docs(objectstack-ui): standardize the inline-vs-dataset chart rule (#2502)

Encode the chart-data expressibility rule into the objectstack-ui skill,
pinned to what the analytics/dataset engine actually supports rather than the
draft prose in the issue.

- SKILL.md (Dataset-Bound Widgets): reframe around the real boundary. Every
  persisted chart is dataset-backed (ADR-0021 single-form cutover), so the
  decision is not "inline vs dataset" but "does the need fit the dataset
  envelope". Adds a Level A ceiling table (fits -> dataset; beyond -> Cube /
  stored rollup field / app code), an iron-rule callout, standardized answers
  for the recurring ambiguous cases (parent->child rollup, computed column,
  lookup filter, dashboard filter -> #2501), and a Level B note that naming a
  dataset is governance, not expressibility. Also corrects the Dashboards
  intro that still described the removed inline widget shape.
- evals/analytics-inline-vs-dataset.json: first working evals for this skill,
  encoding the key decisions in the proven must_contain/must_not_contain
  format.
- dashboard.zod.ts: fix the @example that still showed the removed inline
  widget shape (object/valueField/aggregate), which contradicted the
  now-required `dataset` field declared a few lines below it.

Follow-ups intentionally left out of this change: service-analytics/README.md
documents a retired API surface (HAVING, old IAnalyticsService methods) and
needs a full rewrite; the inline-embedded-dataset schema change and the
save-as-dataset promotion flow are larger, cross-repo design items tracked on
the issue.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016QR9X5jcv9tEQQX5V9UZDu

Author: claude

packages/spec/src/ui/dashboard.zod.ts

@@ -276,26 +276,26 @@ export const GlobalFilterSchema = lazySchema(() => z.object({
  * Dashboard Schema
  * Represents a page containing multiple visualizations.
  * 
- * @example Sales Executive Dashboard
+ * @example Sales Executive Dashboard (ADR-0021: widgets bind a `dataset` and
+ * select `dimensions`/`values` BY NAME — each metric/measure is defined once in
+ * `defineDataset`, never inline on the widget)
  * {
  *   name: "sales_overview",
  *   label: "Sales Executive Overview",
  *   widgets: [
  *     {
  *       title: "Total Pipe",
  *       type: "metric",
- *       object: "opportunity",
- *       valueField: "amount",
- *       aggregate: "sum",
+ *       dataset: "opportunity_metrics",
+ *       values: ["amount_sum"],
  *       layout: { x: 0, y: 0, w: 3, h: 2 }
  *     },
  *     {
  *       title: "Revenue by Region",
  *       type: "bar",
- *       object: "order",
- *       categoryField: "region",
- *       valueField: "total",
- *       aggregate: "sum",
+ *       dataset: "order_metrics",
+ *       dimensions: ["region"],
+ *       values: ["total_sum"],
  *       layout: { x: 3, y: 0, w: 6, h: 4 }
  *     }
  *   ]

skills/objectstack-ui/SKILL.md

@@ -531,8 +531,9 @@ export const CrmApp = App.create({
 ## Dashboards
 
 Dashboards are a grid of widgets (`columns` × `rowHeight`) sharing a
-`dateRange` scrubber and `globalFilters`. Each widget declares an `object`,
-an `aggregate` measure or chart spec, and a `layout: {x,y,w,h}`.
+`dateRange` scrubber and `globalFilters`. Each widget **binds a `dataset`** and
+selects named `dimensions` + `values`, picks a chart `type`, and sets a
+`layout: {x,y,w,h}` (ADR-0021).
 
 ### Widget Types
 
@@ -552,12 +553,15 @@ modifier; date bucketing comes from the bound dataset dimension's
 
 ### Dataset-Bound Widgets
 
-For shared metrics, prefer the ADR-0021 dataset shape over per-widget inline
-queries. A widget binds to `dataset` and selects named `dimensions` and
-`values`; the dataset owns the base object, allowed joins, intrinsic filter,
-dimensions, and certified measures. Reports bind the same way (`dataset` +
-`rows` + `values` + `runtimeFilter`). Full guide: **Guides → Analytics Datasets**
-(`content/docs/data-modeling/analytics.mdx`).
+Every persisted chart is **dataset-backed** (ADR-0021 single-form cutover): a
+dashboard widget, a report, and a list `type:'chart'` view all **bind a `dataset`
+and select named `dimensions` + `values`**; the dataset owns the base object,
+allowed joins, intrinsic filter, dimensions, and certified measures. The legacy
+per-widget inline query (`object` + `categoryField` + `valueField` + `aggregate`)
+**was removed** — a widget now requires `dataset` + `values`; the inline fields are
+dropped and a widget lacking `dataset` fails `os validate`. Reports bind the same
+way (`dataset` + `rows` + `values` + `runtimeFilter`). Full guide: **Guides →
+Analytics Datasets** (`content/docs/data-modeling/analytics.mdx`).
 
 A widget's presentation-scope `filter` flows into the query as the runtime
 filter; keep `filter` on the widget when binding a dataset.
@@ -574,9 +578,59 @@ filter; keep `filter` on the widget when binding a dataset.
 }
 ```
 
-- Dataset-bound widgets need at least one `values` entry.
-- Do not mix `dataset` with inline `object` / `valueField` / `aggregate`
-  unless you are intentionally keeping a legacy inline widget shape.
+**The real decision is not "inline vs dataset" — it is "can a dataset express
+this?"** The shape is already fixed (always a dataset), so what you decide is
+whether the *data need fits the dataset envelope*, and if not, which lower layer to
+escalate to. Decide on **expressibility**; reuse/governance is Level B.
+
+**Level A — the dataset envelope:**
+
+| Fits a dataset → author one | Beyond the envelope → escalate |
+|:--|:--|
+| one base object + **to-one** joins (`include`, ≤3 hops) | a join that **changes grain** / a **to-many** rollup onto the parent |
+| 0..N dimensions; date-bucket `day/week/month/quarter/year` | a **computed dimension** / CASE bucket / numeric bin |
+| measures `count/sum/avg/min/max/count_distinct` | `array_agg`/`string_agg` or any custom-SQL metric |
+| **derived measures** — `ratio/sum/difference/product` of other measures | scalar math on raw fields (`amount*0.8`), aggregate-of-aggregate |
+| WHERE (`$and/$or/$not` on the base object) + measure-scoped filters | **HAVING** (filtering the aggregate result) |
+| `compareTo` (previous period/year) + `totals` (matrix subtotals) | **window** (rank, running total, lag/lead, %-of-total); **union**; reshaping params |
+
+> **The iron rule:** a dataset is a governed, *narrow* semantic layer — NOT a
+> general analytics escape hatch (no raw SQL, no hand-authored joins, no
+> window/having). If the need is in the right column, a dataset **cannot** express
+> it — escalate to a hand-authored **Cube** (raw SQL / explicit joins), a **stored
+> rollup or formula field** on the object (to-many rollups, computed columns), or
+> app code. Do not force it into a dataset: it fails to compile or renders an empty
+> series.
+
+Standardized answers to the recurring ambiguous cases:
+
+- **"Count of child tasks per project."** Base the dataset on the **child** (`task`)
+  and group by the parent lookup (`project`) — grain = child. A rollup onto the
+  **parent** grain ("on `project`, count related tasks") is a to-many rollup:
+  **not** dataset-expressible — use a **stored rollup field** on `project`.
+- **To-one enrichment** ("revenue by `account.industry`") is fine and does **not**
+  change grain — put `account` in `include`, add `account.industry` as a dimension.
+- **Computed column.** Formatting/currency → a measure's `format`/`currency`;
+  arithmetic over declared measures (`margin = difference(revenue, cost)`) → a
+  **derived measure**; CASE / bins / `revenue*0.8` / computed dimensions → **not** a
+  dataset.
+- **Filter by a parent's attribute** → model it as a **dimension** (guaranteed to
+  join); a lookup-path *filter* is not a reliable analytics-path construct.
+- **A dashboard filter driving several charts** (date/region) → **not** a dataset:
+  a dashboard variable + per-chart `filterBindings` broadcast into each chart's
+  WHERE (#2501). A dataset is implied only when a parameter **reshapes** the query
+  (grain/window/join) — and those are beyond the envelope anyway.
+
+**Level B — naming is governance, not expressibility.** An inline dataset draft
+(Studio Live Canvas) and a saved named dataset have **identical** expressibility;
+naming one is a reuse/governance call (canonical/shared metric, RLS, shared
+labels/formats → `defineDataset`). Persisted widgets already require a named
+dataset, so Level B only surfaces in Studio previews and hand-coded react-page
+`<ObjectChart>` blocks (a single-object inline `aggregate`, no dataset binding).
+
+- Dataset-bound widgets need at least one `values` entry, and every
+  `dataset`/`dimensions`/`values` name must resolve to its `defineDataset` —
+  `os validate` fails on an unresolved name (an empty chart otherwise).
 - Studio's Dashboard Widget Inspector can author per-widget `dataset`,
   `dimensions`, and `values`; curated metadata-admin forms merge
   server-only fields back into the payload, so saving through Studio should

skills/objectstack-ui/evals/analytics-inline-vs-dataset.json

@@ -0,0 +1,56 @@
+{
+  "skill_name": "objectstack-ui",
+  "description": "Inline-vs-dataset / dataset-envelope-ceiling decisions (issue #2502). Every persisted chart is dataset-backed (ADR-0021); the real decision is whether the data need fits the dataset envelope, and if not, which lower layer to escalate to. Naming an inline draft vs a saved dataset is governance, not expressibility.",
+  "evals": [
+    {
+      "id": 1,
+      "prompt": "On the sales dashboard, add several widgets that all report the same certified 'revenue' metric (sum of amount) — one grouped by region, one by month, one as a KPI total. Emit the ObjectStack metadata.",
+      "expected_output": "Defines the metric ONCE in a dataset (defineDataset with an amount_sum measure + region and a month-granularity date dimension), then binds each widget to that dataset and selects dimensions/values BY NAME. Does not restate the aggregate inline per widget (the ADR-0021 inline object/valueField/aggregate shape was removed).",
+      "files": [],
+      "assertions": {
+        "must_contain": ["defineDataset", "dataset", "dimensions", "values"],
+        "must_not_contain": ["valueField", "categoryField"]
+      }
+    },
+    {
+      "id": 2,
+      "prompt": "Add a dashboard chart showing 30-day rolling revenue (a moving average / running total over time). Model the data need.",
+      "expected_output": "Recognizes that a window/rolling calculation is BEYOND the dataset envelope (a dataset has no window/having grammar and compareTo only does previous-period/year). Escalates to a hand-authored Cube (raw SQL) or a stored/materialized rolling field — it does not pretend a dataset date-bucket or a plain measure expresses a moving window.",
+      "files": [],
+      "assertions": {
+        "must_contain": ["window", "Cube"],
+        "must_not_contain": ["dateGranularity"]
+      }
+    },
+    {
+      "id": 3,
+      "prompt": "On a projects dashboard I want each project row to show its count of related child tasks (task has a lookup to project). Model it.",
+      "expected_output": "Explains that a to-many rollup ONTO the parent grain is not expressible in a dataset (datasets only do to-one joins; to-many traversal is out of scope). Recommends a stored rollup/summary field on project, OR basing the dataset on the child task and grouping by the project lookup (grain = child). Does not author a to-many join in include.",
+      "files": [],
+      "assertions": {
+        "must_contain": ["rollup", "to-many"],
+        "must_not_contain": []
+      }
+    },
+    {
+      "id": 4,
+      "prompt": "A dashboard has a top-level date-range filter that must re-scope six charts built over different objects. How should each chart pick up the filter? Emit the metadata.",
+      "expected_output": "Uses a dashboard-level variable/filter plus per-chart filterBindings mapping the filter to each widget's own date field (e.g. invoice→created_at, account→signed_at), broadcast into each chart's WHERE (#2501). Charts stay self-contained; it does NOT invent a shared dataset for the filter (a dataset is only implied when a parameter reshapes the query grain/window/join).",
+      "files": [],
+      "assertions": {
+        "must_contain": ["filterBindings"],
+        "must_not_contain": []
+      }
+    },
+    {
+      "id": 5,
+      "prompt": "In a kind:'react' page, drop a quick one-off bar chart of tasks grouped by status. Write the block.",
+      "expected_output": "Uses the inline <ObjectChart objectName=\"task\" aggregate={{ field: '_id', function: 'count', groupBy: 'status' }} /> block directly — an ad-hoc in-page chart does not require a dataset, and the react ObjectChart block has no dataset prop. Naming a dataset would only be for reuse/governance (Level B), not needed here.",
+      "files": [],
+      "assertions": {
+        "must_contain": ["ObjectChart", "aggregate", "groupBy"],
+        "must_not_contain": ["defineDataset"]
+      }
+    }
+  ]
+}