Fix conversion data-loss paths in convert.py

[cursor]

Bug and impact

• convert.py wrote Calibre's temporary HTMLZ output next to the input file as <input-stem>.htmlz and deleted that path after extraction. If a user already had a sibling book.htmlz, conversion overwrote and then deleted user-owned data.
• Default Markdown cleanup also auto-deleted monotonic standalone numeric lines. Books that legitimately contain standalone problem, verse, table, or section numbers such as 1, 2, 3, 4 could lose content before chunking; the manifest then hashed the already-truncated source, so later validation could not detect the loss.

Root cause

• The HTMLZ intermediate path was derived from input_file instead of a private temp directory.
• The LNDS page-number heuristic ran by default, even though standalone numeric lines are ambiguous and may be real book content.

Fix

• Write HTMLZ intermediates under tempfile.TemporaryDirectory(prefix="<book>_htmlz_") and remove the sibling-path cleanup.
• Preserve standalone numeric lines by default unless adjacent to Calibre noise.
• Add explicit --auto-strip-page-numbers for the existing monotonic-page heuristic and keep --strip-page-numbers for aggressive legacy cleanup.
• Update English and Chinese README troubleshooting text.
• Add regression tests for sibling HTMLZ preservation and default numeric-line preservation.

Validation

• python3 -m unittest tests.test_convert -v — 19 passed.
• Post-fix repro: default preserves 1..4; auto_strip_page_numbers=True removes them; conversion uses /tmp/book_htmlz_.../book.htmlz; sibling book.htmlz remains with content USER DATA.
• python3 -m unittest discover -s tests -p 'test_*.py' -v — 207 passed, 3 skipped.
• python3 -m compileall scripts tests — passed.
• Standard Alice mock full pipeline: conversion generated 38 chunks; merge generated HTML/DOCX/PDF. EPUB failed only for the documented Ubuntu Calibre 7.6.0 bytes/str packaging bug.