[lmeyerov]

Summary

• Preserve decoded quoted aliases when lowering ORDER BY keys, including aggregate outputs.
• Preserve binding-kind and quotedness recursively in the shared output-name AST rewriter.
• Resolve a projected Identifier before aggregate-expression text mapping at both whole-key and nested boundaries.
• Prepare ORDER BY expressions through the established row-expression boundary so parameters and structured validation remain intact.
• Keep prior-stage outputs needed only for ORDER BY alive through final RETURN projection.

Boundary Matrix

• Positive: quoted aliases with spaces, escaped backticks, aggregate groups, expression-like aliases versus real arithmetic, hidden plain/computed WITH keys, null/tie/LIMIT, and empty results.
• Repaired semantic siblings: literal (a.num + 1) output alias preserves arithmetic semantics for direct RETURN and explicit WITH re-aliasing; compound ORDER BY forms (+1, unary minus, abs, and genuine count(*) + 1) preserve count/sum output bindings.
• R11 interaction: whole and compound v / `v` resolve to the projected v output, not unrelated aggregate-expression text; genuine sum(w) and explicit group_value controls retain independent ordering.
• R12 interaction: ORDER BY \alias` * $factoraccepts supplied parameters and keeps opposite-ranking controls correct; missing parameters retain structuredmissing-required-field context (field=order_by, value=factor`).
• Negative: unresolved quoted aliases report structured invalid-name-type context; genuine unprojected aggregates remain rejected across the row-column and MATCH+UNWIND paths.
• P0051 caller generalization: four swapped count/sum alias-oracle positives per engine for each UNWIND caller, plus two structured negatives per engine, with opposite independent order controls.
• Engines: pandas/Polars focused matrix; exact-source cuDF compound, direct, hidden-key, mixed-expression, R11 shadowed-aggregate, and R12 parameter siblings.

Evidence

• Rebased onto merged fix: preserve igraph collision attributes across round trips #2207 master bf2edb29822fad6b31f482c7203087325c92c7a1. The squashed rebase keeps the lowering.py, backtick tests, and CI-timeout blobs byte-identical; CHANGELOG.md differs only because master now contains the fix: preserve igraph collision attributes across round trips #2207 entry.
• Exact repaired head: ce0979f48fb66005f1539b3b06878b3543d0ef8a; tree df1286436ebdf8d77df7e430c6c0ea5108e32f6f. It includes accepted R10/R6 from db44b245…, R11 from aa3a49a7…, accepted R12 at 5275db6a…, and the P0051 test-only caller amendment.
• Exact wave-9 parameter gate from the repaired source: 12/12 passed, exit 0.
• Exact wave-8 group/sum alias gate from the repaired source: 12/12 passed, exit 0.
• Exact wave-7 composed swapped-alias regression gate from the repaired source: 24/24 passed, exit 0.
• Host focused backtick suite at the repaired head: 207 passed, 13 deselected (CUDA-only cases run separately); the added caller matrix is 20/20; Ruff clean; git diff --check clean.
• Exact-source DGX cuDF proof (graphistry/test-rapids-official:26.02-gfql, cuDF 26.02.01): both new-caller cuDF nodes pass (2 passed, zero skips), and the full backtick suite is 220 passed. Whole-tree archive SHA-256: 9494c7326c6acc960055abc0b6a597c05ae0f2f4d4dffe198d3baf630ae81415. Receipt: worker/research/round-003/S7-2208-rebase-ce0979f48.md.
• Exact-head CI run 39546451858 is terminal SUCCESS. RTD 35287722 is SUCCESS for this head. It includes TCK, all routes-off lanes, route manifest, and test-polars (3.12) checks. The superseded 86ad494d CI run 39502035223 remains historical green evidence.
• R10/R6 repaired-source evidence at db44b245… remains accepted: wave-7 24/24, host 163 passed/7 deselected, DGX full backtick suite 170 passed, CI run 39454366281, and RTD build 35274857.
• R12 receipt: worker/research/round-002/S7-2205-r12-repair-5275db6a.md; prior R11/R10 receipts remain linked in the plan.

Closes #2205