Editor: view-scoped data context — per-view dataset columns in hints and transforms

docs/IMPLEMENTATION-PLAN.md

@@ -75,9 +75,9 @@ This is the at-a-glance list; keep it in sync with them.
 **Next (flagged for build):**
 
 - **Multi-view data model** ([`multi-view-data-model-scope.md`](exploration/multi-view-data-model-scope.md)) —
-  durable composition support across the data-facing features. M1 (Vega-Lite-fidelity
-  reference classifier, `core/spec-data`) is done; M2–M4 extend the editor data context,
-  Extract, and the data inspector to be view-scoped.
+  durable composition support across the data-facing features. Done: the Vega-Lite-fidelity
+  reference classifier (`core/spec-data`) and the view-scoped editor data context. Remaining:
+  per-view data inspection (`DataInspector` view selector), then view-scoped Extract.
 - **Chart Builder · 3B starter examples** ([`chart-builder-enhancement-scope.md`](exploration/chart-builder-enhancement-scope.md) §3) —
   a small set of curated starters, one per covered FT intent. Reshaped by 3C: a
   builder-openable starter must reference a dataset, so it ships paired sample datasets (or is

docs/architecture/08-vega-editor-techniques.md

@@ -286,16 +286,22 @@ thin app-layer services.
 
 - `spec-transforms` — wrap a view in `layer`/`hconcat`/`vconcat`/`facet`/`repeat`; collapse
   a single-child composition (`unwrapSingleton`). Object-in/object-out.
-- `spec-cursor` — `findViewRange` (cursor offset → the enclosing view's byte range) and
-  `valueKeyAtOffset`/`stringValueAtOffset` (the JSON context at the cursor), over `jsonc-parser`.
-- `spec-fields` — field names a spec's own transforms introduce (their `as`).
-- `spec-inline-data` — the inline rows a spec carries (`data.values`, a `datasets` entry).
+- `spec-data` — the Vega-Lite data model: classify a `data` block (`classifyData`, mirroring
+  `isNamedData`), the library reference name (`libraryRefName`), and the data binding in scope
+  at a cursor path (`dataBindingAtPath`, honoring a view's data inheritance from its parent).
+- `spec-cursor` — `findViewRange` (cursor offset → the enclosing view's byte range),
+  `pathAtOffset` (cursor → JSON path), and `valueKeyAtOffset`/`stringValueAtOffset` (the JSON
+  context at the cursor), over `jsonc-parser`.
+- `spec-fields` — field names a spec's transforms introduce (their `as`), scoped to a
+  cursor's ancestor chain (`derivedFieldNamesAtPath`).
+- `spec-inline-data` — the rows a specific `data` binding carries for profiling
+  (`rowsForDataBinding`: inline `values`, or a self-defined `datasets` entry).
 - `spec-insert` — composition arrays + appending a view to one.
 
 **Services (app, store-aware via `getState`):** `spec-transform-actions` (the
 wrap/simplify/add-view operations and their surfaces), `spec-dataset-hints` (completion,
-hover, inlay providers), `active-dataset` (`dataInfo()` — the columns/types/stats + derived
-fields the draft sees). `SpecEditor` does the wiring.
+hover, inlay providers), `active-dataset` (`dataInfoAt(text, offset)` — the columns/types/stats
+plus derived fields the draft sees at the cursor). `SpecEditor` does the wiring.
 
 Decision rules:
 
@@ -312,12 +318,17 @@ Decision rules:
 - **One edit path.** The lightbulb returns a `WorkspaceEdit` (no editor handle); the toolbar
   and palette use `executeEdits`. Both build the replacement through the same
   serialize-and-reindent step, bracketed by `pushUndoStop`, so ⌘Z restores the prior text.
-- **Field source for hints.** `dataInfo()` reads columns from the named library dataset the
-  draft references, else profiles the spec's **inline data on the fly** (`spec-inline-data` +
-  `core/profile`) — a "ghost dataset" with nothing stored — and adds the spec's derived
-  fields. Memoized by draft text, since providers fire per keystroke and per scroll. Out of
-  scope: data-dependent derived columns (`pivot`/`lookup` output) and `url`/CSV-string inline
-  data, which need the pipeline run or format-aware parsing.
+- **Field source for hints (view-scoped).** `dataInfoAt(text, offset)` resolves the data
+  binding of the cursor's **nearest enclosing view** (`dataBindingAtPath` — a child inherits a
+  parent's data unless it declares its own), then its columns: a named **library dataset**
+  (matched case-insensitively, like the renderer), else the binding's **inline rows** profiled
+  on the fly (`rowsForDataBinding` + `core/profile`) — a "ghost dataset" with nothing stored.
+  Derived fields come from that view's and its ancestors' transforms only
+  (`derivedFieldNamesAtPath`), so a sibling view's `calculate` does not leak in. Profiling is
+  memoized per (draft text, enclosing view), so the inlay provider's many per-line queries
+  profile once. A composed spec whose views bind different datasets therefore gets the right
+  columns per view. Out of scope: data-dependent derived columns (`pivot`/`lookup` output) and
+  `url`/CSV-string inline data, which need the pipeline run or format-aware parsing.
 - **No unknown-field diagnostic.** Hints are additive and forgiving, so over- or
   under-listing costs nothing; a "field not in data" squiggle would false-positive on every
   derived or data-dependent field, so there is deliberately none.

docs/exploration/multi-view-data-model-scope.md

@@ -66,18 +66,21 @@ collection (`spec-fields`), config baking (`spec-config`), standalone export
 
 - **M1 — data-model foundation** ✅ — `core/spec-data` classifier mirroring
   `isNamedData`; `spec-refs` + `rendering` routed through it. Closes clash 1.
-- **M2 — view-scoped editor context** — pure `dataContextAtPath(spec, path)`:
-  climb the cursor's JSON path to the nearest enclosing `data` (honoring
-  Vega-Lite's parent→child data inheritance), classify it, and collect
-  ancestor-chain derived fields. Rework `active-dataset` to be cursor-scoped and
-  thread the offset through the three Monaco providers and the facet/repeat
-  defaults. Resolve columns for every form: library ref, inline `values`,
-  named-inline, self-defined `datasets`, url (no static rows), generator (none).
+- **M2 — view-scoped editor context** ✅ — `dataBindingAtPath` (climb the cursor's
+  JSON path to the nearest enclosing `data`, honoring Vega-Lite's parent→child
+  inheritance) + `derivedFieldNamesAtPath` (ancestor-chain `as` outputs).
+  `active-dataset` is cursor-scoped (`dataInfoAt(text, offset)`), resolving columns
+  for every form (library ref case-insensitive, inline, named-inline, self-defined
+  `datasets`, url/generator → none); the three Monaco providers and the
+  facet/repeat defaults pass the cursor offset.
+- **M4 — multi-view inspection** (pulled ahead of M3) — group the dataflow's
+  datasets into per-view input/resolved pairs (`result-data`) and add a view
+  selector to `DataInspector` (new interactive widget → `/council`). The data
+  inspector executes the live pipeline, so it is the place to review post-transform
+  rows (melt/fold/pivot/aggregate); today it surfaces one heuristic input/resolved
+  pair across the whole composition, with no per-view choice.
 - **M3 — view-scoped extract** — seed Extract from the focused view's inline data
   (reusing the cursor-scope machinery) and rewrite that view's `data`.
-- **M4 — multi-view inspection** — group the dataflow's datasets into per-view
-  input/resolved pairs (`result-data`) and add a view selector to `DataInspector`
-  (new interactive widget → `/council`).
 
 Delivery is incremental, one milestone per commit, verified against real behavior.
 The consolidated data-model contract write-up into `docs/architecture` (05/08)

src/app/services/active-dataset.test.ts

@@ -0,0 +1,113 @@
+import { beforeEach, describe, expect, test } from 'vitest';
+import { createDataset } from '@core/dataset';
+import { useDatasetStore } from '../stores/DatasetStore';
+import { dataInfoAt } from './active-dataset';
+
+/** Seed a library dataset, profiled from inline rows. */
+function seed(name: string, rows: Record<string, unknown>[]): void {
+  useDatasetStore
+    .getState()
+    .add(createDataset({ name, data: rows, format: 'json', source: 'inline' }));
+}
+
+/** An offset inside the first occurrence of `marker` in `text`. */
+function at(text: string, marker: string): number {
+  return text.indexOf(marker) + 1;
+}
+
+const names = (cols: ReadonlyArray<{ name: string }>) => cols.map((c) => c.name).sort();
+
+beforeEach(() => {
+  useDatasetStore.getState().reset();
+});
+
+describe('dataInfoAt — view-scoped data context', () => {
+  test('resolves a library binding case-insensitively', () => {
+    seed('Sales', [{ region: 'W', revenue: 10 }]);
+    const text = '{ "data": { "name": "sales" }, "encoding": { "x": { "field": "MK" } } }';
+    const info = dataInfoAt(text, at(text, 'MK'));
+    expect(info.name).toBe('Sales');
+    expect(names(info.columnTypes)).toEqual(['region', 'revenue']);
+  });
+
+  test('each layer sees its own dataset, not a sibling view’s', () => {
+    seed('Sales', [{ region: 'W', revenue: 10 }]);
+    seed('Regions', [{ region: 'W', population: 100 }]);
+    const text = JSON.stringify(
+      {
+        layer: [
+          { data: { name: 'Sales' }, encoding: { x: { field: 'sales_mk' } } },
+          { data: { name: 'Regions' }, encoding: { y: { field: 'regions_mk' } } },
+        ],
+      },
+      null,
+      2,
+    );
+    const inLayer0 = dataInfoAt(text, at(text, 'sales_mk'));
+    const inLayer1 = dataInfoAt(text, at(text, 'regions_mk'));
+    expect(inLayer0.name).toBe('Sales');
+    expect(names(inLayer0.columnTypes)).toEqual(['region', 'revenue']);
+    expect(inLayer1.name).toBe('Regions');
+    expect(names(inLayer1.columnTypes)).toEqual(['population', 'region']);
+  });
+
+  test('a child without its own data inherits the parent binding', () => {
+    seed('Sales', [{ region: 'W', revenue: 10 }]);
+    const text = JSON.stringify(
+      { data: { name: 'Sales' }, layer: [{ encoding: { x: { field: 'mk' } } }] },
+      null,
+      2,
+    );
+    const info = dataInfoAt(text, at(text, '"mk"'));
+    expect(info.name).toBe('Sales');
+    expect(names(info.columnTypes)).toEqual(['region', 'revenue']);
+  });
+
+  test('profiles an inline binding (ghost dataset, no name)', () => {
+    const text =
+      '{ "data": { "values": [{ "a": 1, "b": 2 }] }, "encoding": { "x": { "field": "MK" } } }';
+    const info = dataInfoAt(text, at(text, 'MK'));
+    expect(info.name).toBeNull();
+    expect(names(info.columnTypes)).toEqual(['a', 'b']);
+  });
+
+  test('profiles a self-defined top-level datasets binding', () => {
+    const text = JSON.stringify(
+      {
+        datasets: { local: [{ c: 1, d: 2 }] },
+        data: { name: 'local' },
+        encoding: { x: { field: 'mk' } },
+      },
+      null,
+      2,
+    );
+    const info = dataInfoAt(text, at(text, '"mk"'));
+    expect(info.name).toBeNull();
+    expect(names(info.columnTypes)).toEqual(['c', 'd']);
+  });
+
+  test('offers ancestor-derived fields, not a sibling view’s', () => {
+    seed('Sales', [{ region: 'W', revenue: 10 }]);
+    const text = JSON.stringify(
+      {
+        data: { name: 'Sales' },
+        transform: [{ calculate: 'datum.revenue * 2', as: 'shared' }],
+        layer: [
+          { transform: [{ calculate: 'x', as: 'inner' }], encoding: { x: { field: 'mk0' } } },
+          { transform: [{ calculate: 'y', as: 'sibling' }], encoding: { y: { field: 'mk1' } } },
+        ],
+      },
+      null,
+      2,
+    );
+    const fields = dataInfoAt(text, at(text, 'mk0')).fields.map((f) => f.name);
+    expect(fields).toContain('shared'); // top-level (inherited)
+    expect(fields).toContain('inner'); // this view
+    expect(fields).not.toContain('sibling'); // the other layer
+  });
+
+  test('a url/generator binding yields no static columns', () => {
+    const url = '{ "data": { "url": "x.csv" }, "encoding": { "x": { "field": "MK" } } }';
+    expect(dataInfoAt(url, at(url, 'MK')).columnTypes).toEqual([]);
+  });
+});

src/app/services/active-dataset.ts

@@ -1,29 +1,35 @@
 /**
- * The data context of the active draft (docs/architecture/08 → editor
- * augmentation): the columns/types/stats the spec can reference, and the fields
- * the editor hints offer. One resolver, shared by the transform actions
- * (facet/repeat field defaults) and the dataset-aware hints
+ * The data context of the active draft **at the cursor** (docs/architecture/08 →
+ * editor augmentation): the columns/types/stats the spec can reference where the
+ * cursor sits, and the fields the editor hints offer. One resolver, shared by the
+ * transform actions (facet/repeat field defaults) and the dataset-aware hints
  * (completion/hover/inlay).
  *
- * Source, in order: a **named library dataset** the draft references (its stored
- * profile), else the spec's **inline data** profiled on the fly (the "ghost
- * dataset" — `core/spec-inline-data` + `core/profile`, nothing stored). On top of
- * either, the spec's **derived** fields (transform `as`, `core/spec-fields`) are
- * added, so a `calculate` output is offered alongside the data columns.
+ * View-scoped: a composed spec (layer/concat/facet/repeat) can bind a different
+ * dataset per view, and Vega-Lite inherits a parent view's data into its children.
+ * So the context is resolved at a JSON path — the data binding of the nearest
+ * enclosing view (`dataBindingAtPath`), plus the fields derived by that view's and
+ * its ancestors' transforms (`derivedFieldNamesAtPath`).
  *
- * The result is memoized by draft text: providers fire often (every keystroke
- * completes, inlay refreshes on scroll), and profiling inline data each time would
- * be wasteful. App layer — reads stores via `getState`, outside React.
+ * Source columns, in order: a named **library dataset** the binding references
+ * (its stored profile, matched case-insensitively like the renderer), else the
+ * binding's **inline rows** profiled on the fly — inline `values` or a self-defined
+ * top-level `datasets` entry (the "ghost dataset", `core/spec-inline-data` +
+ * `core/profile`, nothing stored). URL and generator bindings have no static
+ * columns.
+ *
+ * Profiling is memoized per (draft text, enclosing view) so the inlay provider —
+ * which queries many field lines at one draft — never re-profiles the same inline
+ * data. App layer — reads stores via `getState`, outside React.
  */
 
-import type { Dataset } from '@core/dataset';
 import { type ColumnStats, profileData } from '@core/profile';
-import { derivedFieldNames } from '@core/spec-fields';
-import { inlineDataRows } from '@core/spec-inline-data';
-import { extractDatasetRefs } from '@core/spec-refs';
+import { pathAtOffset } from '@core/spec-cursor';
+import { dataBindingAtPath, libraryRefName, selfDefinedNames } from '@core/spec-data';
+import { derivedFieldNamesAtPath } from '@core/spec-fields';
+import { rowsForDataBinding } from '@core/spec-inline-data';
 import type { ColumnType } from '@core/type-inference';
 import { useDatasetStore } from '../stores/DatasetStore';
-import { useSnippetStore } from '../stores/SnippetStore';
 
 /** A field offered to the editor hints. */
 export interface FieldHint {
@@ -34,7 +40,7 @@ export interface FieldHint {
   derived: boolean;
 }
 
-/** Everything the hints need about the active draft's data. */
+/** Everything the hints need about the active draft's data at a cursor. */
 export interface DataInfo {
   /** The library dataset's name, or null for inline data (the ghost dataset). */
   name: string | null;
@@ -44,7 +50,15 @@ export interface DataInfo {
   fields: ReadonlyArray<FieldHint>;
 }
 
-const EMPTY: DataInfo = { name: null, columnTypes: [], columnStats: [], fields: [] };
+/** The source columns of one view's data binding (the expensive, cached part). */
+interface SourceColumns {
+  name: string | null;
+  columnTypes: ReadonlyArray<{ name: string; type: ColumnType }>;
+  columnStats: ReadonlyArray<ColumnStats>;
+}
+
+const EMPTY_SOURCE: SourceColumns = { name: null, columnTypes: [], columnStats: [] };
+const EMPTY: DataInfo = { ...EMPTY_SOURCE, fields: [] };
 
 function safeParse(text: string): unknown {
   try {
@@ -54,74 +68,122 @@ function safeParse(text: string): unknown {
   }
 }
 
-/** The first library dataset the draft references, or null. */
-function libraryDataset(draftText: string): Dataset | null {
-  const refs = extractDatasetRefs(draftText);
-  if (refs.length === 0) return null;
-  const { datasets } = useDatasetStore.getState();
-  for (const ref of refs) {
-    const ds = datasets.find((d) => d.name === ref);
-    if (ds) return ds;
-  }
-  return null;
+// The parsed draft, memoized by text — both the binding walk and the derived-field
+// walk read it, and providers fire often at one stable draft.
+let parseCache: { text: string; spec: unknown } | null = null;
+function parsedSpec(text: string): unknown {
+  if (parseCache?.text === text) return parseCache.spec;
+  const spec = safeParse(text);
+  parseCache = { text, spec };
+  return spec;
 }
 
-function compute(draftText: string): DataInfo {
-  const spec = safeParse(draftText);
-
-  // Source columns: a named library dataset wins; otherwise profile inline data.
-  let name: string | null = null;
-  let columnTypes: ReadonlyArray<{ name: string; type: ColumnType }> = [];
-  let columnStats: ReadonlyArray<ColumnStats> = [];
-  const library = libraryDataset(draftText);
-  if (library && library.columnTypes.length > 0) {
-    ({ name, columnTypes, columnStats } = library);
-  } else {
-    const rows = inlineDataRows(spec);
-    if (rows) {
-      const profile = profileData(rows, 0);
-      columnTypes = profile.columnTypes;
-      columnStats = profile.columnStats;
-    }
+/** The stored profile of the library dataset a binding references, or null. */
+function libraryProfile(refName: string): SourceColumns | null {
+  const lower = refName.toLowerCase();
+  const ds = useDatasetStore.getState().datasets.find((d) => d.name.toLowerCase() === lower);
+  if (ds && ds.columnTypes.length > 0) {
+    return { name: ds.name, columnTypes: ds.columnTypes, columnStats: ds.columnStats };
   }
+  return null; // unknown name, or a dataset with no profiled columns yet
+}
 
-  const fields: FieldHint[] = columnTypes.map((c) => ({
+/** Source columns for one binding: library profile, else inline rows profiled. */
+function resolveSource(spec: unknown, data: unknown): SourceColumns {
+  const refName = libraryRefName(data, selfDefinedNames(spec));
+  if (refName !== null) return libraryProfile(refName) ?? EMPTY_SOURCE;
+  const rows = rowsForDataBinding(spec, data);
+  if (!rows) return EMPTY_SOURCE; // url / generator / library miss / CSV-string values
+  const profile = profileData(rows, 0);
+  return { name: null, columnTypes: profile.columnTypes, columnStats: profile.columnStats };
+}
+
+// Source columns memoized by (text, enclosing view). Keyed by the binding's anchor
+// path so the inlay provider's many per-line queries within one view profile once.
+// A re-import that changes a dataset's columns under unchanged text serves stale
+// hints until the next keystroke — harmless, since hints are additive.
+let sourceCache: { text: string; byAnchor: Map<string, SourceColumns> } | null = null;
+function sourceColumns(
+  text: string,
+  spec: unknown,
+  anchorKey: string,
+  data: unknown,
+): SourceColumns {
+  if (!sourceCache || sourceCache.text !== text) sourceCache = { text, byAnchor: new Map() };
+  const hit = sourceCache.byAnchor.get(anchorKey);
+  if (hit) return hit;
+  const result = resolveSource(spec, data);
+  sourceCache.byAnchor.set(anchorKey, result);
+  return result;
+}
+
+/**
+ * The shared resolve step: the source columns of the binding in scope at `offset`,
+ * with the parsed spec and cursor path so callers can layer derived fields on top
+ * without re-parsing or re-walking.
+ */
+function resolveAt(
+  text: string,
+  offset: number,
+): {
+  source: SourceColumns;
+  spec: unknown;
+  path: ReadonlyArray<string | number>;
+} {
+  const spec = parsedSpec(text);
+  const path = pathAtOffset(text, offset);
+  const binding = dataBindingAtPath(spec, path);
+  const source = binding
+    ? sourceColumns(text, spec, JSON.stringify(binding.anchorPath), binding.data)
+    : EMPTY_SOURCE;
+  return { source, spec, path };
+}
+
+/** Source columns of the binding in scope at `offset` — no derived-field merge. */
+function sourceAt(text: string, offset: number): SourceColumns {
+  return text.trim() === '' ? EMPTY_SOURCE : resolveAt(text, offset).source;
+}
+
+/** The data context at `offset` in the editor's draft `text`. */
+export function dataInfoAt(text: string, offset: number): DataInfo {
+  if (text.trim() === '') return EMPTY;
+  const { source, spec, path } = resolveAt(text, offset);
+
+  const fields: FieldHint[] = source.columnTypes.map((c) => ({
     name: c.name,
     type: c.type,
     derived: false,
   }));
   const seen = new Set(fields.map((f) => f.name));
-  for (const derived of derivedFieldNames(spec)) {
+  for (const derived of derivedFieldNamesAtPath(spec, path)) {
     if (!seen.has(derived)) {
       fields.push({ name: derived, type: null, derived: true });
       seen.add(derived);
     }
   }
 
-  return { name, columnTypes, columnStats, fields };
+  return {
+    name: source.name,
+    columnTypes: source.columnTypes,
+    columnStats: source.columnStats,
+    fields,
+  };
 }
 
-// Keyed by draft text alone: this assumes a bound dataset's profile is stable
-// for a given draft. A re-import that changes columns under unchanged text serves
-// stale hints until the next keystroke — harmless, since hints are additive.
-let cache: { text: string; info: DataInfo } | null = null;
-
-/** The active draft's data context, memoized by draft text. */
-export function dataInfo(): DataInfo {
-  const text = useSnippetStore.getState().draftText;
-  if (text.trim() === '') return EMPTY;
-  if (cache && cache.text === text) return cache.info;
-  const info = compute(text);
-  cache = { text, info };
-  return info;
+/** Source columns + derived fields available at `offset` (for completion). */
+export function availableFieldsAt(text: string, offset: number): ReadonlyArray<FieldHint> {
+  return dataInfoAt(text, offset).fields;
 }
 
-/** Source columns + inferred types available to the draft (for facet/repeat defaults). */
-export function boundColumns(): ReadonlyArray<{ name: string; type: ColumnType }> {
-  return dataInfo().columnTypes;
+/** Source columns + inferred types in scope at `offset` (for facet/repeat defaults). */
+export function boundColumnsAt(
+  text: string,
+  offset: number,
+): ReadonlyArray<{ name: string; type: ColumnType }> {
+  return sourceAt(text, offset).columnTypes;
 }
 
-/** Source columns plus the spec's derived fields (for completion/inlay). */
-export function availableFields(): ReadonlyArray<FieldHint> {
-  return dataInfo().fields;
+/** The inferred type of a source field `name` at `offset`, or null (for inlay). */
+export function fieldTypeAt(text: string, offset: number, name: string): ColumnType | null {
+  return sourceAt(text, offset).columnTypes.find((c) => c.name === name)?.type ?? null;
 }

src/app/services/spec-dataset-hints.ts

@@ -27,7 +27,13 @@ import { validateExpression } from '@core/expr-validate';
 import { stringValueAtOffset, valueKeyAtOffset } from '@core/spec-cursor';
 import type { ColumnType } from '@core/type-inference';
 import { useSnippetStore } from '../stores/SnippetStore';
-import { availableFields, dataInfo, type DataInfo, type FieldHint } from './active-dataset';
+import {
+  availableFieldsAt,
+  dataInfoAt,
+  fieldTypeAt,
+  type DataInfo,
+  type FieldHint,
+} from './active-dataset';
 
 /** Property values that reference a data field (where column names belong). */
 const FIELD_KEYS = new Set(['field', 'groupby']);
@@ -65,9 +71,11 @@ export function configureSpecDatasetHints(): void {
     provideCompletionItems(model, position) {
       // Field suggestions only edit the draft; nothing to offer on the published view.
       if (useSnippetStore.getState().editorView !== 'draft') return { suggestions: [] };
-      const key = valueKeyAtOffset(model.getValue(), model.getOffsetAt(position));
+      const text = model.getValue();
+      const offset = model.getOffsetAt(position);
+      const key = valueKeyAtOffset(text, offset);
       if (key === null || !FIELD_KEYS.has(key)) return { suggestions: [] };
-      const fields = availableFields();
+      const fields = availableFieldsAt(text, offset);
       if (fields.length === 0) return { suggestions: [] };
 
       const word = model.getWordUntilPosition(position);
@@ -92,7 +100,7 @@ export function configureSpecDatasetHints(): void {
   });
 
   // All three providers annotate the draft only: they derive their field set from
-  // the draft buffer (dataInfo), so gating to the draft view keeps the hints
+  // the draft buffer (dataInfoAt), so gating to the draft view keeps the hints
   // consistent with the text they are computed from. The published view is a
   // read-only reference, where field hints are marginal.
   monaco.languages.registerHoverProvider('json', {
@@ -119,10 +127,10 @@ export function configureSpecDatasetHints(): void {
         }
       }
 
-      // Over a field name: its type + stats.
+      // Over a field name: its type + stats, resolved at this view's data binding.
       const word = model.getWordAtPosition(position);
       if (word) {
-        const info = dataInfo();
+        const info = dataInfoAt(text, offset);
         const hint = info.fields.find((f) => f.name === word.word);
         if (hint) {
           return {
@@ -143,15 +151,19 @@ export function configureSpecDatasetHints(): void {
   monaco.languages.registerInlayHintsProvider('json', {
     provideInlayHints(model, range) {
       if (useSnippetStore.getState().editorView !== 'draft') return { hints: [], dispose() {} };
-      const types = new Map(availableFields().map((f) => [f.name, f.type]));
-      if (types.size === 0) return { hints: [], dispose() {} };
+      const text = model.getValue();
       const hints: monaco.languages.InlayHint[] = [];
       for (let line = range.startLineNumber; line <= range.endLineNumber; line++) {
         // First `field` per line; the compact format keeps each encoding channel
         // (and its lone field) on its own line, so one match per line suffices.
         const match = /"field"\s*:\s*"([^"]+)"/.exec(model.getLineContent(line));
         if (!match) continue;
-        const type = types.get(match[1]);
+        // Resolve the type at this field's own data binding — views in a
+        // composition can bind different datasets. The offset points inside the
+        // field's value so the path resolves to its enclosing view.
+        const valueCol = match.index + match[0].length - match[1].length;
+        const offset = model.getOffsetAt({ lineNumber: line, column: valueCol });
+        const type = fieldTypeAt(text, offset, match[1]);
         if (!type) continue; // unknown or derived (no type to annotate)
         hints.push({
           position: { lineNumber: line, column: match.index + match[0].length + 1 },

src/app/services/spec-transform-actions.ts

@@ -43,7 +43,7 @@ import {
 } from '@core/spec-transforms';
 import { notify } from '../stores/NotificationStore';
 import { useSnippetStore } from '../stores/SnippetStore';
-import { boundColumns } from './active-dataset';
+import { boundColumnsAt } from './active-dataset';
 import { parseSpecObject } from './spec-config-actions';
 
 /** The composition operators the wrap actions offer. */
@@ -53,6 +53,8 @@ type WrapKind = 'layer' | 'hconcat' | 'vconcat' | 'facet' | 'repeat';
 interface Scope {
   range: monaco.Range;
   text: string;
+  /** Whole-document offset of the focused view, for resolving its data binding. */
+  offset: number;
   /** Column the slice starts at (0-based), so re-indented output stays aligned. */
   baseCol: number;
 }
@@ -63,29 +65,31 @@ const isEmptyRange = (r: monaco.IRange): boolean =>
 /** The whole document, as a scope. */
 function wholeDocument(model: monaco.editor.ITextModel): Scope {
   const range = model.getFullModelRange();
-  return { range, text: model.getValue(), baseCol: 0 };
+  return { range, text: model.getValue(), offset: 0, baseCol: 0 };
 }
 
 /**
  * The slice a transform acts on: an explicit selection if present; else the view
- * the cursor sits in (core/spec-cursor); else the whole document.
+ * the cursor sits in (core/spec-cursor); else the whole document. `offset` is the
+ * cursor/selection-start position in the whole document, used to resolve the
+ * focused view's data binding (facet/repeat field defaults).
  */
 function resolveScope(model: monaco.editor.ITextModel, range: monaco.IRange | null): Scope {
   if (!range) return wholeDocument(model);
-  if (!isEmptyRange(range)) {
-    const r = monaco.Range.lift(range);
-    return { range: r, text: model.getValueInRange(r), baseCol: r.startColumn - 1 };
-  }
   const offset = model.getOffsetAt({
     lineNumber: range.startLineNumber,
     column: range.startColumn,
   });
+  if (!isEmptyRange(range)) {
+    const r = monaco.Range.lift(range);
+    return { range: r, text: model.getValueInRange(r), offset, baseCol: r.startColumn - 1 };
+  }
   const node = findViewRange(model.getValue(), offset);
   if (!node) return wholeDocument(model);
   const start = model.getPositionAt(node.offset);
   const end = model.getPositionAt(node.offset + node.length);
   const r = new monaco.Range(start.lineNumber, start.column, end.lineNumber, end.column);
-  return { range: r, text: model.getValueInRange(r), baseCol: start.column - 1 };
+  return { range: r, text: model.getValueInRange(r), offset, baseCol: start.column - 1 };
 }
 
 /** Indent every line after the first by `baseCol`, so a scoped edit stays aligned. */
@@ -106,8 +110,8 @@ function formatScoped(model: monaco.editor.ITextModel, scope: Scope, next: JsonO
 }
 
 /** A categorical column to facet by (first nominal/ordinal), or a placeholder. */
-function defaultFacet(): { field: string; type: string } {
-  const cols = boundColumns();
+function defaultFacet(text: string, offset: number): { field: string; type: string } {
+  const cols = boundColumnsAt(text, offset);
   const categorical = cols.find((c) => {
     const t = defaultFieldType(c.type);
     return t === 'nominal' || t === 'ordinal';
@@ -119,8 +123,12 @@ function defaultFacet(): { field: string; type: string } {
 }
 
 /** Quantitative columns to repeat over + the channel to rewire, with fallbacks. */
-function defaultRepeat(spec: JsonObject): { fields: string[]; channel: string | null } {
-  const cols = boundColumns();
+function defaultRepeat(
+  spec: JsonObject,
+  text: string,
+  offset: number,
+): { fields: string[]; channel: string | null } {
+  const cols = boundColumnsAt(text, offset);
   const numeric = cols
     .filter((c) => defaultFieldType(c.type) === 'quantitative')
     .map((c) => c.name);
@@ -130,8 +138,12 @@ function defaultRepeat(spec: JsonObject): { fields: string[]; channel: string |
   return { fields: fields.length > 0 ? fields : ['field1', 'field2'], channel };
 }
 
-/** Apply a wrap of the given kind to the parsed spec. */
-function buildNext(spec: JsonObject, kind: WrapKind): JsonObject {
+/**
+ * Apply a wrap of the given kind to the parsed (scoped) spec. `text`/`offset` are
+ * the whole document and the focused view's position, used to resolve that view's
+ * data binding for the facet/repeat field defaults.
+ */
+function buildNext(spec: JsonObject, kind: WrapKind, text: string, offset: number): JsonObject {
   switch (kind) {
     case 'layer':
       return wrapInLayer(spec);
@@ -140,11 +152,11 @@ function buildNext(spec: JsonObject, kind: WrapKind): JsonObject {
     case 'vconcat':
       return wrapInConcat(spec, 'v');
     case 'facet': {
-      const { field, type } = defaultFacet();
+      const { field, type } = defaultFacet(text, offset);
       return wrapInFacet(spec, field, type);
     }
     case 'repeat': {
-      const { fields, channel } = defaultRepeat(spec);
+      const { fields, channel } = defaultRepeat(spec, text, offset);
       return wrapInRepeat(spec, fields, channel);
     }
   }
@@ -190,7 +202,8 @@ export function runWrap(editor: monaco.editor.IStandaloneCodeEditor, kind: WrapK
   const target = resolveTarget(editor);
   if (!target) return;
   const { model, scope, spec } = target;
-  writeBack(editor, scope.range, formatScoped(model, scope, buildNext(spec, kind)));
+  const next = buildNext(spec, kind, model.getValue(), scope.offset);
+  writeBack(editor, scope.range, formatScoped(model, scope, next));
   notify({
     kind: 'success',
     title: 'View wrapped',
@@ -381,13 +394,16 @@ export function configureSpecTransformCodeActions(): void {
         return empty;
       }
       if (!isJsonObject(spec)) return empty;
+      const viewSpec = spec;
 
+      const fullText = model.getValue();
+      const build = (kind: WrapKind) => buildNext(viewSpec, kind, fullText, scope.offset);
       const actions: monaco.languages.CodeAction[] = [
-        editAction(model, scope, 'Wrap view in a layer', buildNext(spec, 'layer')),
-        editAction(model, scope, 'Wrap view in horizontal concat', buildNext(spec, 'hconcat')),
-        editAction(model, scope, 'Wrap view in vertical concat', buildNext(spec, 'vconcat')),
-        editAction(model, scope, 'Wrap view in a facet', buildNext(spec, 'facet')),
-        editAction(model, scope, 'Wrap view in a repeat', buildNext(spec, 'repeat')),
+        editAction(model, scope, 'Wrap view in a layer', build('layer')),
+        editAction(model, scope, 'Wrap view in horizontal concat', build('hconcat')),
+        editAction(model, scope, 'Wrap view in vertical concat', build('vconcat')),
+        editAction(model, scope, 'Wrap view in a facet', build('facet')),
+        editAction(model, scope, 'Wrap view in a repeat', build('repeat')),
       ];
       const collapsed = unwrapSingleton(spec);
       if (collapsed)

src/core/spec-cursor.ts

@@ -70,6 +70,16 @@ export function findViewRange(text: string, offset: number): NodeRange | null {
   return null;
 }
 
+/**
+ * The JSON path (object keys and array indices, root → cursor) of the node at
+ * `offset`. Drives the view-scoped data context: the path is walked to find the
+ * nearest enclosing view's data binding and ancestor transforms
+ * (`core/spec-data`, `core/spec-fields`). Error-tolerant, so it works mid-edit.
+ */
+export function pathAtOffset(text: string, offset: number): (string | number)[] {
+  return [...getLocation(text, offset).path];
+}
+
 /**
  * The property key whose *value* the cursor sits in, or null when the cursor is
  * on a key, at the top level, or otherwise not in a value. For an array element

src/core/spec-data.test.ts

@@ -1,5 +1,5 @@
 import { describe, expect, test } from 'vitest';
-import { classifyData, libraryRefName, selfDefinedNames } from './spec-data';
+import { classifyData, dataBindingAtPath, libraryRefName, selfDefinedNames } from './spec-data';
 
 describe('classifyData', () => {
   test('classifies the four Vega-Lite data shapes', () => {
@@ -61,3 +61,50 @@ describe('libraryRefName', () => {
     expect(libraryRefName({ values: [{ a: 1 }] }, none)).toBeNull();
   });
 });
+
+describe('dataBindingAtPath', () => {
+  test('finds the top-level data when no view overrides it', () => {
+    const spec = { data: { name: 'Sales' }, layer: [{ mark: 'bar', encoding: { x: {} } }] };
+    // Cursor deep inside the first layer's encoding — inherits the top-level data.
+    const binding = dataBindingAtPath(spec, ['layer', 0, 'encoding', 'x']);
+    expect(binding).toEqual({ data: { name: 'Sales' }, anchorPath: [] });
+  });
+
+  test('the nearest enclosing view wins (Vega-Lite data inheritance)', () => {
+    const spec = {
+      data: { name: 'Top' },
+      layer: [{ data: { name: 'Inner' }, encoding: { y: { field: 'a' } } }],
+    };
+    const binding = dataBindingAtPath(spec, ['layer', 0, 'encoding', 'y', 'field']);
+    expect(binding).toEqual({ data: { name: 'Inner' }, anchorPath: ['layer', 0] });
+  });
+
+  test('a sibling view does not leak its data', () => {
+    const spec = {
+      layer: [
+        { data: { name: 'A' }, mark: 'bar' },
+        { data: { name: 'B' }, mark: 'line' },
+      ],
+    };
+    expect(dataBindingAtPath(spec, ['layer', 1, 'mark'])).toEqual({
+      data: { name: 'B' },
+      anchorPath: ['layer', 1],
+    });
+  });
+
+  test('resolves through a facet/repeat child spec', () => {
+    const spec = { facet: { field: 'g' }, spec: { data: { name: 'Child' }, mark: 'point' } };
+    expect(dataBindingAtPath(spec, ['spec', 'mark'])).toEqual({
+      data: { name: 'Child' },
+      anchorPath: ['spec'],
+    });
+  });
+
+  test('returns null when no ancestor on the path declares data', () => {
+    expect(dataBindingAtPath({ mark: 'bar' }, ['mark'])).toBeNull();
+  });
+
+  test('descends defensively past a non-object on the path', () => {
+    expect(dataBindingAtPath({ mark: 'bar' }, ['mark', 'type', 'nope'])).toBeNull();
+  });
+});

src/core/spec-data.ts

@@ -72,3 +72,39 @@ export function libraryRefName(data: unknown, selfDefined: ReadonlySet<string>):
   const { name } = data as { name: string };
   return selfDefined.has(name) ? null : name;
 }
+
+/** A `data` block in scope at a cursor, with the path prefix where it was found. */
+export interface DataBinding {
+  /** The raw `data` value (classify it with `classifyData` / `libraryRefName`). */
+  data: unknown;
+  /** The path prefix of the view that declares this `data` — `[]` for the root. */
+  anchorPath: ReadonlyArray<string | number>;
+}
+
+/**
+ * The data binding in scope at a JSON `path` into the spec — the `data` block of
+ * the *nearest enclosing view* (docs/architecture/08 → editor augmentation). This
+ * implements Vega-Lite's data inheritance: a layer/concat/facet/repeat child
+ * inherits its parent's data unless it declares its own, so the binding is the
+ * `data` of the deepest object along the path that has one. Returns `null` when no
+ * ancestor on the path declares `data`.
+ *
+ * `path` is a jsonc-parser location path (object keys and array indices, root to
+ * cursor). The walk descends it defensively, stopping at any non-object/array.
+ */
+export function dataBindingAtPath(
+  spec: unknown,
+  path: ReadonlyArray<string | number>,
+): DataBinding | null {
+  let node: unknown = spec;
+  let binding: DataBinding | null = null;
+  if (isJsonObject(node) && 'data' in node) binding = { data: node.data, anchorPath: [] };
+  for (let i = 0; i < path.length; i++) {
+    if (node === null || typeof node !== 'object') break;
+    node = (node as Record<string | number, unknown>)[path[i]];
+    if (isJsonObject(node) && 'data' in node) {
+      binding = { data: node.data, anchorPath: path.slice(0, i + 1) };
+    }
+  }
+  return binding;
+}

src/core/spec-fields.test.ts

@@ -1,52 +1,70 @@
 import { describe, expect, it } from 'vitest';
-import { derivedFieldNames } from './spec-fields';
+import { derivedFieldNamesAtPath } from './spec-fields';
+
+describe('derivedFieldNamesAtPath', () => {
+  const spec = {
+    transform: [{ calculate: 'x', as: 'shared' }],
+    layer: [
+      { transform: [{ calculate: 'y', as: 'inner' }], mark: 'bar' },
+      { transform: [{ calculate: 'z', as: 'sibling' }], mark: 'line' },
+    ],
+  };
 
-describe('derivedFieldNames', () => {
   it('collects calculate / timeUnit / bin as-names', () => {
-    const spec = {
+    const s = {
       transform: [
         { calculate: 'datum.a + 1', as: 'plusOne' },
         { timeUnit: 'month', field: 'date', as: 'mo' },
         { bin: true, field: 'x', as: ['x_start', 'x_end'] },
       ],
     };
-    expect(derivedFieldNames(spec).sort()).toEqual(['mo', 'plusOne', 'x_end', 'x_start']);
+    expect(derivedFieldNamesAtPath(s, []).sort()).toEqual(['mo', 'plusOne', 'x_end', 'x_start']);
   });
 
   it('reaches into op-list transforms (aggregate/window/joinaggregate)', () => {
-    const spec = {
+    const s = {
       transform: [
         { aggregate: [{ op: 'mean', field: 'v', as: 'meanV' }], groupby: ['g'] },
         { window: [{ op: 'rank', as: 'rk' }] },
         { joinaggregate: [{ op: 'sum', field: 'v', as: 'total' }] },
       ],
     };
-    expect(derivedFieldNames(spec).sort()).toEqual(['meanV', 'rk', 'total']);
+    expect(derivedFieldNamesAtPath(s, []).sort()).toEqual(['meanV', 'rk', 'total']);
   });
 
   it('defaults fold output to key/value when as is omitted', () => {
-    expect(derivedFieldNames({ transform: [{ fold: ['a', 'b'] }] }).sort()).toEqual([
+    expect(derivedFieldNamesAtPath({ transform: [{ fold: ['a', 'b'] }] }, []).sort()).toEqual([
       'key',
       'value',
     ]);
-    expect(derivedFieldNames({ transform: [{ fold: ['a'], as: ['k', 'v'] }] }).sort()).toEqual([
-      'k',
-      'v',
-    ]);
-  });
-
-  it('recurses into per-view transforms and de-duplicates', () => {
-    const spec = {
-      transform: [{ calculate: 'x', as: 'shared' }],
-      layer: [
-        { transform: [{ calculate: 'y', as: 'inner' }], mark: 'bar' },
-        { transform: [{ calculate: 'z', as: 'shared' }], mark: 'line' },
-      ],
-    };
-    expect(derivedFieldNames(spec).sort()).toEqual(['inner', 'shared']);
+    expect(
+      derivedFieldNamesAtPath({ transform: [{ fold: ['a'], as: ['k', 'v'] }] }, []).sort(),
+    ).toEqual(['k', 'v']);
   });
 
   it('returns nothing for a spec with no transforms', () => {
-    expect(derivedFieldNames({ mark: 'bar', encoding: { x: { field: 'a' } } })).toEqual([]);
+    expect(derivedFieldNamesAtPath({ mark: 'bar', encoding: { x: { field: 'a' } } }, [])).toEqual(
+      [],
+    );
+  });
+
+  it('includes ancestor transforms but not sibling-view transforms', () => {
+    // Inside layer 0: sees its own `inner` and the inherited top-level `shared`,
+    // but never layer 1's `sibling`.
+    expect(derivedFieldNamesAtPath(spec, ['layer', 0, 'mark']).sort()).toEqual(['inner', 'shared']);
+    expect(derivedFieldNamesAtPath(spec, ['layer', 1, 'mark']).sort()).toEqual([
+      'shared',
+      'sibling',
+    ]);
+  });
+
+  it('at the root sees only the top-level transforms', () => {
+    expect(derivedFieldNamesAtPath(spec, [])).toEqual(['shared']);
+  });
+
+  it('descends defensively past a non-object on the path', () => {
+    expect(derivedFieldNamesAtPath(spec, ['transform', 0, 'calculate', 'nope'])).toEqual([
+      'shared',
+    ]);
   });
 });

src/core/spec-fields.ts

@@ -8,8 +8,9 @@
  * offered too. This collects the statically-named ones: every transform `as`
  * (`calculate`, `timeUnit`, `bin`, `stack`, `fold`, `flatten`, `regression`, …)
  * and the nested `as` of the op-list transforms (`aggregate`, `window`,
- * `joinaggregate`). Transforms can sit at the top level or inside any view, so
- * the walk recurses the whole spec.
+ * `joinaggregate`), scoped to the transforms in effect at a cursor — the enclosing
+ * view's and its ancestors' (Vega-Lite applies a parent view's transforms upstream
+ * of its children; a sibling view's do not apply).
  *
  * Out of scope by design: **data-dependent** derived columns — `pivot`'s
  * one-column-per-value output and `lookup`'s imported fields — which only exist
@@ -44,25 +45,36 @@ function collectFromTransform(transform: unknown, names: Set<string>): void {
   }
 }
 
+/** Add every field name a node's own `transform` array introduces. */
+function collectNodeTransforms(node: unknown, names: Set<string>): void {
+  if (
+    node !== null &&
+    typeof node === 'object' &&
+    Array.isArray((node as { transform?: unknown }).transform)
+  ) {
+    for (const t of (node as { transform: unknown[] }).transform) collectFromTransform(t, names);
+  }
+}
+
 /**
- * Every field name the spec derives via its transforms, de-duplicated. Accepts a
- * parsed spec object (callers parse the draft once).
+ * The fields derived by transforms *in scope* at a JSON `path` into the spec — the
+ * `as` outputs of every transform on the cursor's enclosing view and its ancestors
+ * (docs/architecture/08 → editor augmentation). Vega-Lite applies a parent view's
+ * transforms upstream of its children, so a field a parent's `calculate` creates is
+ * visible to a nested view; sibling views' transforms are not. The walk descends
+ * `path` (a jsonc-parser location path) defensively, collecting along the way.
  */
-export function derivedFieldNames(spec: unknown): string[] {
+export function derivedFieldNamesAtPath(
+  spec: unknown,
+  path: ReadonlyArray<string | number>,
+): string[] {
   const names = new Set<string>();
-  const walk = (node: unknown): void => {
-    if (Array.isArray(node)) {
-      for (const item of node) walk(item);
-      return;
-    }
-    if (node !== null && typeof node === 'object') {
-      const obj = node as Record<string, unknown>;
-      if (Array.isArray(obj.transform)) {
-        for (const t of obj.transform) collectFromTransform(t, names);
-      }
-      for (const key of Object.keys(obj)) walk(obj[key]);
-    }
-  };
-  walk(spec);
+  let node: unknown = spec;
+  collectNodeTransforms(node, names);
+  for (const key of path) {
+    if (node === null || typeof node !== 'object') break;
+    node = (node as Record<string | number, unknown>)[key];
+    collectNodeTransforms(node, names);
+  }
   return [...names];
 }

src/core/spec-inline-data.test.ts

@@ -1,34 +1,23 @@
 import { describe, expect, it } from 'vitest';
-import { inlineDataRows } from './spec-inline-data';
+import { rowsForDataBinding } from './spec-inline-data';
 
-describe('inlineDataRows', () => {
-  it('reads top-level data.values', () => {
-    const rows = inlineDataRows({ data: { values: [{ a: 1 }, { a: 2 }] }, mark: 'bar' });
-    expect(rows).toEqual([{ a: 1 }, { a: 2 }]);
+describe('rowsForDataBinding', () => {
+  it('reads inline values from the given binding', () => {
+    const spec = { data: { values: [{ a: 1 }, { a: 2 }] } };
+    expect(rowsForDataBinding(spec, spec.data)).toEqual([{ a: 1 }, { a: 2 }]);
   });
 
-  it('falls back to a nested data.values when there is no top-level one', () => {
-    const spec = { layer: [{ data: { values: [{ b: 1 }] }, mark: 'bar' }] };
-    expect(inlineDataRows(spec)).toEqual([{ b: 1 }]);
+  it('resolves a self-defined named binding against top-level datasets', () => {
+    const spec = { datasets: { ds: [{ c: 1 }, { c: 2 }] }, data: { name: 'ds' } };
+    expect(rowsForDataBinding(spec, { name: 'ds' })).toEqual([{ c: 1 }, { c: 2 }]);
   });
 
-  it('prefers top-level over nested', () => {
-    const spec = {
-      data: { values: [{ top: 1 }] },
-      layer: [{ data: { values: [{ nested: 1 }] } }],
-    };
-    expect(inlineDataRows(spec)).toEqual([{ top: 1 }]);
-  });
-
-  it('reads a top-level datasets entry when no data.values exist', () => {
-    const spec = { datasets: { ds: [{ c: 1 }] }, data: { name: 'ds' } };
-    expect(inlineDataRows(spec)).toEqual([{ c: 1 }]);
-  });
-
-  it('returns null for URL data, empty values, or non-object rows', () => {
-    expect(inlineDataRows({ data: { url: 'x.csv' } })).toBeNull();
-    expect(inlineDataRows({ data: { values: [] } })).toBeNull();
-    expect(inlineDataRows({ data: { values: [1, 2, 3] } })).toBeNull();
-    expect(inlineDataRows({ mark: 'bar' })).toBeNull();
+  it('returns null for a library reference, url, generator, or CSV-string values', () => {
+    const spec = { datasets: { ds: [{ c: 1 }] } };
+    expect(rowsForDataBinding(spec, { name: 'LibraryRef' })).toBeNull(); // not self-defined
+    expect(rowsForDataBinding(spec, { url: 'x.csv' })).toBeNull();
+    expect(rowsForDataBinding(spec, { sequence: { start: 0, stop: 5 } })).toBeNull();
+    expect(rowsForDataBinding(spec, { values: 'a,b\n1,2' })).toBeNull();
+    expect(rowsForDataBinding({ mark: 'bar' }, { values: [] })).toBeNull(); // empty rows
   });
 });

src/core/spec-inline-data.ts

@@ -1,20 +1,18 @@
 /**
  * Inline data carried by a spec (docs/architecture/08 → editor augmentation).
  *
- * Portable core. When a spec has no named library dataset, its columns still
- * exist — inline, in `data.values` or a top-level `datasets` entry. This pulls
- * those rows out so the editor can profile them (core/profile) and offer the same
- * field hints a library-bound spec gets — a "ghost dataset" derived from the spec
- * itself, with nothing stored.
+ * Portable core. When a view binds no named library dataset, its columns still
+ * exist — inline, in `data.values` or a self-defined top-level `datasets` entry.
+ * This pulls those rows out so the editor can profile them (core/profile) and offer
+ * the same field hints a library-bound view gets — a "ghost dataset" derived from
+ * the spec itself, with nothing stored.
  *
- * Resolution order mirrors what the renderer sees first: the nearest `data.values`
- * (top level, then any nested view — pruning the `data` payload from the walk like
- * spec-refs does), then a top-level `datasets` entry. URL data has no rows to read
- * statically, and `values` given as a CSV/TSV string needs format-aware parsing —
- * both out of scope here.
+ * URL data has no rows to read statically, and `values` given as a CSV/TSV string
+ * needs format-aware parsing — both out of scope here (they yield null).
  */
 
 import { isJsonObject } from './spec-config';
+import { classifyData, selfDefinedNames } from './spec-data';
 
 /** An array of row objects, or null when the value is not tabular inline data. */
 function asRows(value: unknown): Record<string, unknown>[] | null {
@@ -24,47 +22,23 @@ function asRows(value: unknown): Record<string, unknown>[] | null {
   return null;
 }
 
-/** Rows from a `data` object's `values`, or null. */
-function rowsFromData(data: unknown): Record<string, unknown>[] | null {
-  return isJsonObject(data) ? asRows(data.values) : null;
-}
-
-/** The first `data.values` reachable from `node`, top level before nested. */
-function firstDataValues(node: unknown): Record<string, unknown>[] | null {
-  if (Array.isArray(node)) {
-    for (const item of node) {
-      const rows = firstDataValues(item);
-      if (rows) return rows;
-    }
-    return null;
-  }
-  if (isJsonObject(node)) {
-    const here = rowsFromData(node.data);
-    if (here) return here;
-    for (const key of Object.keys(node)) {
-      if (key === 'data') continue; // its `values` are payload, already taken above
-      const rows = firstDataValues(node[key]);
-      if (rows) return rows;
-    }
-  }
-  return null;
-}
-
-/** The first non-empty table among the spec's top-level `datasets`, or null. */
-function firstNamedDataset(spec: Record<string, unknown>): Record<string, unknown>[] | null {
-  if (!isJsonObject(spec.datasets)) return null;
-  for (const value of Object.values(spec.datasets)) {
-    const rows = asRows(value);
-    if (rows) return rows;
-  }
-  return null;
-}
-
 /**
- * The spec's inline data rows — `data.values` (top level or nested), else a
- * top-level `datasets` entry — or null when the spec carries none.
+ * The tabular rows a *specific* `data` binding provides for static profiling, or
+ * null when it has none readable here: inline `data.values` (a JSON array of row
+ * objects), or a self-defined `{ name }` resolved against the spec's top-level
+ * `datasets`. A library `{ name }` reference (resolved from the dataset store), a
+ * `url`, a generator, or a CSV/TSV string `values` payload yields null — those
+ * carry no statically-readable JSON rows. Used by the view-scoped editor hints to
+ * profile the data bound at the cursor (`dataBindingAtPath` → this).
  */
-export function inlineDataRows(spec: unknown): Record<string, unknown>[] | null {
-  if (!isJsonObject(spec)) return null;
-  return firstDataValues(spec) ?? firstNamedDataset(spec);
+export function rowsForDataBinding(spec: unknown, data: unknown): Record<string, unknown>[] | null {
+  const kind = classifyData(data);
+  if (kind === 'inline') return asRows((data as { values: unknown }).values);
+  if (kind === 'named') {
+    const name = (data as { name: string }).name;
+    if (selfDefinedNames(spec).has(name) && isJsonObject(spec) && isJsonObject(spec.datasets)) {
+      return asRows((spec.datasets as Record<string, unknown>)[name]);
+    }
+  }
+  return null;
 }