@object-ui/core 17.6.0 → 18.0.0

package/dist/utils/dataset-format.js

@@ -6,30 +6,44 @@
  * LICENSE file in the root directory of this source tree.
  */
 import { formatDisplayNumber } from './number-display.js';
+import { formatDate, formatDateTime, formatRelativeDate } from './date-display.js';
+import { resolveMeasureLabel } from './chart-series.js';
+import { composeDrillFilter } from './drill-down.js';
 /**
- * Scale a stored `percent`-field value to its DISPLAY magnitude.
- *
- * Percent fields store a FRACTION (0–1) by convention — a stored `0.75` means
- * 75% (see the percent edit widget `PercentField`, which divides input by 100).
- * A value already in whole-percent form (magnitude ≥ 1, e.g. a `progress` /
- * `completion` field storing `57`) is passed through unchanged. This is the
- * SINGLE source of truth for percent display scaling, shared by the list-view
- * percent cell renderer (`formatPercent` in `@object-ui/fields`) and the dataset
- * measure formatter ({@link formatMeasure}) so a percent renders identically as
- * a row value and as an aggregated metric — the two surfaces can never drift.
- *
- * ⚠️ That last sentence was briefly FALSE, and objectui#4576 is what made it
- * true again. Sharing the SCALING was never enough on its own: between #4553
- * and #4576 the cell renderer got the locale's percent CONVENTION from `Intl`
- * while {@link formatMeasure} appended a literal '%', so a German session read
- * `1.234,5 %` in a list cell and `1.234,5%` in a dashboard measure — the same
- * number, scaled identically, rendered under two conventions. Both ends now go
- * through the locale's own percent affix. If a third surface ever needs percent
- * display, it takes BOTH halves from here — the scaling AND the convention —
- * or this promise breaks again in the same place.
+ * Scale a stored percentage to its DISPLAY magnitude (percentage points), at
+ * the storage the caller STATES.
+ *
+ * `percentScale` is the spec's `PercentScale`: `fraction` is a 0–1 ratio
+ * (`0.75` ⇒ 75, `1` ⇒ 100), `whole` is already in percentage points (`75` ⇒
+ * 75, `1` ⇒ 1). A caller holding a field definition reads it with the spec's
+ * `percentScaleOf` (`@objectstack/spec/data`): a `percent` field stores a
+ * fraction unless it declares a `max` above 1. A caller with no field (a
+ * computed ratio, a server measure column) states the storage it knows: a
+ * `derived: { op: 'ratio' }` measure is a fraction by definition, and the
+ * server annotates a measure column with its `percentScale`.
+ *
+ * ⛔ No storage is inferred from the VALUE any more (objectui#11475). This
+ * function used to be `value > -1 && value < 1 ? value * 100 : value`, a guess
+ * from the magnitude that read neither the field nor its `max`. A fraction-
+ * stored `1` (100%) therefore read `1%` in the list cell while the read-only
+ * form, which reads the declaration, read `100%`; a whole-stored `0.5` read
+ * `50%`. objectui#3136 was the same guess on a dashboard measure. The
+ * parameter is REQUIRED, so every caller says which storage it holds, and a
+ * value outside the union throws rather than falling to a branch silently.
+ *
+ * ⭐ The contract a third percent surface still owes is BOTH halves: this
+ * scaling, at the stored scale, AND the locale's percent convention
+ * (`style: 'percentPoints'`, see {@link formatMeasure} and `formatPercent` in
+ * `@object-ui/fields`). Between objectui#4553 and objectui#4576 the two ends
+ * shared the scaling and not the convention, so a German session read
+ * `1.234,5 %` in a list cell and `1.234,5%` in a dashboard measure.
  */
-export function percentDisplayValue(value) {
-    return value > -1 && value < 1 ? value * 100 : value;
+export function percentDisplayValue(value, percentScale) {
+    if (percentScale === 'fraction')
+        return value * 100;
+    if (percentScale === 'whole')
+        return value;
+    throw new TypeError(`percentDisplayValue: percentScale must be 'fraction' or 'whole' (the spec's PercentScale), got ${String(percentScale)}`);
 }
 /**
  * Format one number in the DISPLAY locale, surviving a malformed locale tag.
@@ -65,6 +79,137 @@ export function percentDisplayValue(value) {
 function formatNumberInLocale(value, locale, options) {
     return formatDisplayNumber(value, { ...options, locale });
 }
+/**
+ * ISO calendar date with NO time part — `2026-07-04`.
+ *
+ * Deliberately anchored at both ends and deliberately narrower than
+ * `Date.parse`, which also accepts locale prose (`March 5, 2026`), bare years
+ * (`2026`) and — the one that matters here — plain NUMERALS. A measure value
+ * is untyped by the time it reaches {@link formatMeasure}, so the shape test
+ * IS the type test, and a loose one would capture the numeric strings and
+ * counts this formatter must leave exactly as they are.
+ */
+const ISO_DATE_ONLY_RE = /^\d{4}-\d{2}-\d{2}$/;
+/**
+ * ISO date carrying a time part — `2026-07-04T07:00:00.000Z`, or the same with
+ * a space separator, with or without seconds/offset. Matched only as far as
+ * `HH:mm`; `Date.parse` decides the rest. What it rejects is an out-of-range
+ * MONTH or CLOCK reading (`2026-13-01T09:30`, `2024-07-04T99:99`), which this
+ * pattern admits — such a value falls through untouched.
+ *
+ * ⛔ A DAY that overflows its month does NOT fall through. `2026-02-30T09:30`
+ * parses and reaches the shared date path, which refuses it with the dash
+ * `formatDateTime` renders for any unparsable value (objectui#10301, the
+ * date-time half of objectui#10026). The refusal is that path's, not this
+ * pattern's. See the note inside {@link formatMeasureDate}.
+ */
+const ISO_DATETIME_RE = /^\d{4}-\d{2}-\d{2}[T ]\d{2}:\d{2}/;
+/**
+ * Render a date-shaped measure value through the display path a list cell
+ * already uses for that field, or return `undefined` to say "not a date" and
+ * leave the caller's own fallthrough in charge.
+ *
+ * ── Why this is a ROUTE and not a formatter (objectui#7178) ──
+ * Every line of date rendering below happens in `./date-display.ts`, which is
+ * where `@object-ui/fields`' `formatDate` / `formatDateTime` now live and what
+ * `DateCellRenderer`, `ObjectGrid`'s date cells and `ObjectGantt`'s tooltips
+ * call. Writing the date convention HERE instead is the mistake this file's
+ * own header is about: objectui#4576, where a percent convention had one copy
+ * in a list cell and another in this function, both correct, and a German
+ * session read `1.234,5 %` in the cell beside `1.234,5%` in the measure. One
+ * date convention, one home, nothing to drift.
+ *
+ * The date/datetime split mirrors `ObjectGantt.tsx`'s sniffed-ISO dispatch
+ * verbatim, because that site is already the repo's answer to this exact
+ * situation: an ISO string arriving with NO field type attached. A measure is
+ * that situation by construction — the analytics result carries a column, not
+ * a field definition — so the absolute locale form is taken rather than
+ * `DateCellRenderer`'s `|| 'relative'` default, which is keyed on knowing the
+ * value is a `date` FIELD in a row.
+ *
+ * ── `format` is honoured on BOTH arms, by different means (objectui#8352) ──
+ * The accepted vocabulary is `'short'` and `'relative'` on EITHER arm, and
+ * every other string — including `'compact'`, including a date PATTERN such as
+ * `'YYYY-MM-DD'` — falls to that arm's default face. The two arms reach that
+ * one vocabulary differently, and the difference is the whole of #8352:
+ *
+ *   - The DATE arm threads `format` onward, into `formatDate`'s positional
+ *     STYLE parameter. That is the same mapping `DateCellRenderer` makes from
+ *     `field.format` and the same one `plugin-dashboard`'s `recordFields`
+ *     makes for a date-shaped format.
+ *   - The DATETIME arm SELECTS a formatter above, because it has nothing to
+ *     thread into: `formatDateTime(value, options?)` has no style parameter,
+ *     so before #8352 this arm could not honour `format` even in principle —
+ *     it dropped it, and `format: 'relative'` on a `Field.datetime` measure
+ *     rendered the absolute face with no diagnostic.
+ *
+ * ⚠️ Threading `format` into `formatDateTime`'s `options.style` is NOT the fix
+ * and was measured before this one was written: that key's vocabulary is
+ * `'compact'` alone, so the pass-through would honour the one word the date
+ * arm does NOT honour while still ignoring both words it does — the defect
+ * inverted, not closed. Widening the signature was refused separately: it is a
+ * published one (see {@link formatDateTime}'s note on why the positional shape
+ * was refused there), and the parity #8352 asks for is reachable from the call
+ * site without touching it.
+ *
+ * Which FACE each honoured word selects is per-arm, because the two arms
+ * render different types and a datetime's own resolution is part of its datum:
+ *
+ *   - `'relative'` → `formatRelativeDate` on both arms, the same function, so
+ *     an in-window value reads `In 2 days` in either column. Its ±7-day
+ *     fallback is inherited rather than re-decided here — an out-of-window
+ *     value renders the absolute DATE face on both arms. Re-deciding that
+ *     window at this call site would put a second copy of the convention in
+ *     this file, which is objectui#4576 exactly.
+ *   - `'short'` → the dense narrow-card face of the arm's own type:
+ *     `formatDate`'s `'short'` (`Sep 11, '26`) for a date, `formatDateTime`'s
+ *     `'compact'` (`9/11/2026 9:30 am`) for a datetime. The datetime face
+ *     keeps the time, and it is byte-identical to what every `datetime` CELL
+ *     paints, so a measure tile and a grid cell showing the same instant
+ *     agree — objectui#4576's rule again, in the other direction.
+ *
+ * See the `format` note on {@link formatMeasure}.
+ */
+function formatMeasureDate(v, format, locale) {
+    if (typeof v !== 'string')
+        return undefined;
+    // `Date.parse` guards both arms, and what it rejects is an out-of-range
+    // MONTH (`2026-13-01`) — which both patterns above admit — ⛔ NOT an
+    // out-of-range DAY. ECMAScript's Date Time String Format accepts `DD` in
+    // `01`-`31` syntactically and `MakeDay` rolls the surplus into the next
+    // month, so a well-shaped impossible date (`2026-02-30`) is PARSEABLE on
+    // both arms: it does not fall through to `String(v)`, it reaches the shared
+    // date path. This paragraph once asserted the opposite — that such a value
+    // keeps falling through — until objectui#8263 measured it.
+    //
+    // ⭐ What that path answers is ITS decision, never this guard's. Since
+    // objectui#10026 it refuses a date-only nonexistent day with the dash it
+    // renders for any unparsable value, so the measure shows that same dash —
+    // agreeing with the list cell ON THE REFUSAL. This guard must not refuse
+    // on its own (by letting the value fall through to `String(v)`, say): the
+    // measure would then disagree with the list cell, the objectui#4576 split.
+    // `agrees with the list cell on an impossible calendar day — on the refusal`
+    // in this file's co-located date suite pins it; read that test for the
+    // full argument.
+    if (ISO_DATE_ONLY_RE.test(v)) {
+        return Number.isNaN(Date.parse(v)) ? undefined : formatDate(v, format, { locale });
+    }
+    if (ISO_DATETIME_RE.test(v)) {
+        if (Number.isNaN(Date.parse(v)))
+            return undefined;
+        // The datetime arm honours the SAME two words the date arm does, by
+        // choosing a formatter here (objectui#8352). It cannot thread `format`
+        // onward the way the date arm does, because `formatDateTime(value,
+        // options?)` takes no style parameter — see this function's docblock for
+        // why that is answered HERE and not by widening a published signature.
+        if (format === 'relative')
+            return formatRelativeDate(v, { locale });
+        if (format === 'short')
+            return formatDateTime(v, { locale, style: 'compact' });
+        return formatDateTime(v, { locale });
+    }
+    return undefined;
+}
 /**
  * Format a MEASURE value. Currency comes from the field's declared `currency`
  * (locale-correct symbol via `Intl`), NOT from a "$" baked into the format
@@ -102,14 +247,70 @@ function formatNumberInLocale(value, locale, options) {
  * is pure, so it moved DOWN into this package (`./number-display.ts`) and
  * `@object-ui/i18n` re-exports it. The duplicate implementation this file used
  * to carry is gone, and with it the drift that duplicate produced — see the
- * percent note in the body, and {@link percentDisplayValue}'s promise, which
- * is true again.
+ * percent note in the body, and the both-halves contract
+ * {@link percentDisplayValue}'s doc comment states.
+ *
+ * ── Date-valued measures (objectui#7178) ──
+ * `min` / `max` over a date or datetime field is a legitimate measure, and it
+ * used to render its stored value verbatim: the guard above returned
+ * `String(v)` for anything non-numeric BEFORE `format` was read, so a KPI tile
+ * showed `2026-07-04T07:00:00.000Z` in `text-2xl` and wrapped it over two
+ * lines. A date-shaped value now routes through {@link formatMeasureDate} to
+ * the display path list cells use; nothing about the numeric path moved.
+ *
+ * ⚠️ What `format` can and cannot say for those values, measured rather than
+ * assumed, and stated PER ARM — a date-shaped measure takes one of two arms in
+ * {@link formatMeasureDate}, and this paragraph used to describe only one of
+ * them while reading as though it described both. That cost a card: an
+ * objectui triage pass read the undifferentiated claim, concluded the defect
+ * was already fixed, and downgraded objectui#8352, which a driven browser run
+ * then refuted. The two arms agree TODAY, but only because #8352 made them; a
+ * universal sentence here is what hid the disagreement before, so it is not
+ * coming back even now that it would be true.
+ *
+ * The shared date path takes a named STYLE, not a date pattern. On BOTH arms
+ * the honoured vocabulary is `'short'` and `'relative'`:
+ *
+ *   - `Field.date` (`2026-09-08`) — honoured by threading `format` into
+ *     `formatDate`'s style parameter. The same words `DateCellRenderer`
+ *     honours from `field.format` for the same field.
+ *   - `Field.datetime` (`2026-09-01T00:00:00.000Z`) — honoured by SELECTING a
+ *     formatter, because `formatDateTime` has no style parameter to thread
+ *     into (objectui#8352). Before that card this arm honoured NOTHING: it
+ *     dropped `format` entirely, so `'relative'` rendered the absolute face.
+ *
+ * Every other string falls to that arm's default face — a PATTERN like
+ * `format: 'YYYY-MM-DD'` is accepted by the schema, reaches this function, and
+ * renders the locale default, and so does `'compact'`, which is
+ * `formatDateTime`'s own vocabulary but not this path's. The pattern part is
+ * not new behaviour introduced here: `plugin-dashboard`'s `recordFields`
+ * already routes a date-shaped `format` into the same style slot and gets the
+ * same locale default. Closing it would mean teaching the shared path a
+ * pattern grammar, which is a change to the path itself and to every list cell
+ * that reads it — not a measure concern, and not either card.
+ *
+ * ⚠️ `'relative'` renders the ABSOLUTE face for a value more than ±7 days from
+ * today, on both arms, by `formatRelativeDate`'s own design. That is correct
+ * output, not a regression — but it also means an out-of-window value is
+ * useless for testing this path, because it renders identically whether
+ * `format` is honoured or dropped. Both readings that missed #8352 were taken
+ * on 40–60-day-old data.
  */
 export function formatMeasure(v, format, currency, percentScale, locale) {
     if (v == null)
         return '—';
-    if (typeof v !== 'number')
-        return String(v);
+    if (typeof v !== 'number') {
+        // Ahead of the `String(v)` short-circuit, which is what made `format` dead
+        // on this path and printed a `min`/`max` over a datetime as a raw
+        // 24-character ISO string (objectui#7178). Deliberately NOT ahead of the
+        // `typeof` test itself: above that line the argument can be a bare number,
+        // and any date test generous enough to consider one would have to decide
+        // whether `1751612400000` is epoch milliseconds — which is how a measure
+        // that counts things starts rendering as a date. Numbers reach the numeral
+        // formatter below byte for byte as before.
+        const asDate = formatMeasureDate(v, format, locale);
+        return asDate ?? String(v);
+    }
     const decimals = format ? (format.split('.')[1]?.match(/0/g)?.length ?? 0) : undefined;
     if (currency) {
         try {
@@ -145,18 +346,24 @@ export function formatMeasure(v, format, currency, percentScale, locale) {
     // A legacy "$" literal in the format string is still honored (explicit author
     // choice) — but it is NOT how a real currency field gets its symbol.
     const legacyDollar = format.includes('$') ? '$' : '';
-    // numeral's "0.0%" multiplies by 100, and a percent field stores a FRACTION
-    // (0.75 ⇒ 75%). Scale to display magnitude the SAME way the list-view cell
-    // renderer does — otherwise an avg of 0.608 renders as "0.6%" instead of
-    // "60.8%", disagreeing with the per-row "75%" the list already shows.
+    // The column's storage is STATED, never read off the value (objectui#11475).
     //
-    // A DECLARED `percentScale` from the server wins outright: the value-magnitude
-    // heuristic below cannot tell a 0–1 ratio of exactly 1 from 1 percentage point
-    // (both are the number 1) and resolves it as "1%", so an SLA rate of full
-    // compliance rendered as "1.0%" instead of "100.0%" (#3136). The heuristic
-    // stays as the fallback for columns that arrive without the annotation.
+    // The server annotates a percentage column with its `percentScale`: a
+    // `derived: { op: 'ratio' }` measure is a fraction by definition, and a
+    // measure over a `percent` field inherits that field's scale
+    // (`AnalyticsResult.fields[].percentScale`). That annotation wins, which is
+    // what fixed an SLA rate of full compliance reading "1.0%" instead of
+    // "100.0%" (#3136).
+    //
+    // A column the server does NOT annotate is, in the contract's words, "not a
+    // percentage", so the only statement left is the author's numeral PATTERN,
+    // and numeral's `%` multiplies by 100: a `'0.0%'` pattern states a fraction.
+    // That is the spec's own reading for a measure over a plain number ("keeps
+    // whatever the author's format string implies", `percentScaleOf`'s doc). It
+    // used to fall to a magnitude guess instead, which read an unannotated `57`
+    // as 57% and an unannotated `0.57` as 57% alike.
     const display = isPercent
-        ? (percentScale ? (percentScale === 'fraction' ? v * 100 : v) : percentDisplayValue(v))
+        ? percentDisplayValue(v, percentScale ?? 'fraction')
         : v;
     // The percent sign is the LOCALE's, not a literal '%' (objectui#4576).
     //
@@ -164,8 +371,8 @@ export function formatMeasure(v, format, currency, percentScale, locale) {
     // a German session read `1.234,5%` from a dashboard measure beside
     // `1.234,5 %` (no-break space — the German percent convention) from a list
     // cell showing the SAME number, because `formatPercent` had gone through
-    // `Intl` since #4553. That contradicted {@link percentDisplayValue}'s own
-    // promise that the two surfaces "can never drift": the SCALING had stopped
+    // `Intl` since #4553. That contradicted the promise {@link percentDisplayValue}
+    // made then, that the two surfaces "can never drift": the SCALING had stopped
     // drifting, the CONVENTION had started. Measured to differ in de, fr, es, ru,
     // sv, cs, fi (no-break space), tr (the sign moves to the FRONT: `%1.234,5`)
     // and ar (its own sign plus U+061C); en, ja, zh were already identical.
@@ -225,12 +432,31 @@ export function formatDimensionValue(v, locale) {
  *
  * `fieldLabel` is injected (rather than imported) so this stays React/i18n-free;
  * callers pass `useSafeFieldLabel().fieldLabel`.
+ *
+ * ## `builtinAggregateLabels` (objectui#7534, sibling of #7258)
+ *
+ * A result field the analytics service minted as a BUILT-IN default measure
+ * carries `builtinAggregate` (objectstack#14492) and a hard-coded English
+ * `label` ('Count'). {@link resolveMeasureLabel} is the ONE resolver for that —
+ * #7258 wired it into `buildChartSeries()`, so a chart legend already read
+ * `计数` on a zh console while the table underneath it, the KPI caption, the
+ * pivot header and the dataset preview all still said `Count`, because they
+ * resolve their titles through here instead.
+ *
+ * Passing the caller's resolved labels closes that gap without a second
+ * resolution order: the discriminator is consulted FIRST (exactly as on the
+ * chart seam), the existing `fieldLabel` convention still applies on top, and
+ * the raw name is still the floor. Omitting the argument reproduces the
+ * previous output byte for byte — an author-declared measure carries no
+ * discriminator and keeps its wire `label` verbatim (objectui#4106), and a
+ * provider-less host resolves no labels and so changes nothing.
  */
-export function buildDatasetFieldHelpers(fields, object, fieldLabel) {
+export function buildDatasetFieldHelpers(fields, object, fieldLabel, builtinAggregateLabels) {
     const fieldByName = new Map((fields ?? []).map((f) => [f.name, f]));
     const measureField = (name) => fieldByName.get(name);
     const headerLabel = (name) => {
-        const fallback = measureField(name)?.label ?? name;
+        const field = measureField(name);
+        const fallback = field ? resolveMeasureLabel(field, builtinAggregateLabels) : name;
         return object && fieldLabel ? fieldLabel(object, name, fallback) : fallback;
     };
     return { measureField, headerLabel };
@@ -241,17 +467,126 @@ export function buildDatasetFieldHelpers(fields, object, fieldLabel) {
  * Each drillable dimension maps to its underlying object field, filtered by the
  * dimension's RAW grouped value (from the server's parallel `drillRawRows`, NOT
  * the visible row which carries the display LABEL — a select/lookup label would
- * mis-filter). An empty/undefined raw value normalizes to `null` (an explicit
- * "is empty" filter). The render-time `runtimeFilter` is ANDed in so the drilled
- * list stays within the same slice the aggregate was computed over.
+ * mis-filter). The render-time `runtimeFilter` is ANDed in so the drilled list
+ * stays within the same slice the aggregate was computed over.
+ *
+ * ## What the EMPTY bucket means, and why it is spelled `{ $null: true }`
+ *
+ * A drill must return exactly the rows the clicked bucket COUNTED. The empty
+ * bucket is the one the aggregate grouped rows with NO VALUE for that dimension
+ * into, so the drill means "this dimension has no value" — NOT "this dimension
+ * holds the literal value null". Those are different row sets, and the
+ * difference is measurable rather than theoretical: against `ValueDataSource`'s
+ * matcher over rows `{owner:'alice'}` / `{owner:null}` / `{}` (no `owner` key),
+ * `['owner','is_null',true]` selects the explicit-null row AND the missing-key
+ * row, while an equality test against `null` selects only the explicit-null one
+ * and drops the row that has no `owner` at all. The bucket counted both, so
+ * is-null is the one that agrees with the number the user clicked.
  *
- * A time-bucketed date dimension (#1752) drills by RANGE, not equality — a
+ * ⇒ every empty authoring — `''`, `null` and `undefined` alike — becomes ONE
+ * spelling, `{ [field]: { $null: true } }`, which `convertFiltersToAST` lowers
+ * to `[field, 'is_null', true]`. The three are not told apart because the
+ * server already merged them into the single bucket the user clicked; there is
+ * no second bucket here for a second spelling to address.
+ *
+ * ⚠️ This value used to be a bare `null`, and that is the defect objectui#9085
+ * records: `convertFiltersToAST`'s OLDEST pinned behaviour is to SKIP a key
+ * whose value is `null` / `undefined` (pinned as "should skip null and
+ * undefined values"), so the constraint was dropped on the way to the wire and
+ * the empty-bucket drill answered with a SUPERSET — every row, silently. The
+ * repair is here at the PRODUCER and ⛔ not at that converter: the skip is
+ * relied on by other producers, and changing it would move every caller's
+ * meaning at once (objectui#9020 ruled it stays). `$null` is not a new
+ * vocabulary either — it is a declared spec operator this dialect already
+ * lowers, and the one spelling that says the same thing on BOTH of
+ * `ValueDataSource`'s routes, where a bare `null` said "only an explicit null"
+ * on the plain-object route and "no constraint at all" on the AST route.
+ *
+ * ⚠️ Known BOUNDARY, unchanged by this and deliberately not widened: the drill
+ * "escape hatch" (the host's `openRecordList`, which serializes a drill filter
+ * into `filter[...]` URL params) has NO spelling for is-null — its operator
+ * vocabulary is equality plus four range bounds — so it drops this condition
+ * exactly as it already dropped the bare `null`, byte-for-byte the same query
+ * string. That surface is a superset today and stays one; closing it needs a
+ * URL-dialect operator on both the write and the read side, which is its own
+ * card.
+ *
+ * A time-bucketed date dimension (objectstack-ai/objectstack#1752) drills by RANGE, not equality — a
  * humanized bucket ("2026-Q2") can't be exact-matched, so the server sends a
  * half-open `[gte, lt)` per date dim in `rawRanges` instead of a raw value.
  * Each becomes an ObjectQL range operator object (`{ $gte, $lt }`) so the drill
  * scopes the list to the clicked time bucket instead of every bucket (which the
  * old date-dim skip degraded to — a superset).
  *
+ * ## The widget filter is CONJOINED, not spread (objectui#9137)
+ *
+ * ⛔ `runtimeFilter` is composed through {@link composeDrillFilter}, never by
+ * spreading it into the object literal below. Spreading is correct only for the
+ * OBJECT arm; this parameter's declared producer sends the ARRAY arm. The
+ * dashboard's `DashboardWidgetSchema.filter` docblock states it outright —
+ * "objectui passes an ObjectQL FilterNode array here, not the spec's
+ * `FilterCondition` envelope" — and `DatasetWidget`'s guard admits it
+ * (`typeof [] === 'object'` and `Object.keys(['x']).length === 1` are both
+ * true), so the array arm reached this line unfiltered. Spreading an ARRAY
+ * yields INDEX keys: `[['region','=','emea']]` became
+ * `{ '0': ['region','=','emea'] }`, which is not a condition anyone can honour.
+ *
+ * ⚠️ MEASURED, and it is NOT the silent superset the sibling sites had. An
+ * index key is not dropped on the way to the wire — every sink in circulation
+ * refuses it LOUDLY, because a bare array is not a legal equality comparand
+ * (objectui#8530 / objectui#8514): `convertFiltersToAST` THROWS
+ * `FilterOperatorError`, `driver-sql` answers `400 INVALID_FILTER`, and
+ * `ValueDataSource`'s matcher selects NOTHING. Against rows scoped
+ * `region = 'emea'`, drilling the `stage = 'won'` bucket returned `(none)`
+ * where the widget alone returned three rows and the conjunction returns the
+ * right two. So the widget's conditions were not merely lost — the drill was
+ * dead in the array arm. The failure is fail-CLOSED here, and this is the third
+ * site of the class objectui#8944 (chart) and objectui#9024 (pivot) removed.
+ *
+ * Three properties of THIS site, each confirmed rather than assumed — it is not
+ * a copy-paste of the other two:
+ *
+ * 1. The return stays a NON-optional `Record<string, unknown>`, because callers
+ *    do not expect `undefined`. `composeDrillFilter` answers `undefined` when
+ *    both sources are empty, so that one case is spelled `?? {}` — the same
+ *    empty object this function already returned for a bucket with no dims.
+ * 2. Range operators SURVIVE. A time-bucketed dim contributes
+ *    `{ $gte, $lt }`, and the composition re-spells it as two `$and` children
+ *    on the same field (`[{d:{$gte}},{d:{$lt}}]`) rather than losing a bound;
+ *    the selected rows are identical, and `serializeDrillFilterParams` walks
+ *    `$and` (objectui#8944) so both bounds still reach `filter[d][gte|lt]`.
+ * 3. `runtimeFilter` is declared `unknown`, which is what the two consumers
+ *    actually hand it — `any` from the dashboard, `Record<string, unknown>`
+ *    from `DatasetReportRenderer`. Widening the PARAMETER is what admits both
+ *    without a cast; ⛔ neither consumer's own type is changed by this.
+ *
+ * ⚠️ Conjunction changes one behaviour deliberately: a field named by BOTH the
+ * widget filter and the clicked bucket used to be overwritten by the bucket,
+ * and now both conditions apply. That is the point — a drill may narrow the
+ * widget's scope and may never widen it — and it is the posture the chart and
+ * pivot sites already ship.
+ *
+ * ## ⭐ Only the COMPOSING leg moves; the identity leg is left byte-identical
+ *
+ * The sink is asked only when there IS a second source. That is this site's
+ * one departure from `ObjectChart` / `ObjectPivotTable`, which call it
+ * unconditionally, and it is deliberate rather than a shortcut: the spread
+ * this replaces only ever ran when `runtimeFilter` was present, so the
+ * no-widget-filter path carries no defect to repair.
+ *
+ * ⚠️ Routing it anyway was MEASURED and rejected. `composeDrillFilter`'s own
+ * note — "a lone surviving source lowers back to exactly the flat object the
+ * spread produced" — holds only for a source of ONE condition; with two it
+ * lowers to `$and`, because `convertFiltersToAST` emits one node per condition
+ * and `parseFilterAST` cannot fold two nodes back onto one key. A `{ $gte,
+ * $lt }` range is two conditions on its own, so even a single date bucket
+ * re-spells. Sending the identity leg through the sink therefore re-shaped
+ * every multi-condition dataset drill — including the controls objectui#9085,
+ * objectui#4056, objectstack#5473 and objectstack-ai/objectstack#1752 pin on paths this card does not
+ * touch — for ZERO change in the rows selected or the `filter[...]` params
+ * emitted. ⛔ Do not "simplify" this to an unconditional call without redoing
+ * that measurement.
+ *
  * Shared by the dashboard `DatasetWidget` and the report renderer so a drill
  * filters identically (and correctly, including lookups) on both surfaces.
  */
@@ -259,7 +594,12 @@ export function buildDatasetDrillFilter(rawRow, drillDims, dimensionFields, runt
     const drillFilter = {};
     for (const d of drillDims) {
         const raw = rawRow?.[d];
-        drillFilter[dimensionFields[d]] = raw === '' || raw === undefined ? null : raw;
+        // `null` is in this test, not only `''` / `undefined`: JSON cannot carry
+        // `undefined`, so a SQL NULL grouped value arrives over the wire AS `null`
+        // — it is the empty bucket's most common shape, and leaving it out would
+        // fix the defect for the spellings the wire rarely uses.
+        drillFilter[dimensionFields[d]] =
+            raw === '' || raw === null || raw === undefined ? { $null: true } : raw;
     }
     if (rawRanges) {
         for (const r of Object.values(rawRanges)) {
@@ -267,5 +607,8 @@ export function buildDatasetDrillFilter(rawRow, drillDims, dimensionFields, runt
                 drillFilter[r.field] = { $gte: r.gte, $lt: r.lt };
         }
     }
-    return runtimeFilter ? { ...runtimeFilter, ...drillFilter } : drillFilter;
+    // ⛔ The composition is `composeDrillFilter`'s, never a spread. The identity
+    // leg is deliberately left alone — see "only the composing leg moves" above.
+    // `?? {}` is property 1: the empty/empty answer stays an empty object.
+    return runtimeFilter ? composeDrillFilter(runtimeFilter, drillFilter) ?? {} : drillFilter;
 }