docs: bring the READMEs up to date with M2

[alxmrs]

The front page still said "M0 landed — the scalar core is implemented. The rest is in progress." That was true in July. Since then ddx-core reached 0.2.1 on crates.io, ddxdb 0.1.0 shipped to PyPI, the DataFusion adapter became publishable, and the JAX oracle landed.

The bigger gap: no way in for Python

The only runnable example was a Rust snippet. pip install ddxdb now works and is the lowest-friction entry point by a wide margin, so it leads:

import ddxdb
ddxdb.rewrite_sql("SELECT grad(sin(x), x) AS d FROM t")
# 'SELECT (cos(x)) AS d FROM t'

...followed by the point that makes it interesting — rewrite_sql is text in, text out, so it already works with any engine that accepts SQL, not just the ones ddx ships an adapter for. Rust follows.

Also

• A status table saying where each crate actually is, with registry links. This replaces per-line milestone tags in the layout block, which meant status lived in two places and drifted in both.
• A correctness section. "Fail loud, never silently wrong" is the claim this project stands on, and nothing on the front page backed it up. It now names the three things that do: the JAX oracle, the property soak, and the conventions pinned from both sides where ddx and JAX differ on purpose.
• Dropped ddx-duckdb/ from the layout — it listed a directory that does not exist.
• Badges for the two published artifacts, so current versions are visible without leaving the page.

Every claim verified, not asserted

A README is the easiest place for aspirational text to survive, so each one was checked:

claim	how
the Rust snippet	compiled and run against ddx-core 0.2.1 from crates.io, in a scratch crate outside the workspace
the Python snippet	run against ddxdb installed from PyPI into a clean venv with --no-cache
"a soak that runs nightly"	nightly-fuzz.yml, cron 0 16 * * *
badge URLs	both return 200
CONTRIBUTING.md link	exists

The Rust check is the one worth calling out: building it outside the workspace is what proves a reader copying it gets working code, rather than code that only compiles here.

Also: the ddxdb page on PyPI

Two fixes to python/ddxdb/README.md, which is the long description PyPI renders. They ship with the next ddxdb release, not this merge.

It compared against a DuckDB function that does not exist. The client-side rewrite was justified by contrast with "DuckDB's in-database ddx('<sql>') table function" — an M5 design that ships nowhere. A reader landing on PyPI has no way to know that, so it reads as a comparison between two available options and sends them looking for one. The property is real and kept: the test proves a temp table, which exists only on the caller's connection, is visible to the rewrite. It is now stated on its own terms, with the alternative framed as a design rather than a feature.

It had not caught up with supported_functions() — the section was still headed "One other function." Now documented, with the caveat that a name being present does not by itself make an expression differentiable, so the typed errors remain the general answer to "can ddx handle this?".

Docs only — no behaviour changes. 130 oracle tests and 51 wheel tests unchanged.

🤖 Generated with Claude Code

[alxmrs]

Can we say it's necessary, not load bearing?

[alxmrs]

This is too much text about the oracle test suite. Let's omit a lot of this or make it shorter. It's distracting for the purposes of the contrib guide.

[alxmrs]

I think we can get rid of this whole section (enumerating the checks). This is pretty self evident, no?

[alxmrs]

We can omit this cause it is only needed to be done once, and it is done.

[alxmrs]

Wait, we don't have DuckDB support yet.

[alxmrs]

Let's at least say "eventually".

[alxmrs]

LGTM for now.