chore(scratch): archives des scripts exploratoires de choix d'OCR

Conservés comme trace de recherche — non documentés, non factorisés,
ne pas dépendre de ce dossier depuis le code de production.

- test_glm_ocr.py          : benchmark GLM-OCR 0.9B (écarté pour
                             faiblesse sur dp_libelle, praticien et
                             colonne Recodage).
- test_got_ocr.py          : tests GOT-OCR2.0 (échec sur tableaux
                             denses à en-têtes verticaux).
- test_paddle.py           : tentative PaddleOCR (incompatible avec
                             paddlepaddle installé).
- test_surya.py            : tentative Surya (incompatible
                             transformers 5.6).
- test_qwen_vl.py          : Qwen2.5-VL-7B (excellent mais 220s/page,
                             écarté faute de VRAM et vitesse).
- test_qwen_vl_3b.py       : Qwen2.5-VL-3B (retenu, 3s/page, qualité
                             > GLM-OCR sur les champs critiques).
- test_prompt_ab.py        : A/B test prompts Accord/Désaccord.
- test_prompt_crop*.py     : prompts + crop ciblé checkboxes (échec
                             → module pipeline/checkboxes.py).
- test_prompt_recueil_*.py : prompts page recueil (consignes verbeuses
                             dégradent la sortie, cf. discussion).
- README.md                : index du dossier.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

scratch/test_glm_ocr.py

@@ -0,0 +1,101 @@
+"""Test GLM-OCR 0.9B sur une page OGC isolée."""
+import sys
+import time
+import torch
+from transformers import AutoProcessor, AutoModelForImageTextToText
+from pdf2image import convert_from_path
+import os
+
+PDF_PATH = sys.argv[1] if len(sys.argv) > 1 else "2018 CARC/OGC 7.pdf"
+PAGE_NUM = int(sys.argv[2]) if len(sys.argv) > 2 else 1
+MODE = sys.argv[3] if len(sys.argv) > 3 else "text"  # text | table | json
+OUTPUT_MD = sys.argv[4] if len(sys.argv) > 4 else "test_glm_result.md"
+
+# Prompt selon mode
+JSON_SCHEMA_OGC = """Extrais les informations de cette fiche médicale OGC et réponds en JSON strict :
+{
+  "etablissement": "",
+  "finess": "",
+  "date_debut_controle": "",
+  "n_ogc": "",
+  "n_champ": "",
+  "dates_sejour": "",
+  "codage_etablissement": {
+    "dp": "",
+    "dr": "",
+    "das": [{"code": "", "position": ""}]
+  },
+  "codage_recodage": {
+    "dp": "",
+    "dr": "",
+    "das": [{"code": "", "position": ""}]
+  },
+  "actes_etablissement": [{"code": "", "position": ""}],
+  "actes_recodage": [{"code": "", "position": ""}],
+  "ghm_etablissement": "",
+  "ghs_etablissement": "",
+  "ghm_recodage": "",
+  "ghs_recodage": "",
+  "accord_desaccord": "",
+  "praticien_conseil": ""
+}"""
+
+PROMPTS = {
+    "text":  "Text Recognition:",
+    "table": "Table Recognition:",
+    "json":  JSON_SCHEMA_OGC,
+}
+prompt_text = PROMPTS[MODE]
+
+print(f"PDF: {PDF_PATH}  page: {PAGE_NUM}  mode: {MODE}")
+print(f"--- Chargement GLM-OCR 0.9B ---")
+t0 = time.time()
+MODEL_PATH = "zai-org/GLM-OCR"
+processor = AutoProcessor.from_pretrained(MODEL_PATH, trust_remote_code=True)
+model = AutoModelForImageTextToText.from_pretrained(
+    MODEL_PATH,
+    torch_dtype="auto",
+    device_map="auto",
+    trust_remote_code=True,
+)
+print(f"Modèle chargé en {time.time()-t0:.1f}s")
+print(f"VRAM utilisée : {torch.cuda.memory_allocated()/1e9:.2f} Go")
+
+print(f"--- Conversion PDF page {PAGE_NUM} ---")
+pages = convert_from_path(PDF_PATH, 300, first_page=PAGE_NUM, last_page=PAGE_NUM)
+tmp = f"/tmp/glm_page_{PAGE_NUM}.png"
+pages[0].save(tmp, "PNG")
+print(f"Image: {tmp} ({pages[0].size})")
+
+messages = [{
+    "role": "user",
+    "content": [
+        {"type": "image", "url": tmp},
+        {"type": "text", "text": prompt_text},
+    ],
+}]
+
+print(f"--- Génération (mode={MODE}) ---")
+t0 = time.time()
+inputs = processor.apply_chat_template(
+    messages,
+    tokenize=True,
+    add_generation_prompt=True,
+    return_dict=True,
+    return_tensors="pt",
+).to(model.device)
+inputs.pop("token_type_ids", None)
+
+generated_ids = model.generate(**inputs, max_new_tokens=8192)
+output_text = processor.decode(
+    generated_ids[0][inputs["input_ids"].shape[1]:],
+    skip_special_tokens=False,
+)
+print(f"Génération en {time.time()-t0:.1f}s ({len(output_text)} chars)")
+
+with open(OUTPUT_MD, "w", encoding="utf-8") as f:
+    f.write(f"# {os.path.basename(PDF_PATH)} — page {PAGE_NUM} — mode={MODE}\n\n")
+    f.write(output_text)
+print(f"--- Sauvé dans : {OUTPUT_MD} ---")
+print("\n--- Aperçu (1000 premiers chars) ---\n")
+print(output_text[:1000])