extract_pages_markdown_bytes() flags every page as needs_ocr on a plain text-based PDF, contradicting detect_pdf_bytes()

Summary

extract_pages_markdown_bytes() classifies every page as needs_ocr=True with empty
markdown on a PDF that detect_pdf_bytes() / process_pdf_bytes() classify as
pdf_type="text_based" with pages_needing_ocr=[] — on the exact same input bytes. The two
entry points disagree completely on whether the document needs OCR.

This makes extract_pages_markdown_bytes() unusable for per-page reliability checks right
now: it reports OCR-needed for pages that plainly are not scanned/image-based.

Environment

• pdf-inspector 0.2.6 (PyPI wheel, manylinux_2_17_x86_64)
• Python 3.12, Linux x86_64

Minimal repro

# pip install pdf-inspector pymupdf
import fitz
import pdf_inspector

doc = fitz.open()
rect = fitz.Rect(72, 72, 523, 750)
for i in range(3):
    page = doc.new_page()
    page.insert_textbox(rect, f"Page {i + 1}. This is a normal paragraph of body text.")
pdf_bytes = doc.tobytes()
doc.close()

det = pdf_inspector.detect_pdf_bytes(pdf_bytes)
print("detect_pdf_bytes:", det.pdf_type, det.confidence, det.pages_needing_ocr)

per_page = pdf_inspector.extract_pages_markdown_bytes(pdf_bytes)
print("extract_pages_markdown_bytes.pages_needing_ocr:", per_page.pages_needing_ocr)
for p in per_page.pages:
    print(f"  page={p.page} needs_ocr={p.needs_ocr} markdown_len={len(p.markdown)}")

Actual output

detect_pdf_bytes: text_based 0.5 []
extract_pages_markdown_bytes.pages_needing_ocr: [1, 2, 3]
  page=0 needs_ocr=True markdown_len=0
  page=1 needs_ocr=True markdown_len=0
  page=2 needs_ocr=True markdown_len=0

(confidence here is 0.5, not high — but pages_needing_ocr=[] from detect_pdf_bytes()
already says no page needs OCR, directly contradicting extract_pages_markdown_bytes()
flagging all three.)

(Note pages_needing_ocr on the per-page result also appears to be 1-indexed — [1, 2, 3]
for a 3-page doc — while PageMarkdown.page is documented as 0-indexed in the .pyi stub;
worth double-checking that indexing is consistent across the per-page API surface too.)

Expected behavior

extract_pages_markdown_bytes() should agree with detect_pdf_bytes() /
process_pdf_bytes() on whether a given text-based PDF needs OCR — at minimum, it shouldn't
flag every page of a plain multi-page text document as needing OCR with empty Markdown.

Start by running the minimal reproduction against extract_pages_markdown_bytes(), then compare its result with detect_pdf_bytes() and process_pdf_bytes() on the same bytes. Trace the per-page extraction and OCR classification entry points; done means a plain text-based multi-page PDF returns Markdown, does not flag every page for OCR, agrees with the other APIs, and has consistent page indexing.