Keep GFQL AUTO index construction source-native

[lmeyerov]

What changed

GFQL resident-index construction now keeps source-native frames under engine='auto':

• eager Polars nodes and edges select the Polars index engine;
• pandas remains pandas;
• cuDF remains cuDF;
• mixed, missing-frame, and Polars LazyFrame inputs retain the legacy pandas fallback;
• explicit engines and global AUTO query routing are unchanged.

This avoids the legacy AUTO path converting resident eager-Polars frames to pandas during index construction, which prevented the later explicit Polars hop from serving that resident index and reintroduced a full-frame conversion tax.

This is a clean current-master delivery of the reviewed change from #1767. It supersedes #1767 only as the delivery vehicle; #1767 is intentionally left open.

Validation

• Exact reviewed three-file patch: 168 additions, 2 deletions; source/test blobs and patch hash match the converged fix(gfql): index AUTO preserves resident polars frames (kills the polars-hop O(E) tax) #1767 review input.
• Review converged with 0 blockers, 0 important findings, and 1 deferred suggestion.
• Guarded RAPIDS 25.02 and 26.02 targeted source-native class: 10 passed each.
• Guarded RAPIDS 26.02 full resident-index test file: 155 passed.
• Same-build parity-asserted A/B: seeded Polars hop 2.934 ms → 0.139 ms median (21.17×); index build 21.967 ms → 21.367 ms; scan-oracle and base/head value parity held.

Fresh GitHub CI for this clean PR is required before manual merge.

[lmeyerov]

Closing as a duplicate of #1767: the two diffs are byte-identical (same 3 files, +168 lines; gh pr diff output compares equal ignoring blob hashes). #1767 is the earlier of the two (2026-07-22 vs 2026-07-23) and has the fuller description of the polars-hop O(E) tax it kills, so keeping that one.

Both are still ~200 commits behind master and neither has been re-measured since the #1758-#1763 abstraction-tax stack and #1776/#1777 landed, so #1767 needs a restack and a fresh A/B before it can be judged.