_ALL.csv com RTX PRO 6000 (serving fixado + embed/rerank/stt/vision/ocr); +rerank_ALL +ocr_pages_ALL

OCR Spark completo: 1/10/50 × conc 1/4/8/16 (12 pts) — gap fechado

OCR L40S atualizado (10 pts; 50p conc8/16 em coleta)

OCR 4090 completo: +conc16 (11 pts, fechado o gap via TMPDIR fix no tupi)

OCR: +RTX PRO 6000 (12 pts, 99-101 tok/s)

v0.6: +B200 OCR tensor-parallel (TP 1/2/4 — achado: TP piora OCR-8B)

OCR L40S ampliado: conc 1/4/8/16 (10 pts; satura suave 48→44 tok/s)

OCR NuMarkdown completo: RTX 4090 + L40S + Spark (19 pontos, n_ok>0)

v0.5: B200 (Blackwell + tensor-parallel) + DGX Spark (GB10) + OCR multi-página + L40S-NGC; coluna engine