fix: handle plain ibis backends in join backend rebinding

[hachej]

Summary

• Fixes ValueError: Don't know how to handle type <class 'ibis.expr.types.relations.Table'> when using joins on plain ibis backends (Snowflake, Databricks, etc.)
• _rebind_join_backends now gracefully skips xorq-specific node traversal when the expression tree contains plain ibis tables
• Adds regression test reproducing the exact YAML join scenario from the issue

Closes #221

[hachej]

Claude Code Review

Verdict: Approve — the fix is correct, minimal, and well-tested.

ops.py — _rebind_join_backends

Correctness: The two try/except blocks are properly scoped:

• ImportError guard on import — handles the (unlikely) case where xorq is not installed at all. Good defensive coding.
• (ValueError, TypeError) around walk_nodes — catches the exact ValueError from xorq's to_node() when it encounters a plain ibis.expr.types.relations.Table. The TypeError is a reasonable defensive addition for future xorq changes. Scope is tight (only wraps the walk_nodes call, not the op().replace() logic below).

Semantics: Skipping rebinding for plain ibis backends is correct — rebinding only exists to unify xorq Backend instances created by separate from_ibis() calls. Plain ibis backends (Snowflake, Databricks) share a connection natively and don't need this.

Edge case — mixed xorq left + plain ibis right: If walk_nodes succeeds on the left (xorq-wrapped) but the right table is plain ibis, the replacer function would still run on the right via right_tbl.op().replace(replacer) at line 3755. However, since isinstance(op, xorq_rel.DatabaseTable) would be False for plain ibis ops, the replacer is effectively a no-op. Safe but worth noting for future maintainers.

test_plain_ibis.py — test_rebind_join_backends_with_plain_ibis_table

Good: Directly calls the static method with raw plain ibis tables, hitting the exact ValueError code path from GH-221. Asserts identity (is) which confirms no unnecessary copying. This test fails on main without the fix — confirmed true regression coverage.

Minor nit: create_table with inline DataFrames creates persistent tables on the module-scoped plain_ibis_con. Since the fixture is scope="module", a table name collision with another test that reuses the connection could theoretically occur. Not a real risk given current test names, but "rebind_l" / "rebind_r" are sufficiently unique.

No concerns

• No regressions to xorq-backed join paths (happy path unchanged)
• No serialization changes (Codex flagged a pre-existing issue, tracked in Serialization: missing cardinality in pre-existing tagged joins defaults to 'one' #223)

🤖 Review by Claude Code

[goober]

Facing the same issue for Athena.
@hachej are you planning to merge this PR or are there outstanding issues that you need any assistance with?