fix(sql): add missing DuckDB aggregate names, guard against future drift

[domoritz]

Fixes #1142.

Note: This fix was created by Claude (an agent team ran a binder-error audit of the SQL layer; the fix went through automated implementation, simplification, and adversarial review passes). It is a draft until @domoritz has reviewed it and marks it ready.

What changed

Commit 1 — the fix. Adds the 8 missing names to aggregateNames (alphabetical placement): arg_max_nulls_last, arg_min_nulls_last, argmax, argmin, count_if, group_concat, listagg, sem. Regression tests in test/visitors.test.ts ("Verbatim aggregate detection") build a markQuery-shaped query per name and assert via toBeValidQuery that no GROUP BY is emitted over the aggregate alias and the result binds in DuckDB.

Commit 2 — the guard. New test/aggregate-names.test.ts diffs aggregateNames in both directions against duckdb_functions() WHERE function_type = 'aggregate' on the live test-fixture DuckDB, so this drift class (#1051, and now #1142) is caught mechanically whenever the DuckDB dependency is bumped instead of by user bug reports. Two commented exclusion sets carry the intent: window-only functions that DuckDB reports as aggregates (usable only with OVER, which isAggregateExpression handles separately) plus internal sum_no_overflow, and macro-backed names (geomean, geometric_mean, histogram_values, weighted_avg) that still expand to aggregates. Failure messages print the exact ready-to-paste arrays. A minimal queryFixture helper was added to test/util/validate.ts.

Deliberately not added: the window-only pseudo-aggregates and sum_no_overflow (see exclusion rationale in the test).

Verification

• pnpm --filter @uwdata/mosaic-sql test: 274/274 pass
• pnpm --filter @uwdata/mosaic-core test: 63/63 pass (aggregate detection feeds core preaggregation)
• typecheck + lint clean
• Revert checks: with src/ast/aggregate.ts reverted, all 8 new detection tests fail with the binder error above, and the drift test fails printing exactly the 8 missing names; with a bogus name injected, the drift test fails in the remove direction.

(“F-FN-01” in the commit messages is the internal finding ID from the audit.)

🤖 Generated with Claude Code

[derekperkins]

When I was just working on the function allowlist, duckdb_functions() wasn't quite enough for a full inventory, as several functions are classified as macros. I ended up creating a dedicated AGENTS.md to assist with future updates.

That being said, this is probably sufficient here. I mentioned in my original allowlist PR, it might be useful to put function information in json format elsewhere so any package can look at the function introspection. Right now every core and core extension function is enumerated in the go server, split into compute-only and elevated permissions groups.

[domoritz]

@derekperkins good points. Let's merge this for now and when we have a good inventory of functions, we can refer to that from a test similar to how we use duckdb_functions().