Drop support for Pandoc 1.x/2.x (require Pandoc >= 3.0)

[yihui]

Motivation

Follow-up to #2622. In that PR, the new extract-data-uri.lua filter (which replaced --extract-media) was noted by @bastistician to be more complex than necessary because it had to cope with Pandoc 2.x quirks (unreliable pandoc.mediabag for data URIs, missing pandoc.utils.sha1, missing pandoc.system.make_directory). @yihui agreed it may be time to drop Pandoc 2.x, and @cderv confirmed that current Ubuntu releases ship Pandoc 3+.

Pandoc 3.0 was released over three years ago. This PR raises the minimum required version from 1.14 to 3.0 and removes the version guards and fallbacks that only existed for older Pandoc.

Changes

Lua filters

• extract-data-uri.lua: decode data URIs via pandoc.mediabag.fetch() instead of the hand-rolled base64 decoder, and use pandoc.utils.sha1() unconditionally for the extracted filename. This is the simplification Replace --extract-media with Lua filter to fix _files/ cleanup regression #2622 asked for (net −52 lines in this file).
• shared.lua: make_directory() always uses pandoc.system.make_directory (added in 2.19); dropped the Windows/mkdir -p shell fallbacks. pandoc_type() always uses pandoc.utils.type (added in 2.17).
• Dropped the Pandoc 2.1 version guards from the bundled filters (number-sections, latex-div, pagebreak).

R

• Removed pandoc2.0() and the many pandoc_available("1.x"/"2.x") branches and dead code paths, including: the pandoc-citeproc filter path, --latex-engine, --atx-headers, the markdown_github variant, the subtitle.tex patch (Pandoc < 2.6), fix_horiz_rule() (< 2.8), the accessible-code-block HTML dependency (2.7.3–2.9.2.1), the beamer template patches (< 1.17.3), and the ioslides Windows codepage workaround.
• render() now requires Pandoc 3.0.

Resources / tests

• Removed the now-unused inst/rmd/latex/subtitle.tex and inst/rmd/h/accessibility/empty-anchor.js.
• Updated tests to drop the corresponding old-Pandoc branches and skips.

Net: +69 / −464 lines.

Testing

Affected test files pass locally against Pandoc 3.6.3 (test-pandoc, test-lua-filters, test-md_document, test-github_document, test-powerpoint_presentation, test-pdf_document, test-html_document, test-beamer_presentation, test-ioslides_presentation). The #2604/#2620 regression tests added in #2622 still pass with the simplified filter.

🤖 Generated with Claude Code

[cderv]

My main concern with this idea of dropping support entirely is the support for organization, and entreprise that moves very slowly ! Old system where pandoc 2 would still be used, possibly shipped through an old RStudio.

It is hard to know for sure but this is a worry.

The problem with the lua script is really that ideally we need pandoc 2.8 which is from 2019 - which is 7 years now

rmarkdown/inst/rmarkdown/lua/shared.lua

Lines 78 to 79 in 194c2a1

	if pandocAvailable({2, 8}) and pandoc.system and pandoc.system.make_directory then
	pcall(pandoc.system.make_directory, path, true)

Pandoc 3 is only from 2023 and major bump. 🤷‍♂️

I see this allows to replace some pandoc 2.19 requirements too - so maybe this is oldest we could keep with this change ? pandoc 2.19 requirement , and not pandoc 3 ?

In a way we remove Pandoc 1 support, and we bump minor to latest pandoc 2 minor, and soon will support pandoc 3 only ?

I guess for such change I was thinking we would need to do a deprecation mechanism.

• Do a release by throwing a message or a warning in all system that use Pandoc 2, saying we will change to pandoc 3 , and maybe reach out in an issue to explain to us why pandoc 3 would be a problem ?

R Markdown is widely use. In many private and unknown environnment - so I am extra careful. Especially because R works in a way where using latest available CRAN package is a standard, and locking package version is not widely used (I think).

So I wonder of the possible side effect, and if this is worth the risk.
do we gain much to finally makes this Pandoc version support bump ?
If we want to go through, what about this idea of going in a deprecation cycle like manner ?

I'll ask around at Posit if we know organization with such constraint.