fix: preserve panel labels and figure_inner_text in post-reference backmatter zone

This commit is contained in:
Research Assistant 2026-06-17 22:36:51 +08:00
parent 06f3c3e948
commit af11f6d985
2 changed files with 24 additions and 0 deletions

View file

@ -4682,6 +4682,12 @@ def normalize_document_structure(
new_role="media_asset",
reason="image/media block in post_reference_backmatter_zone promoted to media_asset",
)
# ponytail: panel_label + figure_inner_text guard; if other specialized roles
# appear in backmatter zone, extend this list
if block.get("marker_signature", {}).get("type") == "panel_label":
continue
if block.get("role") == "figure_inner_text":
continue
elif role not in _BACKMATTER_NON_BODY_ROLES:
old_role = block.get("role")
if old_role != "backmatter_body":

View file

@ -143,3 +143,21 @@ def test_keep_formal_caption_seed_for_numbered_display_legend() -> None:
}
assert _should_keep_formal_caption_seed(block) is True
def test_panel_label_in_post_ref_zone_not_converted_to_backmatter() -> None:
"""Panel labels like (a) in post_reference_backmatter_zone stay figure_inner_text.
Regression: guard in normalize_document_structure prevents backmatter
normalization from overwriting figure_inner_text role.
"""
from paperforge.worker.ocr_roles import assign_block_role
block = {
"block_label": "text",
"block_content": "(a)",
"block_bbox": [90, 90, 120, 120],
"page": 10,
}
role = assign_block_role(block, page_blocks=[], page_width=1200, page_height=1600)
assert role.role == "figure_inner_text"