fix(rule): fire HNSW rewrite through SubqueryAlias

[rohan-hotdata]

Summary

• Table-aliasing SQL clients (ibis/SQLGlot, ORMs, BI tools) alias every table and wrap projections in subqueries. DataFusion keeps these SubqueryAlias nodes when the alias is referenced downstream, and USearchRule only descended through Projection/Filter — so it silently declined and fell back to brute-force search for any aliased query, including ibis's actual generated SQL.
• peel_alias/peel_alias_capturing see through SubqueryAlias at each matcher entry point. The rewrite still resolves internally against the real table reference (so it type-checks against USearchNode's own schema), then re-wraps the result in the same outer alias so the exposed schema matches the original qualifiers exactly. A trailing schema-equality check declines cleanly (no rewrite, exact fallback) rather than emit a plan DataFusion's post-optimizer invariant check would reject.
• One shape is a known, safe limitation: a single alias directly on the scan with no outer wrap over the Sort leaves the original schema with mixed qualifiers (one column tied to the alias, another unqualified) that can't be reproduced by a single Projection/SubqueryAlias. The rule declines safely there — identical to pre-fix behavior for that shape, just reached via a different path. Documented in tests/subquery_alias.rs.

Test plan

• cargo fmt --check, cargo clippy --all-targets -- -D warnings, cargo test --features sqlite-provider — all clean (96 tests).
• New tests/subquery_alias.rs — 7 tests covering the double-alias (ibis) shape, single-alias + WHERE, vector-in-output decline, and metric-mismatch decline.
• New tests in tests/execution.rs (exec_ibis_shape_double_alias_*) run the actual ibis-compiled SQL (verified byte-for-byte against a live ibis 12.0.0 instance compiling the semantic_search helper sketch) against real vector data, asserting correct nearest-neighbor results — not just plan shape.
• Verified against a real RuntimeDB build: patched runtimedb's Cargo.toml to this branch locally, built the Linux release binary, built/loaded the runtimedb:local image into the local kind cluster, rolled out the deployment, and ran the exact ibis-shaped double-SubqueryAlias query through the live HTTP API against a real uploaded table + cosine/L2 index. EXPLAIN showed USearchExec in the physical plan and SubqueryAlias: t1 correctly wrapping the USearch logical node; the real query returned correct results (closest row at distance 0.0).

🤖 Generated with Claude Code

[claude]

Reviewed the SubqueryAlias-peeling logic in src/rule.rs. The design is sound: peeling sees through alias chains at each matcher entry point, the rewrite resolves against the real table reference, re-wraps in the captured outer alias, and the trailing schema-equality check declines cleanly for any shape it can't reproduce exactly (no invalid plan reaches DataFusion's invariant check). Verified the filter-predicate qualifier is a non-issue since the planner strips qualifiers before compiling physical filters, and that the rewritten plan can't re-fire. Tests cover both the fire and decline paths, including real execution against ibis-compiled SQL.