v1.2.2

Adds PowerPoint and Excel input through Docling, closes the remaining
symlink and hard-link gaps in directory scans, and fixes content loss in
the text lane, picture handling across formats, and several metadata and
CLI edge cases found in review.

Added

• PowerPoint (.pptx) and Excel (.xlsx) input (#59, #63, issue #25).
  Both convert through Docling, so they need the high-fidelity extra:
  pip install "any2md[high-fidelity]". The deck or workbook title is
  the H1, each slide or sheet is an H2 section, sheets render as Markdown
  tables, and hidden sheets and empty slides add no section. A slide's
  title comes from its title placeholder wherever it sits on the slide.
  Without Docling, each file fails with an install hint and the batch
  continues. Legacy binary .ppt and .xls files are reported with a
  hint to save them in the newer format.
• Pictures follow one rule in every format (#63, #65). A picture with
  a caption, Word caption paragraph, Docling description, or alt text gets
  a *Figure: text* line. With --save-images, each picture is written
  to images/<output stem>/imgN.<ext> and linked below its figure line,
  for PDF, DOCX (Docling and mammoth lanes), PPTX, and XLSX. The mammoth
  lane no longer inlines base64 images. Images are published only after
  the Markdown is written, and --force regenerates them and removes
  stale ones.

Security

• Scanned files are bound to no-follow opens (#57, #62). A file found
  by a directory scan is opened relative to the scan root without
  following a symlink at any path component, checked as a regular file on
  the open descriptor, and read through that descriptor. A file or parent
  directory swapped for a symlink after the scan can no longer redirect
  the read. Hard-linked files found by a scan are skipped with a
  SKIP (hard link) line. Files named on the command line still convert.
• OOXML archive guard (#59, #64, #65). DOCX, PPTX, and XLSX archives
  are checked before any parser reads them: total unpacked size, per-part
  compression ratio, and member count. XLSX workbooks also have per-part
  and total limits on the XML that openpyxl parses, applied by format and
  relationship type so renamed parts cannot slip past them.
• Docling and library warnings stay off the terminal (#59, #63, #65).
  Warnings raised by Docling during a conversion are counted toward
  --strict and the warning total instead of printing raw text. Warnings
  from other libraries keep their normal handling.

Fixed

• Content loss in the text lane (#55, #61). Stage T10 dropped short
  lines that did not follow a finished sentence, including the opening
  title line of a TXT file. It now follows the documented rule, keeps
  short lines after code blocks, and never drops figure or image lines. A
  standalone title-case or ALL-CAPS first line of a TXT file becomes the
  H1 and the title, and a TXT document gets one H1.
• Fenced code is left alone (#55). Text-lane stages and the shared
  cleanup stages (whitespace, quotes, footnotes, entity decoding, link
  stripping, validation) no longer rewrite or drop lines inside fenced
  code. A ~~~ fence counts the same as a backtick fence.
• Block structure is read the way CommonMark reads it (#49, #52, #58,
  #61). Heading ranking and fence detection share one scanner that models
  fences, indentation and tabs, list items, block quotes, lazy
  continuation, setext headings, thematic breaks, and raw HTML blocks.
  Headings inside code, quotes, or HTML blocks are never re-ranked, and a
  form feed before a block marker no longer hides it. The scan runs once
  per distinct text, which roughly halves text-lane time on large bodies.
• TOML dates and times in --meta-file and .any2md.toml (#54) are
  written as quoted ISO 8601 strings at any depth instead of failing the
  file. A list of tables now fails the file with a clear message instead
  of writing Python-style text.
• Conflicting --meta keys (#40) such as organization=X plus
  organization.name=Y now fail with a clear error instead of silently
  dropping one value. Thanks to @Qrytics for the fix.
• Output name collisions (#60, #64). Two inputs in one run that map to
  the same .md name, such as a/x.pdf and b/x.pdf under -r, no
  longer overwrite or silently skip each other: the later one fails with
  an output name collision line. A name is claimed only when a file is
  converted, and the same file named twice converts once.
• Text-lane setext artifacts (#64). A === or --- line under text
  in pymupdf4llm or mammoth output is kept as a rule or plain text rather
  than read as an extra heading by the validator.
• Large scanned files on macOS (#64, #65) are read in bounded pieces,
  so a file over 2 GB with a raised --max-file-size no longer fails.
• Legacy-format skips are counted in the run summary (#64).

Changed

• CI dependencies (#11, #12, #32, #33): actions/checkout 7.0.0,
  actions/upload-artifact 7.0.1, actions/download-artifact 8.0.1, and
  github/codeql-action 4.36.2, each pinned by commit SHA.
• Code-scanning cleanup (#56): empty except blocks, unused globals,
  dead comparisons, and test import style, with no behavior change.