Migrate DISJOIN to a registered expander and solve the star-REPLACE portability limitation — Closes #143, #153

[conradbzura]

Summary

Migrate DISJOIN from the emit-time BaseGIQLGenerator.giqldisjoin_sql string special-case (left by epic #114) to a generic registered AST expander, and make the full-row passthrough capability-driven so non-canonical encodings produce portable SQL on engines that lack SELECT * REPLACE.

Register one expand_disjoin against GenericTarget, so every target resolves to it through the registry's generic chain. The expander assembles the same __giql_dj_* WITH-CTE subquery, parses it back into a sqlglot expression, and returns that node for the active target's serializer to render — dissolving the emit-time string special-case. A single capability branch on ctx.capabilities.supports_star_replace selects the passthrough projection form: emit t.* REPLACE (...) on a target that supports it (DuckDB), and the portable t.* EXCEPT (start, end), <recomputed start>, <recomputed end> form otherwise, which every * EXCEPT-capable engine plans. The portable branch is what adds DataFusion support for non-canonical DISJOIN passthrough. Input canonicalization stays owned by CanonicalizeCoordinates (pass 2, #122) — the expander consumes already-canonical 0-based half-open columns and only round-trips the output back into the target's declared encoding.

Fix #153 by aliasing all four projected columns (kc / ks / ke / pos) in every __giql_dj_cuts UNION branch. Previously the bare de-canonicalized end column and the end-cut expression collided under one output name in the default 0-based half-open identity case; DuckDB tolerated the duplicate, but DataFusion rejected it as a non-unique projection name. With every branch aliased, the projection is internally unique on strict engines and behaviour-preserving on DuckDB. As a result, the cross-target oracle's previously-pinned _pending_153 expected-failure is promoted to a real three-target identity test.

Part of epic #137 wave 3; carries the shared ExpanderRegistry.snapshot()/restore() seam that sibling wave-3 PRs also have (dedupe on merge).

Closes #143, #153

Proposed changes

Registry save/restore seam (src/giql/expander.py)

Add the public ExpanderRegistry.snapshot() / ExpanderRegistry.restore() methods, first introduced for these fixtures. snapshot() returns a fresh shallow copy of the (target, operator) → expander registrations; restore() drops all current entries and re-installs exactly the snapshot contents. This lets an isolating test fixture (or a plugin) capture the import-time baseline, mutate the process-wide REGISTRY around a body, and hand the baseline back afterward so the built-in expanders survive a fixture that would otherwise clear() them permanently.

giql.expanders package + DISJOIN expander (src/giql/expanders/__init__.py, src/giql/expanders/disjoin.py)

Add the giql.expanders package whose __init__ auto-imports every submodule via pkgutil.iter_modules, so dropping a <operator>.py into the package registers its expander as an import side effect without editing the package file. Add disjoin.py with the @register(GenericTarget, GIQLDisjoin) expander and its helpers (_build_disjoin_sql, _disjoin_passthrough, _disjoin_output_encoding, _disjoin_resolution), carrying over the original resolution-unpacking and historical diagnostics verbatim. The passthrough is the capability-driven form described in the summary; the identity 0-based half-open case stays a plain t.* fast path.

#153 alias fix (isolated in disjoin.py's __giql_dj_cuts assembly)

Alias kc / ks / ke / pos in all three __giql_dj_cuts UNION branches. This is an isolated, cherry-pickable change: it only adds aliases to existing projections and does not depend on the migration or the capability branch.

GIQLDisjoin.GIQL_EXPAND flip + legacy deletion (src/giql/expressions.py, src/giql/generators/base.py, src/giql/transpile.py)

Flip GIQLDisjoin.GIQL_EXPAND from the disabled sentinel to True, so the ExpandOperators pass replaces the node with the expander's AST. Delete BaseGIQLGenerator.giqldisjoin_sql and the DISJOIN-only generator helpers (_disjoin_resolution, _disjoin_passthrough, _disjoin_output_encoding) plus the now-unused GIQLDisjoin import. Wire import giql.expanders in transpile.py so the registry is populated before the first transpile.

Test updates

Update test_disjoin_transpilation.py, test_canonicalizer.py, and test_expander.py for the registry-driven path, and add the capability-passthrough and snapshot()/restore() coverage. Two execute-on-engine harnesses now transpile with the engine dialect: test_usage_patterns.py (_execute) and coordinate_space/conftest.py (giql_query) pass dialect=engine/dialect="duckdb", because a non-canonical DISJOIN passthrough emits * EXCEPT for the generic target and * EXCEPT is not DuckDB-runnable — the SQL must be shaped for the engine it executes on. Promote the cross-target oracle's test_disjoin_on_datafusion_unsupported_pending_153 expected-failure to test_disjoin_agrees_across_all_targets, a real three-target identity test.

Test cases

#	Test Suite	Given	When	Then	Coverage Target
1	TestDisjoinCanonicalization	A self-mode DISJOIN over a 1-based closed target, transpiled for the generic target	Transpiling to SQL	The passthrough de-canonicalizes via a portable * EXCEPT projection that drops and re-projects the interval columns	Portable passthrough on engines without REPLACE
2	TestDisjoinCanonicalization	A self-mode DISJOIN over a 1-based closed target, transpiled for the DuckDB target	Transpiling to SQL	The passthrough de-canonicalizes via a star REPLACE on the final projection	REPLACE passthrough on DuckDB
3	TestDisjoinTranspilation	A DISJOIN over a registered target	Transpiling to SQL	Emits a parenthesized WITH-CTE subquery with the disjoin_chrom / disjoin_start / disjoin_end columns	Registry-expanded CTE shape
4	TestDisjoinTranspilation	A DISJOIN with the reference omitted	Transpiling to SQL	Defaults the reference to the target set and skips the coverage EXISTS clause	Self-reference coverage skip
5	TestDisjoinTranspilation	A DISJOIN whose reference is a distinct table or shadowing CTE	Transpiling to SQL	Emits the coverage EXISTS clause against the reference	Coverage filter emission
6	TestDisjoinTranspilation	A DISJOIN target or reference using the reserved __giql_dj_ prefix, or an unknown reference name	Transpiling to SQL	Re-raises DISJOIN's historical diagnostics verbatim	Diagnostic parity
7	TestExpanderRegistry	A registry with one entry captured by snapshot()	A second entry is registered afterward	The snapshot holds only the first entry, being a copy not a live view	snapshot() independence
8	TestExpanderRegistry	A snapshot taken, then the registry cleared and a different entry registered	restore() is called with the snapshot	The original entry resolves again and the post-snapshot entry is gone	restore() semantics
9	TestExpandOperatorsPass	A flagged operator with a registered expander	The pass transforms the AST	Dispatches to the registered expander and replaces the node	GIQL_EXPAND dispatch
10	TestNoOpWhenFlagsOff	A DISJOIN query with pass 2 bypassed but pass 3 kept	Comparing canonicalizer output to the expanded baseline	Pass 2 contributes nothing, the byte-identical comparison isolating it	Canonicalizer no-op isolation
11	TestCrossTargetOracleDisjoin	Two overlapping intervals on chr1	The oracle runs generic, datafusion, and duckdb targets	Every target returns identical sub-segments, proving DISJOIN runs on DataFusion and agrees with DuckDB	Three-target identity (#143 / #153)
12	TestCrossTargetOracleDisjoin	Two overlapping intervals	DISJOIN splits them on DuckDB	Returns the expected split sub-segments	DuckDB split correctness