fix(objectql): HAVING 的 $nin / $notContains 对无值行 NULL-safe (#5905)

.changeset/having-filter-null-safe-negative-operators.md

@@ -0,0 +1,5 @@
+---
+'@objectstack/objectql': patch
+---
+
+HAVING 求值对齐 #5298 的 NULL-safe 裁决:聚合行上没有值的列现在满足 `$nin` 与 `$notContains`,与 driver-sql / formula / service-analytics 一致(此前 HAVING 是唯一仍判否的求值面)。

packages/objectql/src/having-filter.test.ts

@@ -4,10 +4,12 @@
  * HAVING evaluator (#4286 step 3) — semantics over AGGREGATED rows.
  *
  * The namespace is the aggregated row's own columns (aggregation aliases +
- * groupBy projections); operator semantics mirror the Filter Protocol's
- * memory evaluation, EXCEPT that an unknown operator throws — ignoring one
- * would silently return unfiltered aggregates, the exact silently-inert
- * failure (#4286, ADR-0078) enforcement exists to end.
+ * groupBy projections); operator semantics follow the Filter Protocol, with two
+ * deliberate divergences from driver-memory's matcher: an unknown operator
+ * throws — ignoring one would silently return unfiltered aggregates, the exact
+ * silently-inert failure (#4286, ADR-0078) enforcement exists to end — and the
+ * negation-carrying operators are NULL-safe per #5298 (see the grid at the
+ * bottom of this file, #5905).
  */
 
 import { describe, it, expect } from 'vitest';
@@ -78,3 +80,113 @@ describe('matchesHaving — the unknown-operator refusal', () => {
     expect(matchesHaving({ k: 'Alpha' }, { k: { $regex: '^alp', $options: 'i' } })).toBe(true);
   });
 });
+
+/**
+ * [#5905] The no-value grid for the negation-carrying operators.
+ *
+ * #5298 ruled (option A, 2026-08-06) that "the column has no value" SATISFIES a
+ * test for "not this value", and PR #5962 landed it on driver-sql, formula,
+ * service-analytics and the `FILTER_LOGIC_*` conformance table. HAVING is the
+ * fifth evaluation face of the same vocabulary and was not in that PR's
+ * inventory, so it stayed the lone holdout — and no conformance table would
+ * have caught it, because `FILTER_LOGIC_CASES` does not drive the HAVING path
+ * (verified: `packages/objectql` imports it nowhere). This grid IS that
+ * coverage.
+ *
+ * Two no-value shapes, deliberately separated, because on this face they did
+ * NOT arrive at the old answer by the same route:
+ *
+ * - NULLED — the key is present with `null`. The early-exit guard tests
+ *   `=== undefined`, so it never fired here; `$nin` was already NULL-safe and
+ *   `$notContains` was not (`typeof null !== 'string'` ⇒ judged false).
+ * - MISSING — the key is absent, so the aggregated row reads `undefined`. The
+ *   early-exit guard fired first and answered false for BOTH operators, before
+ *   either arm was reached.
+ *
+ * The positive-operator rows are the control: `$in` / `$contains` must keep
+ * REJECTING both no-value shapes. Widening the exemption list too far would
+ * turn them green, which is the failure this pair is here to catch.
+ */
+describe('no-value rows and the negation-carrying operators (#5905 / #5298 option A)', () => {
+  const NULLED = { customer_id: 'nulled', tag: null, total: 100 };
+  const MISSING = { customer_id: 'missing', total: 100 };
+  const VALUED_OUT = { customer_id: 'valued_out', tag: 'gamma', total: 100 };
+  const VALUED_IN = { customer_id: 'valued_in', tag: 'alpha', total: 100 };
+  const GRID = [NULLED, MISSING, VALUED_OUT, VALUED_IN];
+
+  describe('$nin', () => {
+    it('a NULLED column satisfies $nin', () => {
+      expect(matchesHaving(NULLED, { tag: { $nin: ['alpha', 'beta'] } })).toBe(true);
+    });
+
+    it('a MISSING column satisfies $nin', () => {
+      expect(matchesHaving(MISSING, { tag: { $nin: ['alpha', 'beta'] } })).toBe(true);
+    });
+
+    it('a present value OUTSIDE the list still satisfies $nin (unchanged)', () => {
+      expect(matchesHaving(VALUED_OUT, { tag: { $nin: ['alpha', 'beta'] } })).toBe(true);
+    });
+
+    it('a present value INSIDE the list still fails $nin (unchanged)', () => {
+      expect(matchesHaving(VALUED_IN, { tag: { $nin: ['alpha', 'beta'] } })).toBe(false);
+    });
+
+    it('applyHaving keeps both no-value rows and drops only the listed value', () => {
+      expect(applyHaving(GRID, { tag: { $nin: ['alpha', 'beta'] } }).map((r) => r.customer_id))
+        .toEqual(['nulled', 'missing', 'valued_out']);
+    });
+  });
+
+  describe('$notContains', () => {
+    it('a NULLED column satisfies $notContains', () => {
+      expect(matchesHaving(NULLED, { tag: { $notContains: 'lph' } })).toBe(true);
+    });
+
+    it('a MISSING column satisfies $notContains', () => {
+      expect(matchesHaving(MISSING, { tag: { $notContains: 'lph' } })).toBe(true);
+    });
+
+    it('a present value WITHOUT the substring still satisfies $notContains (unchanged)', () => {
+      expect(matchesHaving(VALUED_OUT, { tag: { $notContains: 'lph' } })).toBe(true);
+    });
+
+    it('a present value WITH the substring still fails $notContains (unchanged)', () => {
+      expect(matchesHaving(VALUED_IN, { tag: { $notContains: 'lph' } })).toBe(false);
+    });
+
+    it('applyHaving keeps both no-value rows and drops only the containing value', () => {
+      expect(applyHaving(GRID, { tag: { $notContains: 'lph' } }).map((r) => r.customer_id))
+        .toEqual(['nulled', 'missing', 'valued_out']);
+    });
+  });
+
+  describe('the control: positive operators still reject a no-value column', () => {
+    it('$in rejects NULLED and MISSING', () => {
+      expect(matchesHaving(NULLED, { tag: { $in: ['alpha', 'beta'] } })).toBe(false);
+      expect(matchesHaving(MISSING, { tag: { $in: ['alpha', 'beta'] } })).toBe(false);
+    });
+
+    it('$contains rejects NULLED and MISSING', () => {
+      expect(matchesHaving(NULLED, { tag: { $contains: 'lph' } })).toBe(false);
+      expect(matchesHaving(MISSING, { tag: { $contains: 'lph' } })).toBe(false);
+    });
+
+    it('$ne — already exempt before #5905 — is unchanged for both shapes', () => {
+      expect(matchesHaving(NULLED, { tag: { $ne: 'alpha' } })).toBe(true);
+      expect(matchesHaving(MISSING, { tag: { $ne: 'alpha' } })).toBe(true);
+      expect(matchesHaving(VALUED_IN, { tag: { $ne: 'alpha' } })).toBe(false);
+    });
+  });
+
+  /**
+   * The NULL-safety lives at the LEAF, so `$not` inverts it rather than
+   * inheriting it — the same design driver-sql writes down for its own
+   * `$not` rewrite (a nested negation totalises its own operand). A no-value
+   * row satisfies `$nin`, therefore it does NOT satisfy `$not: { $nin }`.
+   */
+  it('$not inverts the leaf answer instead of re-applying the guard', () => {
+    expect(matchesHaving(MISSING, { $not: { tag: { $nin: ['alpha'] } } })).toBe(false);
+    expect(matchesHaving(NULLED, { $not: { tag: { $notContains: 'lph' } } })).toBe(false);
+    expect(matchesHaving(VALUED_IN, { $not: { tag: { $nin: ['alpha'] } } })).toBe(true);
+  });
+});

packages/objectql/src/having-filter.ts

@@ -19,13 +19,27 @@
 // item's `alias` for structured entries — after date bucketing). It is an
 // ordinary FilterCondition over those columns: implicit equality, the
 // comparison / set / null / existence / string operators, and `$and` / `$or` /
-// `$not` composition. Operator semantics mirror the Filter Protocol as
-// driver-memory's matcher implements it, with ONE deliberate divergence:
+// `$not` composition. Operator semantics follow the Filter Protocol, with TWO
+// deliberate divergences from driver-memory's matcher — the face this module
+// was originally written against:
 //
-// AN UNKNOWN OPERATOR THROWS. The memory matcher ignores operators it does not
-// know; here an ignored operator would silently return UNFILTERED aggregates —
-// the precise failure mode (#4286, ADR-0078) this module exists to end. The
-// rejection names the operator and the supported set.
+// 1. AN UNKNOWN OPERATOR THROWS. The memory matcher ignores operators it does
+//    not know; here an ignored operator would silently return UNFILTERED
+//    aggregates — the precise failure mode (#4286, ADR-0078) this module exists
+//    to end. The rejection names the operator and the supported set.
+//
+// 2. [#5905] THE NEGATION-CARRYING OPERATORS ARE NULL-SAFE. `$ne` / `$nin` /
+//    `$notContains` are satisfied by a row whose column HAS NO VALUE — "the
+//    column has no value" satisfies a test for "not this value". That is the
+//    ruling the maintainer took on #5298 (option A, 2026-08-06), landed by
+//    PR #5962 across driver-sql, formula, service-analytics and the
+//    `FILTER_LOGIC_*` conformance table. HAVING is the FIFTH evaluation face of
+//    the same vocabulary and was not in that PR's inventory, which left this
+//    file as the lone holdout (#5905) — and the only face no conformance table
+//    covers, since `FILTER_LOGIC_CASES` does not drive the HAVING path.
+//    driver-memory / driver-mongodb still answer the old way only because
+//    #5499 freezes them; the divergence is against a frozen face, not against
+//    the ruling.
 
 import type { FilterCondition } from '@objectstack/spec/data';
 
@@ -49,6 +63,27 @@ function unknownOperator(op: string, where: 'logical' | 'condition'): Error {
   );
 }
 
+/**
+ * [#5905] Operators whose answer for a column with NO VALUE is decided by the
+ * operator's own arm below, not by the early exit in {@link checkCondition}.
+ *
+ * That exit exists so a POSITIVE test (`$gt`, `$in`, `$contains`, …) can never
+ * be accidentally satisfied by a column the aggregated row does not carry. The
+ * operators listed here are the ones for which "no value" is a real answer
+ * rather than an accident:
+ *
+ * - `$exists` / `$null` — answering about absence IS their whole job;
+ * - `$ne` / `$nin` / `$notContains` — they carry their own negation, and #5298
+ *   ruled (option A) that a value-less column satisfies them, on every backend.
+ *
+ * `$nin` and `$notContains` were missing from this list, which is the defect
+ * #5905 records: the exit fired first and answered FALSE for them, so the arms
+ * below — which would have answered TRUE — were never reached.
+ */
+const NO_VALUE_ANSWERED_BY_OPERATOR: ReadonlySet<string> = new Set([
+  '$exists', '$ne', '$null', '$nin', '$notContains',
+]);
+
 /**
  * Filter aggregated rows by the query's `having` condition. An absent or empty
  * condition returns the rows unchanged (same vacuous-filter convention as
@@ -109,7 +144,7 @@ function checkCondition(value: any, condition: any): boolean {
   for (const op of keys) {
     if (op === '$options') continue; // consumed by $regex below
     const target = (condition as Record<string, any>)[op];
-    if (value === undefined && op !== '$exists' && op !== '$ne' && op !== '$null') return false;
+    if (value === undefined && !NO_VALUE_ANSWERED_BY_OPERATOR.has(op)) return false;
     switch (op) {
       // eslint-disable-next-line eqeqeq
       case '$eq': if (value != target) return false; break;
@@ -134,7 +169,16 @@ function checkCondition(value: any, condition: any): boolean {
         if (target === false && value == null) return false;
         break;
       case '$contains': if (typeof value !== 'string' || !value.includes(target)) return false; break;
-      case '$notContains': if (typeof value !== 'string' || value.includes(target)) return false; break;
+      // [#5905] The mirror of `$contains`, NOT its copy-with-a-negated-test.
+      // `$contains` fails a non-string value because "contains" cannot hold for
+      // something that is not text; `$notContains` SUCCEEDS for the same value
+      // for the same reason — it cannot contain the substring. Reusing the
+      // `typeof value !== 'string' ⇒ false` guard here (what this line used to
+      // do) made a value-less column fail BOTH an operator and its negation,
+      // the two-valued reading #5298 ruled out. This is `formula`'s shape
+      // (`matches-filter.ts`: `!(typeof actual === 'string' && …)`), which
+      // driver-sql's polarity table already follows for the same operator.
+      case '$notContains': if (typeof value === 'string' && value.includes(target)) return false; break;
       case '$startsWith': if (typeof value !== 'string' || !value.startsWith(target)) return false; break;
       case '$endsWith': if (typeof value !== 'string' || !value.endsWith(target)) return false; break;
       case '$regex': {