fix(parity): close P10/P6/P12 + make history test reliable cross-platform

[TimelordUK]

Summary

Clears three DuckDB-parity gaps and fixes a long-standing flaky test. Parity board moves 72 → 76 AGREE, 10 → 6 GAP; each fixed case has its corpus expect dropped so the --check gate now locks the gain.

Two logically separate commits:

1. fix(parity) — P10, P6, P12

P	Gap	Fix
P10	HAVING NOT (…) errored (Unsupported expression type … Not)	Added a Not arm to ArithmeticEvaluator::evaluate, negating via the existing to_bool helper with SQL three-valued logic on the NULL edge (NOT NULL → NULL).
P6	INTERSECT / EXCEPT unimplemented (return Err stubs)	Implemented both in the set-op loop. DISTINCT-by-default, filtering the materialized left table against the right using the same Debug row-key apply_distinct uses for UNION, deduping inline.
P12	WITH rejected in expression position	parse_subquery already dispatched a leading WITH to the CTE parser; widened the subquery-detection guards from Token::Select to Token::Select | Token::With in primary.rs (scalar + tuple IN) and comparison.rs (IN / NOT IN). Retires previously-unreachable CTEHoister arms.

Corpus note: the except case threshold moved 2000 → 3000 — every region has a sale > 2000, so the original form was trivially empty and would have AGREEd for the wrong reason; > 3000 leaves {Oceania}, a genuine left-minus-right difference.

docs/SQL_PARITY.md entries flipped to 🟢 FIXED with rationale.

2. test(history) — reliable cross-platform

history_protection_integration sandboxed the history dir by redirecting HOME/APPDATA, but dirs::data_dir() on Windows resolves via the Win32 known-folder API and ignores those env vars, so the test silently read the developer's real history and failed. Env-var redirection is also process-global, so #[serial] (which only serializes serial-tagged tests) couldn't stop a concurrent test from writing into the sandbox and tripping the history-protection guard — hence pass-alone / fail-in-suite flakiness.

Fix: added CommandHistory::with_history_file(path) and had new() delegate to it. The test now builds against an explicit tempdir file — no env var, no #[serial], no cross-test races. Also added a SQL_CLI_DATA_DIR override to AppPaths (honored directly, so it works on every platform) as a legitimate way to relocate/sandbox app data.

Testing

• Parity harness: 76 AGREE / 1 DIFFER / 6 GAP, contract holds (83 cases).
• cargo test --lib: 687 passed, 0 failed.
• cargo test --test integration: 407 passed, 0 failed — verified green across three repeated full runs (previously flaky).
• cargo fmt clean.

🤖 Generated with Claude Code