Add readers for the other formats supported by pandas

[timvink]

Closes #42

New readers

read_parquet, read_orc, read_html, read_xml, read_hdf, read_sas, read_spss and read_stata, each with a matching pd_read_* macro. That completes the list in #42.

Two formats need a bit of handling of their own:

• pd.read_html() returns every table it finds, so read_html() inserts the first one. Use match to select another.
• pd.read_hdf() returns a pd.Series when a Series was stored, which is converted to a frame.

Refactor

Each reader used to be a copy of the same two functions. They are now built from a LOADERS registry by markdown_reader() and dataframe_reader(), so a new format is one line. Adding a reader to select_readers, the JSON schema and the macros happens automatically.

get_keywords() and friends now accept more than one function, because a reader can take keyword arguments from several: read_yaml() has its own encoding on top of the arguments of pd.json_normalize(). Everything a reader does not accept is still passed on to .to_markdown().

Dependencies

• pandas>=1.3 instead of >=1.1, because pd.read_xml() was added in 1.3.
• lxml as a dev dependency, needed by pd.read_html() and pd.read_xml(). It is the only change in uv.lock.

Build times

Adding readers used to make builds slower, because every reader ran its own regex over every page. All selected readers are now searched for in one pattern, so a page is scanned once no matter how many readers there are. On a 15KB page:

	8 readers	16 readers
a regex per reader (before)	1.28 ms	2.54 ms
one regex (this PR)	0.16 ms	0.15 ms

That is about a second saved per 1000 pages compared to master, and select_readers is no longer needed to keep builds fast.

Tables are inserted with a replacement function now, instead of as a regex replacement string. A value like C:\1 path used to be expanded as a backreference and silently became C: path; it is now inserted as it is. The same goes for a value like hi\nthere, which used to break the table across two rows, as tests/fixtures/basic_setup/docs/page_carriage_return.md shows.

Tests

tests/fixtures/pandas_readers covers parquet, orc, stata, sas and xml (with the standard library parser), pandas_readers_html covers read_html() and read_xml() with the default parser.

The read_spss and read_hdf tests are skipped unless pyreadstat and tables are installed, which they are not in CI. Both pass locally with uv run --with pyreadstat --with tables pytest. I left those two out of the dev dependencies because they are heavy and have patchy wheel coverage on python 3.8.

Documented in docs/readers.md, including which readers need an extra package installed. read_spss and read_hdf have no live example there for the same reason.

🤖 Generated with Claude Code

https://claude.ai/code/session_01JCJnK1rDt8rmDsXTFc7AHE