feat(jdbc): single edge→target JOIN for unbounded g.V().out().has() (stacked on #172) - #173

Stacked on #172 (feat/unbounded-adjacency-scan) — review/merge that first; this PR's diff is only the delta.

What

#172 eliminates the source-vertex scan by rewriting unbounded g.V().out(L).has(k,v) to g.E().hasLabel(L).inV().has(k,v) — an edge scan plus a separate deferred target fetch (two query phases). This PR collapses that into one SQL JOIN when the backend supports it:

select v.* from edges e join <target> v on e.<endpoint> = v.id where <target has()>

DB-side filtering, a single round-trip, and only matching rows returned/materialized (vs. reading every edge into the app first).

How

• SearchVertexQuery.allSources — "match all sources", i.e. no source-id bound. Distinct from an empty vertex list, which aborts (= match none).
• RowEdgeSchema.getJoinSearch / RowController fallback honour allSources (edge-label predicates only, never abort), reusing the perf(jdbc): adjacency join push-down for vertex-return hops (out/in/both + filter/order/limit) #167 join push-down.
• SupportsUnboundedAdjacency marks RowController as allSources-capable. The strategy swaps in the JOIN start step only when a controller advertises it; every other backend keeps feat(core): skip source-vertex scan for unbounded g.V().out()/in()/both() via edge-scan rewrite #172's backend-agnostic g.E().<adjacency>() rewrite — no ES regression.
• UniGraphUnboundedAdjacencyStep — a start step implementing Orderable / ReceivesPredicatesHolder / PropertyFetcher, so the order/limit/has()/property strategies fold into it exactly as they do for UniGraphVertexStep. Handles both the JOIN result and a deferred fallback (inner edge schemas, where searchJoin declines) that hydrates+filters neighbours and preserves duplicate ids (both()).

Two latent bugs surfaced + fixed

Both were only reachable once an adjacency query can be the first query in a traversal — which this start step makes possible:

1. RowController.search(SearchVertexQuery) never flushed the write buffer before searchJoin's direct fetch (only the generic search() flushes). A first-query join read stale data — this caused 56 feature-suite failures until fixed. Now flushes up front.
2. UniGraphPropertiesStrategy GroupCountStep branch dereferences propertyFetchers without the null-guard its sibling branches have. Worked around by making the step a PropertyFetcher (so the lookup is non-null); the unguarded line is noted for a follow-up hardening.

Testing

• JdbcUnboundedAdjacencyTest (11): single-join-not-two-queries, multi-hop chaining from the join step, out/in/both, has() push-down, guards, flag gate.
• JdbcJoinPushdownTest (11), JdbcAdjacencyFilterTest (9): unchanged.
• Full feature + structure suites run with the flag ON produce a failing set byte-identical to flag-off and to master (65 failures, same names) — verified by diffing failure names. Zero regressions, zero new errors.

Follow-ups

• Null-guard UniGraphPropertiesStrategy GroupCountStep branch (independent latent bug).
• outE()/inE()/bothE() and an ES allSources path.

🤖 Generated with Claude Code