fix(core): prevent unbindable preaggregation queries, harden preagg tests

[domoritz]

Fixes #1189, #1187, #1188, #1190. Part of the audit tracked in #1170.

Note: This fix was created by Claude (an agent team ran a binder-error audit of the SQL layer; each fix went through automated implementation, simplification, and adversarial review passes, and the combined branch was re-reviewed for assembly fidelity).

The preaggregator emits an unbindable CREATE TABLE in four different shapes. Since #1090 (and now #1158) the coordinator recovers by falling back to the client's query, so every one of these fails silently as a lost optimization — charts are correct, the speedup just never happens. Relatedly, preaggregator.test.ts mostly asserted !!info, which is also true when the create fails and the fallback answers; that is why none of these were caught. One commit per issue, plus two droppable refactors:

1. F-PA-03: Preaggregation mean-centering subquery drops the schema of a qualified base table #1189 fix: Qualify preagg mean-centering subquery table refs — getBase keeps the full serialized table path ("s"."data"), restoring the optimization for schema-qualified tables (and fixing cross-schema base-identity comparison as a side effect).
2. F-PA-01: Preaggregation emits nested aggregates when mean-centering CTE-derived columns #1187 fix: Bail out of preaggregation when mean-centering CTE-derived columns — four explicit guards (divergent bases / aggregate-valued / window-valued / unresolvable-through-subqueries operands) throw into the existing bail channel.
3. F-PA-02: Preaggregation places window functions without aggregate inputs in GROUP BY #1188 fix: Skip preaggregation for window functions without aggregate inputs — a plain window function among selection outputs skips the optimization instead of landing in GROUP BY (conservative interim relative to Window functions do not work in query with aggregate #259's CTE-hoisting idea).
4. F-PA-04: Preaggregation does not push verbatim selection fields into CTE client queries #1190 fix: Skip preaggregation when verbatim active columns meet subqueries — preaggregateInfo returns null when verbatim selection fields can't be pushed into client CTEs.
5. refactor(core): Share one node-presence test across preagg bail-outs — dedupes the walk helpers into containsNode (new preagg/contains-node.ts).
6. refactor(core): Use a symbol for getBase's divergent-base result — replaces the NaN sentinel; callers read expr === DIVERGENT instead of typeof expr === 'number'.

Commits 5–6 are separately droppable (reviewer-verified they revert cleanly in order) if you'd rather keep this PR fixes-only.

Relationship to #1158

Complementary, verified empirically per fix: #1158 recovers after a failed view create; these guards prevent the invalid CREATE TABLE from being issued at all (cases 2–4), removing the error/log noise the recovery path would absorb. Case 1 is the only one that makes preaggregation work rather than bail.

Test hardening

The schema-qualified test probes the materialized view directly; the bail-out tests assert the non-optimized route plus correct values. Six new tests total (core 72/72); revert checks reproduce exactly the expected per-fix failures.

🤖 Generated with Claude Code