Complete the DataFusion target and finalize capability-driven canonicalization — Closes #145

[conradbzura]

Summary

Choose the coordinate-canonicalization emit strategy from the active target's capabilities instead of hardcoding a DuckDB-only star REPLACE, completing the DataFusion target (epic #137). The CanonicalizeCoordinates pass now receives the target's Capabilities, and both the wrapper-CTE projection and the NEAREST row passthrough emit * REPLACE when supports_star_replace holds (DuckDB) and the portable * EXCEPT (start, end), <start>, <end> form otherwise (the generic / DataFusion family). This removes the two remaining hardcoded REPLACE assumptions, so a non-canonical coordinate encoding transpiles to engine-runnable SQL on DataFusion. DuckDB output is byte-unchanged; the generic/DataFusion EXCEPT form is row-equivalent (column order differs, since EXCEPT re-appends the recomputed interval columns) and is verified end-to-end on the real DataFusion engine across all four coordinate encodings, custom interval-column names, and strand. DataFusion serialization is finalized on the generic sqlglot output and its capability values promoted from provisional to verified. The SELECT *-over-correlated-NEAREST column leak (#160) remains deferred to the later query-level seam (#146).

Closes #145

Proposed changes

• canonicalizer.py — canonicalize_coordinates(expression, capabilities=None) threads capabilities into _canonical_projection, which branches * REPLACE vs portable * EXCEPT; None preserves the historical REPLACE form for direct callers.
• transpile.py — passes target.capabilities into the pass.
• generators/base.py — _nearest_passthrough gains the same capability gate (resolving its TODO(#142)).
• expanders/nearest.py — _distance_and_filters threads ctx.capabilities to the passthrough from both the LATERAL and decorrelated-fallback forms.
• targets.py — DataFusionTarget docstring promoted from "provisional" to verified, recording which capability values the cross-target oracle validates.

Test cases

#	Test Suite	Given	When	Then	Coverage Target
1	TestCanonicalProjectionCapabilities	A non-canonical operand and capabilities lacking star-replace	canonicalize_coordinates is called directly	The wrapper emits the portable * EXCEPT form	EXCEPT branch via the public capabilities arg
2	TestCanonicalProjectionCapabilities	A non-canonical operand and REPLACE-capable capabilities	canonicalize_coordinates is called directly	The wrapper emits * REPLACE	REPLACE branch via explicit capabilities
3	TestCanonicalProjectionCapabilities	A non-canonical operand and no capabilities argument	canonicalize_coordinates is called	The wrapper defaults to * REPLACE	capabilities=None historical-default contract
4	TestCanonicalProjectionCapabilities	An AST already canonicalized once	The pass runs a second time	No second wrapper CTE is synthesized	Idempotence of the EXCEPT wrapper
5	TestCanonicalProjectionCapabilities	A non-canonical DISJOIN	Transpiling with dialect="datafusion"	The wrapper and passthrough emit * EXCEPT with no leaked operator	DataFusion dialect threads its capabilities
6	TestNearestTargetCanonicalization / TestBaseGIQLGenerator	A non-canonical NEAREST target	Transpiling for DuckDB and generic	DuckDB emits * REPLACE, generic emits * EXCEPT, sharing the distance CASE	Wrapper + passthrough both branches via transpile
7	TestDisjoinEncodingSweepOnDataFusion	Overlapping intervals in each of the four encodings	The DISJOIN oracle runs all three targets	Every target returns identical de-canonicalized sub-segments	EXCEPT canonicalization on the real DataFusion engine, all encodings
8	TestNearestEncodingSweepOnDataFusion	A non-canonical NEAREST target per encoding (incl. custom columns)	The oracle runs datafusion vs duckdb	Both return the same nearest interval and distance	EXCEPT NEAREST passthrough on DataFusion, all encodings + custom columns
9	TestDataFusionSchemaAxes	Custom-column DISJOIN and stranded MERGE	The oracle runs all three targets	Every target agrees	Custom-column and strand axes on DataFusion