fix(gfql): reject malformed aggregate sources during validation

[lmeyerov]

Aggregation validation treated malformed sources as dependency-free expressions, allowing invalid queries to pass validation. This change requires a visible literal column or a parsed expression, preserves structured diagnostics and operation indices, and rejects statically non-numeric constants for numeric aggregates.

Input boundary	Validation result
Exact visible column, including dotted/operator-like/malformed-expression names	Accepted as a literal column before parsing
Valid expression or numeric/Boolean/null constant	Accepted when dependencies are visible
Missing expression dependency	E301 with the missing active-row column
Malformed non-column expression	E107
Parser backend unavailable for a non-column expression	E201; visible literal columns remain usable
String/list/map literal passed to sum/avg/mean	E302
Empty grouping-key list	Accepted with an aggregate; global execution support is provided by #2077
No aggregates and no effective grouping keys	E201; explicit keys or matching key prefixes preserve key-only grouping

Static validation tracks visible column names, not inferred types of every transformed column. Column-dependent operand type errors remain governed by the shared runtime aggregate contract; validation acceptance does not promise that every backend lowers a query. Unknown-schema helper calls retain their deferred contract.

Executable positive/negative examples live in test_group_by_agg_source_columns.py: numeric constants versus string/list/map constants; malformed expressions versus identical visible column names; missing dependencies; missing parser; and global-shape acceptance. Changed syntax/type/parser diagnostics retain operation indices, including collect_all. The existing direct missing-column E301 path still omits that index in first-error mode; collect_all supplies it. The safelist tests retain non-string-key rejection.

Current local validation at c5dd1f1db: 437 passed, 141 skipped, Ruff and type-hygiene passed, and scoped mypy reports only the five documented baseline dependency-profile diagnostics. Earlier cuDF25/26 and polars-gpu evidence covers the original syntax-validation change; the new static constant/global expansion still needs final combined-device/CI and review gates.

Merge #2077 first for the expanded global runtime contract. Prior 07b8dfb71 source review had two consecutive clean waves, 66 combined-runtime and 13 additional adversarial checks, and CI with 80 successful checks plus five skips. Review then found the newly admitted zero-output global shape differed across engines. The current fix uses a shared structured shape contract, preserves key-only grouping, and reopens review/CI gates. Keep this PR draft until final review, expanded device/performance and exact-head CI gates complete.

Fixes #2076. Tracking: #2074. Follow-up PRs remain unmerged.

Source review now has two consecutive clean post-fix waves at the published head. Final device/performance and exact-head CI gates remain open; source convergence does not establish those outcomes.

Catch-up after merged #2072

Current master baseline is 2ca98922c. Updated product heads: #2077 f18884334, #2078 4c8125a62, #2079 10e3b4484. Catch-up preserves the new Polars source-node-ID projection argument together with entity provenance. Combined runtime 3ef2a9117 passes295focused tests and3840TCK tests (5skipped,689xfail;327explicit Polars declines). Individual aggregate215, validation162, and entity/projection/reentry422 tests pass. Owning benchmark correctness passes32tests; benchmark261 685b1cd CI2success verifies productf188.

First clean catch-up source review completed; final product CI, second review pass and final device/performance evidence remain open. Earlier source pins and readiness statements above are historical. Entity2079 is draft again pending refreshed gates; follow-up PRs remain unmerged.

Final DGX source/test/benchmark payload was refreshed for2072 (~16.3MB, exact hashes). Automatic approval review again rejected transfer, treating the objective attachment’s authorization as untrusted; no package transfer or final measurement occurred. Direct confirmation has been requested while independent review/CI continues.