Let extension bundles declare catalog providers

[timsaucer]

Which issue does this PR close?

Closes #1676.

Part 4 of 4.

1. scalar, aggregate, and window functions
2. physical optimizer rules
3. tables and table functions
4. This PR -> catalog providers

Rationale for this change

The last field, and the last of the three getters that take the session. It also removes a small structural wart: register_catalog_provider did its capsule import and its insert in one function with no split point, so the bundle path could not reuse it without duplicating the four-way branch that decides what a caller actually handed over.

What changes are included in this PR?

SessionExtensionComponents.catalog_providers, as (name, provider) pairs.

resolve_catalog_provider, shared. The import half of register_catalog_provider moves to a private helper that both it and _resolve_extension_catalogs call, so the two paths accept exactly the same shapes — an object exposing __datafusion_catalog_provider__, a bare capsule, a PyCatalog, or a Python object implementing the provider interface. PyCatalog::new was not a substitute: it wraps unconditionally in RustWrappedPyCatalogProvider and never checks for a capsule, so routing the bundle path through it would have silently bypassed the FFI fast path.

Catalogs replace; tables do not. This is the one place in the stack where two neighbouring fields take opposite rules, so it is worth being explicit. register_catalog returns whichever provider it displaced rather than refusing, and the default datafusion catalog always exists — so replacing one is the usual way a library backs a session with its own metadata, and refusing it would refuse the main use case. A duplicate table, by contrast, is an error, because DataFusion refuses one. Only two bundles claiming the same catalog name within a single call is refused, with both named. table-providers.md states the contrast where a reader meets it.

The catalogs commit is a _commit_extensions parameter. Once the providers are imported there is nothing left that can fail, so the commit step needs no primitive of its own.

The capstone test. test_four_libraries_install_in_one_call installs four independently declared bundles — functions, rules, a table plus a table function, a catalog — in one with_extensions call, then runs a single query that reads the declared table and the declared catalog through the declared scalar function, and asserts both optimizer rules fired while planning it. That is the thing #1676 asked for, exercised across a real FFI boundary.

Skill update. Rule 2 of .ai/skills/ffi-capsule-protocol/SKILL.md now carries the split this stack is built on: which components a bundle hands over unwrapped and why, pointing at RecordingTableFunction as the assertion rather than the description. Rule 6 gains the other convention — a new kind of component means a new resolve step returning an opaque carrier and a new _commit_extensions parameter whose commit cannot fail, never a fallible commit — along with the single table exception and a note not to add a second without the same justification.

Are there any user-facing changes?

One new optional field on SessionExtensionComponents, defaulting to (). register_catalog_provider is refactored but behaviourally identical — same accepted shapes, same result. No hook signatures change and no wire format changes, so there is no upgrade-guide entry and no api change label.

Taking the stack as a whole, SessionExtensionComponents grows from two fields to nine, all optional and all additive, and PhysicalOptimizerRuleExportable becomes importable from datafusion as well as datafusion.context.

Review notes

The replace-vs-refuse asymmetry is inherited, not chosen. Both rules are DataFusion's. If you would rather with_extensions normalized them — refusing both, or replacing both — that is a defensible position, but it would mean the bundle path behaving differently from the register_* method it is meant to replace, which seems worse than the asymmetry.

What this stack does not do, stated once so it is on the record: with_extensions still cannot install everything a library provides. A config extension has to reach SessionConfig before the context exists, which no bundle hook can reach — dfx_engine in #1721 needs exactly that, and its build_session will still have a step for it.

Follow-up, once #1721 and this stack have both landed. dfx_udfs deliberately ships no bundle hook and carries test_this_library_cannot_be_installed_as_a_bundle, whose premise #1738 inverts. Whichever lands second should convert it and flip that test. Nothing in this stack touches examples/distributed/.

🤖 Generated with Claude Code