[TimelordUK]

Why

Examples smoke failures were being cleaned up by deleting or editing the offending SQL. That erased the evidence: three statements removed in 2240b73 / 676309f were each demonstrating a real, open engine gap. This restores them behind the statement-level -- [SKIP] directive so the queries stay in the tree as end-to-end acceptance tests without reddening the run.

No Rust changed.

Restored and marked

Example	Gap	Tracked as
chemistry.sql	OR in JOIN ... ON	P27 (open)
jsonl_logs.sql	ORDER BY ... NULLS LAST	P13 stage 2 (open)
statistical_analysis.sql	scalar subquery in a CTE	P38 (new)

-- [SKIP] is checked in non_interactive.rs before parsing, so unparseable SQL behind it costs nothing. Each P* entry gained an Acceptance test line pointing at the example, so whoever picks the fix up knows where the end-to-end check lives.

For jsonl_logs.sql the NULLS-free statement still runs and the original sits beside it skipped — the demo and the tracked gap, both. Also reverts an unintended whitespace change from 2240b73.

Two new findings

Both have corpus cases; both reference behaviours were confirmed against DuckDB rather than assumed.

P37 — a window function inline in WHERE silently returns zero rows. Success exit code, header printed, no error, on a 20-row table. DuckDB rejects the query outright (Binder Error: WHERE clause cannot contain window functions). Bucketed OURS_ONLY because we "succeed" where the reference errors — the bucket is right, the behaviour is not.

It is a sibling of P15 with one fix site: ExpressionLifter only walks the SELECT list, so an inline window is never hoisted and reaches recursive_where_evaluator.rs as a raw WindowFunction. From QUALIFY that path errors (P15); from WHERE it filters every row out instead. Silent, so it leads the open fix order. Found in expander_rewriters.sql, which advertises this as a working transformation.

P38 — a scalar subquery in a CTE body's SELECT list is never evaluated. Narrower than it first appeared; two controls locate it:

Shape	Result
Top-level amount - (SELECT AVG(…))	✅ AGREE
Same expression inside a CTE body	🚫 GAP
Bare (SELECT AVG(…)) in a CTE, no arithmetic	🚫 error
amount * 2 inside a CTE	✅ AGREE

Neither scalar subqueries nor CTE arithmetic is broken in general — two SELECT-list evaluation paths disagree on which expression types they support. The entry flags checking ENGINE_REFACTORING.md before adding the arm twice.

Smoke suite triage: 13 failures → 1

Twelve examples can never run in CI — they need a live server, post to placeholder hosts (api.example.com, tradesapi, trade-server), or are nvim-plugin macro templates (one contains no runnable statement at all). All now carry file-level -- [TEST:SKIP] <reason>, which test_examples.py::should_skip_file reads and reports.

expander_rewriters is deliberately left failing — it is P26, a real gap, and skipping it would bury it. Smoke failures are already non-blocking (the suite exits 0), so CI stays green while the bug stays visible.

Verification

• Parity contract holds at 172 cases, 141 AGREE (up 2 — the new controls).
• Examples: 152/153 pass, all 33 formal tests pass, suite exits 0.
• No Rust changed, so no cargo fmt / clippy delta.

🤖 Generated with Claude Code

https://claude.ai/code/session_01PpidyXocQt835yaJpKBSHk