Snapshot URL datasets locally on add; preview tabular data as a table

docs/IMPLEMENTATION-PLAN.md

@@ -268,7 +268,8 @@ the reference.
 - Snippet `datasetRefs` maintained on publish; library shows dataset icon +
   Linked Datasets; dataset detail shows Linked Snippets (bidirectional name link, §09F).
 - **Extract-to-Dataset** flow from the editor (§03F).
-- URL-sourced datasets fetched at render time.
+- URL datasets fetched once on add and **snapshotted** locally (profiled like inline;
+  render from the snapshot, refreshable on demand — see §05 / §04 rendering contract).
 
 **Tests**
 

docs/architecture/02-persistence.md

@@ -6,19 +6,20 @@ How Astrolabe stores data in the browser, and the rules that keep that storage t
 
 ## 1. The Infrastructure-Adapter Principle
 
-**Rule: nothing outside `src/app/infrastructure/` ever touches `indexedDB`, `localStorage`, `window`, or `location` directly.** Every browser-storage interaction goes through a typed adapter module that exposes plain async functions returning domain objects.
+**Rule: nothing outside `src/app/infrastructure/` ever touches `indexedDB`, `localStorage`, `window`, `location`, or `fetch` directly.** Every browser interaction (storage _and_ network) goes through a typed adapter module that exposes plain async functions returning domain objects.
 
 ```
 src/
 ├── core/                    # portable engine — NO browser APIs, NO React
 ├── app/
-│   ├── stores/              # Zustand stores; calls infrastructure, never IDB
+│   ├── stores/              # Zustand stores; calls infrastructure, never IDB/fetch
-│   ├── services/            # business logic; calls infrastructure, never IDB
+│   ├── services/            # business logic; calls infrastructure, never IDB/fetch
-│   └── infrastructure/      # the ONLY place that imports indexedDB/localStorage
+│   └── infrastructure/      # the ONLY place that imports indexedDB/localStorage/fetch
 │       ├── snippet-store.ts     # IndexedDB: snippets (metadata + drafts)
 │       ├── dataset-store.ts     # IndexedDB: datasets (heavy payloads)
 │       ├── settings-store.ts    # localStorage: UserSettings
-│       └── ux-prefs.ts          # localStorage: app/UI prefs (sort, panel layout)
+│       ├── ux-prefs.ts          # localStorage: app/UI prefs (sort, panel layout)
+│       └── remote-data.ts       # network: fetch a URL dataset's body (the ONLY fetch)
 ```
 
 ### Why this boundary exists
@@ -29,7 +30,15 @@ src/
 - **Failure containment.** Quota errors, corrupt JSON, and missing keys are handled at the boundary and converted into typed results (or sane fallbacks), so the rest of the app never sees a raw `DOMException`.
 
 > **Do:** `import { saveSnippet } from '@/app/infrastructure/snippet-store'`
-> **Don't:** `indexedDB.open(...)` or `localStorage.getItem(...)` anywhere in a component, store, or service.
+> **Don't:** `indexedDB.open(...)`, `localStorage.getItem(...)`, or `fetch(...)` anywhere in a component, store, or service.
+
+### Background vs. interactive adapters
+
+Most adapters are driven by **background subscribers** (arch 01 §5, _Effects_): a store changes, a startup subscriber writes it through to IndexedDB — the store never calls the adapter itself. The **network** adapter is the exception. Fetching a URL dataset is a user-initiated action with its own pending/error UI, so the **component** calls `remote-data.ts` directly (components may call adapters — cf. `navigator.clipboard`) and hands the fetched body to **pure** store actions (`DatasetStore.commitUrlSnapshot` / `refreshDataset`). The store never fetches, so it stays browser-free and unit-testable on already-fetched text.
+
+> **Rule:** keep `fetch` behind `remote-data.ts`; orchestrate the URL-dataset fetch in the _component_ (busy state + the "paste data inline instead" recovery), not the store. Store commit actions only ever receive already-fetched text.
+
+**URL-dataset snapshot lifecycle** (the files a change to it touches): add / Refresh → component fetches (`infrastructure/remote-data.ts`) → `core/dataset.snapshotFromText` shapes the body by sniffed format → `DatasetStore.commitUrlSnapshot` / `refreshDataset` snapshots + profiles it _exactly like inline data_ → render resolves it cached-first in `core/rendering.resolvedData` (a live-URL fallback applies only while a URL dataset is still unfetched). See spec §05 for the behavior.
 
 ---
 
@@ -221,6 +230,8 @@ export async function saveSnippet(s: Snippet): Promise<void> {
 > **Do:** default `version` to the earliest shape (`1`) when the field is absent.
 > **Don't:** branch on the presence of individual fields scattered through the app to detect "old data." Centralize that knowledge in the migration function.
 
+> **Mirror shape changes in the import normalizer.** Imported records are built from a file, not read from IndexedDB, so they **never pass through `migrate<Entity>`** — `core/import-normalize.ts` upgrades them independently. A migration that changes a field's _shape_ must be applied in both places or import produces a malformed record. (E.g. the dataset v1→v2 URL-snapshot reshaping — address moves from `data` into `url`, `data` cleared — lives in **both** `migrateDataset` and `normalizeDataset`.)
+
 ---
 
 ## 5. localStorage Preferences (Settings & App/UI Prefs)

docs/architecture/06-type-inference.md

@@ -183,7 +183,10 @@ how the UI shows **"N/A"** — see §3.2.
 
 ### 3.1 What gets profiled
 
-Profiling applies only to **tabular inline data**:
+Profiling applies to any **tabular payload**, whether pasted inline or fetched
+from a URL (the snapshot model stores a URL dataset's data locally, so it profiles
+through the same path as inline data — `snapshotFromText` shapes the fetched body,
+then `computeDatasetProfile` runs):
 
 - **JSON** that is an array of objects.
 - **CSV** (comma-separated, header row).
@@ -191,8 +194,9 @@ Profiling applies only to **tabular inline data**:
 
 Everything else is **not profiled**:
 
-- **URL datasets** — the library holds only the link, not the data, so there is
+- **Unfetched URL datasets** — a URL reference with no snapshot yet (e.g. one
-  nothing to scan. Counts are `null` / N/A.
+  migrated from an older record), so there is nothing to scan. Counts are `null` /
+  N/A until it is refreshed.
 - **Non-tabular data** — a single JSON object, TopoJSON, or anything we can't
   read as rows-of-columns. Counts are `null` / N/A.
 

docs/manual-verification.md

@@ -44,6 +44,23 @@
 - [ ] Toasts stack, auto-dismiss, and fade as specified; the live-preview busy
       indicator appears for slow (>~1s) renders and clears after.
 
+## URL datasets (remote data snapshot)
+
+> Needs a real network, real CORS, and real offline — tests mock the fetch.
+
+- [ ] Add a dataset by URL from a CORS-friendly host (e.g. a GitHub raw `.csv` or a
+      vega-datasets URL) — Save shows "Fetching…", then the dataset appears profiled
+      (rows, columns, size) with a "Fetched <time>" line and the source address.
+- [ ] Go offline (DevTools → Network → Offline) and reload — a snippet that references
+      that URL dataset still renders from the local snapshot (no fetch at render time).
+- [ ] Add a URL that blocks cross-origin requests (or while offline) — the form shows a
+      cause-specific error and a "Paste data inline instead" button that switches to an
+      inline paste keeping the name/comment; no broken record is saved.
+- [ ] Refresh a URL dataset whose source changed — rows/size and the "Fetched" time
+      update; refreshing while offline raises an error toast and leaves the snapshot intact.
+- [ ] Build Chart from a fetched URL dataset works (columns are known); from an
+      unfetched URL reference the builder has no schema until Refresh.
+
 ## Visual sweep of the M6 surfaces
 
 - [ ] Library search / sort / empty states, the storage monitor, About & Donate

docs/spec/04-live-preview.md

@@ -16,8 +16,8 @@ When a spec uses inline data, the preview renders it directly. When a spec inste
 
 - A spec may point at a dataset from the library by name instead of embedding the data inline.
 - Before rendering, the preview substitutes the referenced dataset's stored contents into the spec.
-- URL-sourced datasets are fetched as needed at render time.
+- URL-sourced datasets render from their **local snapshot** (fetched when the dataset was created or last refreshed; see _Datasets_), so the preview does not fetch at render time and works offline. A URL dataset that has never been fetched has no snapshot and instead renders against its **live URL** as a fallback, until it is refreshed.
-- If a referenced dataset cannot be found or fetched, the preview shows a readable error (see _Error Display_) rather than a broken chart.
+- If a referenced dataset cannot be found, the preview shows a readable error (see _Error Display_) rather than a broken chart.
 
 ## Fit / Sizing Modes
 
@@ -42,12 +42,13 @@ Before the chart is drawn, the spec shown in the editor is transformed into the
 
 **1. Dataset reference resolution.** Any named-data reference (`data` with a `name`) is replaced in-place with the referenced dataset's actual contents, shaped by the dataset's source and format (see _Datasets_):
 
-| Dataset source / format | The reference's `data` becomes                                   |
+| Dataset source / format   | The reference's `data` becomes                                                   |
-| ----------------------- | ---------------------------------------------------------------- |
+| ------------------------- | -------------------------------------------------------------------------------- |
-| URL (any format)        | a URL reference to the dataset's address, tagged with its format |
+| URL, fetched (any format) | its snapshot, inlined and tagged exactly like the inline rows below              |
-| Inline JSON             | the parsed values, inlined                                       |
+| URL, not yet fetched      | a live URL reference to the dataset's address, tagged with its format (fallback) |
-| Inline CSV / TSV        | the raw text, inlined, tagged with its format (CSV or TSV)       |
+| Inline JSON               | the parsed values, inlined                                                       |
-| Inline TopoJSON         | the value, inlined, tagged as TopoJSON                           |
+| Inline CSV / TSV          | the raw text, inlined, tagged with its format (CSV or TSV)                       |
+| Inline TopoJSON           | the value, inlined, tagged as TopoJSON                                           |
 
 - Resolution recurses into nested sub-specs (layered and concatenated specs, and a parent spec's child `spec`), so references anywhere in the spec are resolved.
 - If a referenced dataset does not exist, rendering fails with a "dataset not found" error (see _Error Display_).

docs/spec/05-datasets.md

@@ -28,7 +28,7 @@ A two-pane modal:
 Each list item shows:
 
 - The dataset **name**.
-- A **meta line** combining: source ("URL" prefix for URL datasets), row count when known, the **format label** (JSON / CSV / TSV / TOPOJSON), and **size** (human-readable, e.g. B / KB / MB). For URL datasets where counts are not yet known, only the source and format label are shown.
+- A **meta line** combining: source ("URL" prefix for URL datasets), row count when known, the **format label** (JSON / CSV / TSV / TOPOJSON), and **size** (human-readable, e.g. B / KB / MB). A URL dataset that has not yet been fetched shows "not fetched" in place of the figures it does not have.
 - A **usage badge** when one or more snippets reference the dataset, indicating how many.
 
 Clicking an item selects it and shows its detail. Per-item actions (delete, plus copy-reference and build-chart) live in the detail pane for the selected dataset.
@@ -38,9 +38,9 @@ Clicking an item selects it and shows its detail. Per-item actions (delete, plus
 A dataset has one of two source types, chosen when creating it:
 
 - **Inline** — the data itself is pasted in and stored directly in the library.
-- **URL** — the dataset stores a remote URL (http/https). The data is not copied locally; it is fetched on demand when a referencing spec is rendered (see _Live Preview_).
+- **URL** — the dataset is fetched once from a remote http/https address when it is created, and the fetched data is **snapshotted** into the library, along with the address (kept so the snapshot can be re-fetched). A URL dataset then works offline and renders from its local copy; it is never re-fetched at render time, only on demand (see _Refresh_, below).
 
-For inline datasets the library holds the full data and can profile it. For URL datasets the library holds only the link, so row/column/size figures are typically not computed up front and show as "N/A".
+Either way the library holds the full data and profiles it. A URL dataset additionally records the address it was fetched from and when it was last fetched. Because the fetch happens in the browser, an address the browser cannot reach — offline, or one that blocks cross-origin requests — cannot be snapshotted; see _Actions → New / Create New_ for how that failure is handled.
 
 ## Supported Formats
 
@@ -59,18 +59,18 @@ When the user pastes inline data, the app auto-detects the format and reports a
 - Otherwise, multi-line text with a header row is detected as TSV (when tab-separated) or CSV (when comma-separated) — medium confidence.
 - Unrecognized input yields no format (low confidence); saving is blocked with a message asking the user to check the input.
 
-The detected format and source are shown as badges in the create form so the user can confirm or override the source (Inline/URL) before saving. For URL datasets the format is inferred from the URL's file extension (`.csv`, `.tsv`, `.json`, `.topojson`) and shown as a hint.
+The detected format and source are shown as badges in the create form so the user can confirm or override the source (Inline/URL) before saving. For URL datasets the create form shows a format hint inferred from the URL's file extension (`.csv`, `.tsv`, `.json`, `.topojson`); the dataset's actual format is determined from the **fetched content** when it is created, falling back to the extension when the content is ambiguous.
 
 ## Profiling
 
-For tabular inline data (JSON array-of-objects, CSV, TSV) the app computes and stores a profile:
+For tabular data (JSON array-of-objects, CSV, TSV) — whether pasted inline or fetched from a URL — the app computes and stores a profile:
 
 - **Row count** and **column count**.
 - The list of **column names**.
 - An **inferred type per column**: number, text/string, date, or boolean. Type inference looks at the column's values: all-numeric becomes number, all `true`/`false` becomes boolean, otherwise string; empty cells are ignored.
 - **Size** in bytes of the stored data.
 
-A **truncated data preview** of the raw data is also retained for display. URL datasets and non-tabular data are not profiled (counts show "N/A").
+A **truncated data preview** of the raw data is also retained for display. A fetched URL dataset is profiled exactly like inline data; only non-tabular data and a URL dataset that has **not yet been fetched** are unprofiled (counts show "N/A").
 
 ## Detail Panel
 
@@ -78,8 +78,9 @@ The detail pane for a selected dataset shows:
 
 - **Name**.
 - **Comment** (optional free-text notes), when present.
+- **Source** (URL datasets only): the address the snapshot was fetched from, and when it was last fetched (or "Not fetched yet"), alongside the **Refresh** action.
 - **Overview**: statistics (rows, columns, size), the **column list** with each column's name and inferred type shown with a simple type indicator, and created/modified timestamps.
-- **Preview**: a truncated rendering of the data (raw text for CSV/TSV/URL, pretty-printed for JSON/TopoJSON).
+- **Preview**: a sample of the data. Tabular datasets (CSV, TSV, or a JSON array-of-objects — inline or fetched) render as a **table** of the first rows under the profiled column names, with a note when more rows exist than are shown. Non-tabular payloads (a single JSON object, TopoJSON) render as pretty-printed JSON, and a URL dataset that has not yet been fetched shows a short placeholder.
 - **Linked Snippets**: the list of snippets that reference this dataset by name. This is the dataset side of bidirectional dataset↔snippet linking (see _Snippet Library_).
 
 ## Actions
@@ -89,13 +90,14 @@ A destructive or off-screen outcome raises a confirming toast; an action whose r
 - **Copy Reference** — copies the by-name reference object to the clipboard, ready to paste into a spec:
   `{ "data": { "name": "MyDataset" } }`
   The clipboard write is invisible, so it is confirmed _inline on the control_ ("Copied"), announced politely to assistive technology — not a toast.
-- **New / Create New** — opens the create form in the detail pane with fields: **name** (required, unique), **source** toggle (Inline / URL), the **data** (a paste area for inline, a URL field for URL source), and an optional **comment**. Save is disabled until a name and valid data/URL are present. On success the new dataset is shown selected in the detail pane — that visible result is the confirmation, so no toast is raised (a dataset created _off-screen_ via Extract does toast; see _Spec Editor_).
+- **New / Create New** — opens the create form in the detail pane with fields: **name** (required, unique), **source** toggle (Inline / URL), the **data** (a paste area for inline, a URL field for URL source), and an optional **comment**. Save is disabled until a name and valid data/URL are present. For a **URL** dataset, saving fetches and snapshots the address; the Save control shows a busy state while fetching. If the fetch fails — offline, blocked by the host (cross-origin), not found, empty, or timed out — the form shows a readable, cause-specific error and offers a one-click **"Paste data inline instead"** that switches the form to an inline paste (keeping the name and comment), rather than saving a broken record. On success the new dataset is shown selected in the detail pane — that visible result is the confirmation, so no toast is raised (a dataset created _off-screen_ via Extract does toast; see _Spec Editor_).
-- **Edit** — rename, edit the comment, and update the data (re-paste inline data or refresh the URL). Updating inline data re-profiles it; the modified timestamp advances.
+- **Refresh** (URL datasets) — re-fetches the dataset's address and re-snapshots it, re-profiling the data and advancing the last-fetched time. The updated figures are the visible confirmation, so success raises no toast; a failed refresh raises an error toast. Refresh is also how a URL dataset that has not yet been fetched (e.g. one migrated from an older version) acquires its snapshot.
+- **Edit** — rename, edit the comment, change the source data, and (for URL datasets) change the address. Updating inline data re-profiles it; changing a URL dataset's address re-fetches and re-snapshots it; editing only a URL dataset's name or comment does **not** re-fetch. The modified timestamp advances.
 - **Delete** — asks for confirmation ("Delete \"Name\"? This cannot be undone."), then removes the dataset and clears the selection.
 
 ## Build Chart From Dataset
 
-From a selected dataset the user can launch the visual _Chart Builder_ (see _Chart Builder_) pre-targeted at that dataset, producing a new snippet whose spec references the dataset by name.
+From a selected dataset the user can launch the visual _Chart Builder_ (see _Chart Builder_) pre-targeted at that dataset, producing a new snippet whose spec references the dataset by name. This relies on the dataset's profiled columns, so a URL dataset must have been fetched first — an unfetched URL reference has no schema to map (use _Refresh_ to fetch it).
 
 ## Extract Inline Data → Dataset
 

docs/spec/09-data-model.md

@@ -34,29 +34,33 @@ A snippet carries two specs at once. `draftSpec` is the editable working copy; `
 
 A **Dataset** is a named, reusable data source that snippets can reference by name instead of inlining data. Datasets are managed in the _Datasets_ manager and support multiple formats and two source kinds.
 
-| Field         | Type                      | Meaning                                                                                                                          |
+| Field         | Type                      | Meaning                                                                                                                                                                               |
-| ------------- | ------------------------- | -------------------------------------------------------------------------------------------------------------------------------- |
+| ------------- | ------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
-| `id`          | number                    | Unique numeric identifier.                                                                                                       |
+| `id`          | number                    | Unique numeric identifier.                                                                                                                                                            |
-| `version`     | number                    | Schema version of this record, used for read-time migration (see _Schema versioning_ below).                                     |
+| `version`     | number                    | Schema version of this record, used for read-time migration (see _Schema versioning_ below).                                                                                          |
-| `name`        | string                    | Unique, human-readable name; the key snippets reference via `datasetRefs`.                                                       |
+| `name`        | string                    | Unique, human-readable name; the key snippets reference via `datasetRefs`.                                                                                                            |
-| `data`        | JSON value                | The payload. For `source = url`: the URL string. For `source = inline`: the raw CSV/TSV text, or the parsed JSON/TopoJSON value. |
+| `data`        | JSON value                | The payload, shaped by format: raw CSV/TSV text, or the parsed JSON/TopoJSON value. For `source = url` this is the **fetched snapshot**, or `null` before the first successful fetch. |
-| `format`      | string                    | One of `json`, `csv`, `tsv`, `topojson`.                                                                                         |
+| `format`      | string                    | One of `json`, `csv`, `tsv`, `topojson`.                                                                                                                                              |
-| `source`      | string                    | One of `inline` (data embedded in the record) or `url` (data fetched from a remote address).                                     |
+| `source`      | string                    | One of `inline` (data pasted into the record) or `url` (data fetched once from a remote address and snapshotted into the record).                                                     |
-| `comment`     | string                    | Free-form user note about the dataset.                                                                                           |
+| `url`         | string (url only)         | The remote address a `url` dataset was fetched from, retained so it can be re-fetched ("Refresh"). Absent for inline datasets.                                                        |
-| `rowCount`    | number or null            | Number of data rows, or null when unknown/not applicable.                                                                        |
+| `fetchedAt`   | ISO-timestamp or null     | For `url` datasets: when the snapshot was last fetched, or `null` if never fetched. Absent for inline datasets.                                                                       |
-| `columnCount` | number or null            | Number of columns, or null when unknown/not applicable.                                                                          |
+| `comment`     | string                    | Free-form user note about the dataset.                                                                                                                                                |
-| `columns`     | string[]                  | Column names, in order.                                                                                                          |
+| `rowCount`    | number or null            | Number of data rows, or null when unknown/not applicable.                                                                                                                             |
-| `columnTypes` | array of `{ name, type }` | Per-column inferred type. `name` is the column; `type` is one of `number`, `string`, `date`, `boolean`.                          |
+| `columnCount` | number or null            | Number of columns, or null when unknown/not applicable.                                                                                                                               |
-| `size`        | number                    | Approximate payload size in bytes.                                                                                               |
+| `columns`     | string[]                  | Column names, in order.                                                                                                                                                               |
-| `created`     | ISO-timestamp string      | When the dataset was first added.                                                                                                |
+| `columnTypes` | array of `{ name, type }` | Per-column inferred type. `name` is the column; `type` is one of `number`, `string`, `date`, `boolean`.                                                                               |
-| `modified`    | ISO-timestamp string      | When the dataset was last changed.                                                                                               |
+| `size`        | number                    | Approximate payload size in bytes.                                                                                                                                                    |
+| `created`     | ISO-timestamp string      | When the dataset was first added.                                                                                                                                                     |
+| `modified`    | ISO-timestamp string      | When the dataset was last changed.                                                                                                                                                    |
 
-The `rowCount`, `columnCount`, `columns`, `columnTypes`, and `size` fields are derived summaries computed when data is added or updated; they support previews and type display without re-parsing the full payload.
+The `rowCount`, `columnCount`, `columns`, `columnTypes`, and `size` fields are derived summaries computed when data is added or updated — including when a `url` dataset is fetched or refreshed; a fetched URL snapshot profiles exactly like inline data. They support previews and type display without re-parsing the full payload.
 
 ### Schema versioning
 
 Both **Snippet** and **Dataset** records carry a numeric `version` recording the shape of that individual record. When a record is read from storage it is migrated up to the current shape before the app uses it; new writes always store the current version. A record written before versioning existed (no `version` field) is treated as version `1`. This is distinct from the storage container's own layout version, and from the _Import & Export_ envelope `version` (which describes the file format, not a record). Records exported via _Import & Export_ include their `version`.
 
+The current **Dataset** version is `2`. The v1→v2 migration reflects the URL-snapshot model: a v1 `url` dataset stored its address in `data`, so migration moves that address into the new `url` field and clears `data` to `null` — the record becomes an _unfetched reference_ that renders against its live URL until the user refreshes it, at which point the fetched snapshot is stored and profiled.
+
 ## C. UserSettings
 
 **UserSettings** holds persisted user preferences as a single structured record. The semantics and UX of each option are covered in _Settings_; the shape below is the storage contract.

src/app/components/DatasetsModal.module.css

@@ -235,6 +235,32 @@
   color: var(--text-secondary);
 }
 
+/* URL dataset provenance: the source address + last-fetched time (spec §05). */
+.sourceMeta {
+  display: flex;
+  flex-wrap: wrap;
+  align-items: baseline;
+  gap: var(--space-2) var(--space-4);
+  margin: 0;
+  font-size: 12px;
+  color: var(--text-secondary);
+}
+
+.sourceLink {
+  color: var(--accent);
+  text-decoration: none;
+  overflow-wrap: anywhere;
+}
+
+.sourceLink:hover {
+  text-decoration: underline;
+}
+
+.sourceLink:focus-visible {
+  outline: 2px solid var(--focus);
+  outline-offset: 2px;
+}
+
 .section {
   display: flex;
   flex-direction: column;
@@ -335,6 +361,46 @@
   white-space: pre;
 }
 
+/* Tabular preview — a scrollable grid with a sticky header (spec §05 → Preview). */
+.previewTableWrap {
+  max-height: 260px;
+  overflow: auto;
+  border: var(--border-width) solid var(--border);
+  border-radius: var(--radius);
+}
+
+.previewTable {
+  border-collapse: collapse;
+  width: 100%;
+  font-family: var(--font-mono);
+  font-size: 12px;
+  line-height: 1.4;
+}
+
+.previewTable th,
+.previewTable td {
+  text-align: left;
+  padding: var(--space-2) var(--space-3);
+  border-bottom: var(--border-width) solid var(--border);
+  white-space: nowrap;
+  max-width: 240px;
+  overflow: hidden;
+  text-overflow: ellipsis;
+}
+
+.previewTable thead th {
+  position: sticky;
+  top: 0;
+  z-index: 1;
+  background: var(--layer-01);
+  font-weight: 600;
+  color: var(--text-secondary);
+}
+
+.previewTable tbody tr:last-child td {
+  border-bottom: none;
+}
+
 .muted {
   margin: 0;
   font-size: 13px;
@@ -439,3 +505,10 @@
   font-size: 13px;
   color: var(--support-error);
 }
+
+/* The inline-fallback recovery under a failed URL fetch — left-aligned, not
+   stretched, so it reads as a secondary recovery beneath the error message. */
+.fallback {
+  align-self: flex-start;
+  margin-top: calc(-1 * var(--space-2));
+}

src/app/components/DatasetsModal.tsx

@@ -12,12 +12,14 @@
  * Snippets stay reactive without a stored back-pointer.
  */
 
-import { useState } from 'react';
+import { useMemo, useState } from 'react';
 import { useShallow } from 'zustand/react/shallow';
-import { datasetReference, type DataSource, type Dataset } from '@core/dataset';
+import { datasetReference, tabularRows, type DataSource, type Dataset } from '@core/dataset';
 import { detectFormat, detectFormatFromUrl, type DataFormat } from '@core/format-detection';
 import { datasetUsageCounts, snippetsReferencingDataset } from '@core/relationships';
 import { closeModal, openModal, resnapshot } from '../modals/ModalCoordinator';
+import { fetchRemoteData } from '../infrastructure/remote-data';
+import { remoteFetchErrorMessage } from '../services/remote-data-errors';
 import { confirm } from '../stores/ConfirmStore';
 import { notify } from '../stores/NotificationStore';
 import {
@@ -50,6 +52,17 @@ const SOURCE_OPTIONS: ReadonlyArray<SegmentedOption<DataSource>> = [
   { value: 'url', label: 'URL' },
 ];
 
+/** Rows shown in the tabular preview before truncating (spec §05 → Detail Panel). */
+const PREVIEW_ROW_LIMIT = 50;
+
+/** One table cell's text: blank for empty, the string as-is, else JSON (numbers,
+ *  booleans, nested values). Avoids `String()` on objects ("[object Object]"). */
+function cellText(value: unknown): string {
+  if (value == null) return '';
+  if (typeof value === 'string') return value;
+  return JSON.stringify(value);
+}
+
 export function DatasetsModal() {
   const datasets = useDatasetStore(useShallow((s) => s.datasets));
   const view = useDatasetStore((s) => s.view);
@@ -116,11 +129,15 @@ function DatasetListItem({
   onSelect: () => void;
 }) {
   // Meta line: source ("URL" prefix), row count when known, format label, size.
+  // A fetched URL snapshot reads like an inline dataset (rows + size); an unfetched
+  // URL reference shows "not fetched" in place of figures it doesn't have yet.
+  const unfetched = dataset.source === 'url' && dataset.data == null;
   const parts: string[] = [];
   if (dataset.source === 'url') parts.push('URL');
-  if (dataset.rowCount !== null) parts.push(`${dataset.rowCount} rows`);
+  if (unfetched) parts.push('not fetched');
+  else if (dataset.rowCount !== null) parts.push(`${dataset.rowCount} rows`);
   parts.push(formatLabel(dataset.format));
-  if (dataset.source !== 'url') parts.push(humanBytes(dataset.size));
+  if (!unfetched) parts.push(humanBytes(dataset.size));
 
   return (
     <li className={`${styles.item} ${active ? styles.itemActive : ''}`}>
@@ -151,16 +168,46 @@ function DatasetDetail({
 }) {
   const startEdit = useDatasetStore((s) => s.startEdit);
   const remove = useDatasetStore((s) => s.remove);
+  const refreshDataset = useDatasetStore((s) => s.refreshDataset);
   const selectSnippet = useSnippetStore((s) => s.selectSnippet);
   const [copied, setCopied] = useState(false);
+  const [refreshing, setRefreshing] = useState(false);
 
   const linked = snippetsReferencingDataset(snippets, dataset.name);
 
+  // Tabular data (CSV/TSV/JSON-array, inline or fetched) previews as a table of the
+  // first rows under the profiled columns; non-tabular payloads fall back to text.
+  // Memoized on the record so a large CSV isn't re-parsed on unrelated re-renders.
+  const previewRows = useMemo(
+    () => tabularRows(dataset.data, dataset.format, PREVIEW_ROW_LIMIT),
+    [dataset],
+  );
+
   const handleEdit = () => {
     startEdit();
     resnapshot();
   };
 
+  // Re-fetch a URL dataset's source and re-snapshot it. The visible result (updated
+  // rows + "Fetched" time) is the confirmation, so success raises no toast; only a
+  // failure surfaces one (spec §05 → Actions; docs/architecture/10 → Toast copy).
+  const handleRefresh = async () => {
+    if (!dataset.url) return;
+    setRefreshing(true);
+    try {
+      const { text } = await fetchRemoteData(dataset.url);
+      refreshDataset(dataset.id, { text });
+    } catch (err) {
+      notify({
+        kind: 'error',
+        title: "Couldn't refresh dataset",
+        message: remoteFetchErrorMessage(err, 'retry'),
+      });
+    } finally {
+      setRefreshing(false);
+    }
+  };
+
   const handleCopy = async () => {
     const text = JSON.stringify(datasetReference(dataset.name), null, 2);
     try {
@@ -212,6 +259,16 @@ function DatasetDetail({
           <span role="status" className="visually-hidden">
             {copied ? 'Reference copied to clipboard' : ''}
           </span>
+          {dataset.source === 'url' && (
+            <button
+              type="button"
+              className={styles.action}
+              onClick={() => void handleRefresh()}
+              disabled={refreshing}
+            >
+              {refreshing ? 'Refreshing…' : 'Refresh'}
+            </button>
+          )}
           <button type="button" className={styles.action} onClick={handleEdit}>
             Edit
           </button>
@@ -237,6 +294,19 @@ function DatasetDetail({
 
       {dataset.comment && <p className={styles.comment}>{dataset.comment}</p>}
 
+      {dataset.source === 'url' && (
+        <p className={styles.sourceMeta}>
+          <a className={styles.sourceLink} href={dataset.url} target="_blank" rel="noreferrer">
+            {dataset.url}
+          </a>
+          <span>
+            {dataset.fetchedAt
+              ? `Fetched ${new Date(dataset.fetchedAt).toLocaleString()}`
+              : 'Not fetched yet'}
+          </span>
+        </p>
+      )}
+
       <section className={styles.section}>
         <h4 className={styles.sectionTitle}>Overview</h4>
         <dl className={styles.stats}>
@@ -254,7 +324,7 @@ function DatasetDetail({
           </div>
           <div>
             <dt>Size</dt>
-            <dd>{dataset.source === 'url' ? 'N/A' : humanBytes(dataset.size)}</dd>
+            <dd>{dataset.data == null ? 'N/A' : humanBytes(dataset.size)}</dd>
           </div>
         </dl>
 
@@ -277,7 +347,43 @@ function DatasetDetail({
 
       <section className={styles.section}>
         <h4 className={styles.sectionTitle}>Preview</h4>
-        <pre className={styles.preview}>{previewText(dataset)}</pre>
+        {previewRows ? (
+          <>
+            {/* TODO: a11y — this scroll container isn't keyboard-reachable; when the
+                preview overflows, keyboard-only users can't scroll it (WCAG 2.1.1).
+                A council pass should decide the pattern (tabindex=0 + aria-label region,
+                or a different overflow treatment) before we lean on it more. */}
+            <div className={styles.previewTableWrap}>
+              <table className={styles.previewTable}>
+                <thead>
+                  <tr>
+                    {dataset.columns.map((col, ci) => (
+                      <th key={ci} scope="col">
+                        {col}
+                      </th>
+                    ))}
+                  </tr>
+                </thead>
+                <tbody>
+                  {previewRows.map((row, ri) => (
+                    <tr key={ri}>
+                      {dataset.columns.map((col, ci) => (
+                        <td key={ci}>{cellText(row[col])}</td>
+                      ))}
+                    </tr>
+                  ))}
+                </tbody>
+              </table>
+            </div>
+            {dataset.rowCount != null && dataset.rowCount > previewRows.length && (
+              <p className={styles.muted}>
+                Showing the first {previewRows.length} of {dataset.rowCount} rows.
+              </p>
+            )}
+          </>
+        ) : (
+          <pre className={styles.preview}>{previewText(dataset)}</pre>
+        )}
       </section>
 
       <section className={styles.section}>
@@ -307,19 +413,22 @@ function DatasetDetail({
   );
 }
 
-/** A truncated rendering of the data: raw for csv/tsv/url, pretty JSON otherwise. */
+/** A truncated rendering of the data: raw for csv/tsv, pretty JSON otherwise. */
 function previewText(dataset: Dataset): string {
   const MAX = 2000;
+  // No snapshot yet (an unfetched URL reference) — there is nothing to preview.
+  if (dataset.data == null) {
+    return dataset.source === 'url' ? 'Not fetched yet — use Refresh to load the data.' : '';
+  }
   let text: string;
-  if (dataset.source === 'url') {
+  if (dataset.format === 'csv' || dataset.format === 'tsv') {
-    text = String(dataset.data);
+    // CSV/TSV payloads are raw text; fall back to JSON for any non-string value.
-  } else if (dataset.format === 'csv' || dataset.format === 'tsv') {
+    text = typeof dataset.data === 'string' ? dataset.data : JSON.stringify(dataset.data);
-    text = String(dataset.data);
   } else {
     try {
       text = JSON.stringify(dataset.data, null, 2);
     } catch {
-      text = String(dataset.data);
+      text = typeof dataset.data === 'string' ? dataset.data : '[unserializable data]';
     }
   }
   return text.length > MAX ? `${text.slice(0, MAX)}\n…` : text;
@@ -328,20 +437,59 @@ function previewText(dataset: Dataset): string {
 function DatasetFormView({ editing }: { editing: boolean }) {
   const form = useDatasetStore((s) => s.form);
   const formError = useDatasetStore((s) => s.formError);
+  const selected = useDatasetStore(selectSelectedDataset);
   const updateForm = useDatasetStore((s) => s.updateForm);
   const cancelForm = useDatasetStore((s) => s.cancelForm);
   const save = useDatasetStore((s) => s.save);
+  const commitUrlSnapshot = useDatasetStore((s) => s.commitUrlSnapshot);
   // Save stays disabled until a name and valid data/URL are present (spec §05).
   const canSave = useDatasetStore(selectCanSave);
 
+  // URL datasets fetch on save (snapshot model): `fetching` drives the button's
+  // busy state; `fetchError` holds a failed fetch's message + the inline fallback.
+  const [fetching, setFetching] = useState(false);
+  const [fetchError, setFetchError] = useState<string | null>(null);
+
   // Live format/source hint from the current input (spec §05 → Auto-detection).
   const detected =
     form.source === 'url'
       ? { format: detectFormatFromUrl(form.input.trim()), confidence: 'url' as const }
       : detectFormat(form.input);
 
-  const handleSave = () => {
+  const handleSave = async () => {
-    if (save()) resnapshot(); // committed — re-baseline so a later close won't prompt
+    // Inline saves are synchronous; only URL datasets touch the network.
+    if (form.source !== 'url') {
+      if (save()) resnapshot(); // committed — re-baseline so a later close won't prompt
+      return;
+    }
+    const url = form.input.trim();
+    // A metadata-only edit (same URL, snapshot already present) needs no re-fetch.
+    const metaOnly =
+      editing &&
+      selected?.source === 'url' &&
+      selected.data != null &&
+      url === (selected.url ?? '');
+    if (metaOnly) {
+      if (save()) resnapshot();
+      return;
+    }
+    // Create, changed URL, or inline→URL: fetch once, then snapshot + commit.
+    setFetchError(null);
+    setFetching(true);
+    try {
+      const { text } = await fetchRemoteData(url);
+      if (commitUrlSnapshot({ text })) resnapshot();
+    } catch (err) {
+      setFetchError(remoteFetchErrorMessage(err));
+    } finally {
+      setFetching(false);
+    }
+  };
+
+  // Recovery from a failed fetch: switch to an inline paste, keeping name + comment.
+  const handlePasteInline = () => {
+    setFetchError(null);
+    updateForm({ source: 'inline', input: '' });
   };
 
   const handleCancel = () => {
@@ -370,7 +518,10 @@ function DatasetFormView({ editing }: { editing: boolean }) {
           label="Dataset source"
           options={SOURCE_OPTIONS}
           value={form.source}
-          onChange={(source) => updateForm({ source })}
+          onChange={(source) => {
+            setFetchError(null);
+            updateForm({ source });
+          }}
         />
       </div>
 
@@ -381,7 +532,10 @@ function DatasetFormView({ editing }: { editing: boolean }) {
             type="url"
             className={styles.input}
             value={form.input}
-            onChange={(e) => updateForm({ input: e.target.value })}
+            onChange={(e) => {
+              setFetchError(null);
+              updateForm({ input: e.target.value });
+            }}
             placeholder="https://example.com/data.csv"
           />
         ) : (
@@ -418,23 +572,34 @@ function DatasetFormView({ editing }: { editing: boolean }) {
         />
       </label>
 
-      {formError && (
+      {(formError || fetchError) && (
         <p className={styles.formError} role="alert">
-          {formError}
+          {formError ?? fetchError}
         </p>
       )}
+      {/* A failed fetch is recoverable by pasting the data inline (the user's choice
+          when CORS or offline blocks the URL) — offered as a direct one-click path. */}
+      {fetchError && (
+        <button
+          type="button"
+          className={`${styles.action} ${styles.fallback}`}
+          onClick={handlePasteInline}
+        >
+          Paste data inline instead
+        </button>
+      )}
 
       <div className={styles.formActions}>
-        <button type="button" className={styles.action} onClick={handleCancel}>
+        <button type="button" className={styles.action} onClick={handleCancel} disabled={fetching}>
           Cancel
         </button>
         <button
           type="button"
           className={`${styles.action} ${styles.primary}`}
-          disabled={!canSave}
+          disabled={!canSave || fetching}
-          onClick={handleSave}
+          onClick={() => void handleSave()}
         >
-          {editing ? 'Save changes' : 'Create dataset'}
+          {fetching ? 'Fetching…' : editing ? 'Save changes' : 'Create dataset'}
         </button>
       </div>
     </div>

src/app/infrastructure/dataset-migrations.test.ts

@@ -0,0 +1,73 @@
+import { describe, expect, test } from 'vitest';
+import { CURRENT_DATASET_VERSION } from '@core/dataset';
+import { migrateDataset } from './dataset-migrations';
+
+describe('migrateDataset', () => {
+  test('stamps the current version and fills missing fields with defaults', () => {
+    const d = migrateDataset({ id: 7, data: 'a,b\n1,2', format: 'csv', source: 'inline' });
+
+    expect(d.id).toBe(7);
+    expect(d.version).toBe(CURRENT_DATASET_VERSION);
+    expect(d.name).toBe('Untitled');
+    expect(d.comment).toBe('');
+    expect(d.columns).toEqual([]);
+    expect(d.columnStats).toEqual([]);
+  });
+
+  test('an inline record carries no url/fetchedAt keys', () => {
+    const d = migrateDataset({ id: 1, data: [{ a: 1 }], format: 'json', source: 'inline' });
+    expect(d.source).toBe('inline');
+    expect('url' in d).toBe(false);
+    expect('fetchedAt' in d).toBe(false);
+  });
+
+  test('v1→v2: a legacy URL record moves its address out of data into url', () => {
+    // Pre-v2 shape: the URL lived in `data`, there was no `url`/`fetchedAt`, and the
+    // profile was N/A. It becomes an unfetched reference: url set, data cleared.
+    const d = migrateDataset({
+      id: 2,
+      data: 'https://example.com/data.csv',
+      format: 'csv',
+      source: 'url',
+      rowCount: null,
+      columns: [],
+    });
+
+    expect(d.version).toBe(CURRENT_DATASET_VERSION);
+    expect(d.source).toBe('url');
+    expect(d.url).toBe('https://example.com/data.csv');
+    expect(d.data).toBeNull();
+    expect(d.fetchedAt).toBeNull();
+    expect(d.rowCount).toBeNull();
+  });
+
+  test('a v2 URL snapshot is preserved (data, url, and fetchedAt all kept)', () => {
+    const d = migrateDataset({
+      id: 3,
+      version: 2,
+      data: 'a,b\n1,2',
+      url: 'https://example.com/data.csv',
+      fetchedAt: '2026-06-10T00:00:00.000Z',
+      format: 'csv',
+      source: 'url',
+      rowCount: 1,
+      columns: ['a', 'b'],
+    });
+
+    expect(d.data).toBe('a,b\n1,2');
+    expect(d.url).toBe('https://example.com/data.csv');
+    expect(d.fetchedAt).toBe('2026-06-10T00:00:00.000Z');
+    expect(d.rowCount).toBe(1);
+  });
+
+  test('preserves unknown fields so a newer build round-trips without loss', () => {
+    const d = migrateDataset({
+      id: 1,
+      data: [],
+      format: 'json',
+      source: 'inline',
+      futureField: 42,
+    });
+    expect((d as unknown as Record<string, unknown>).futureField).toBe(42);
+  });
+});

src/app/infrastructure/dataset-migrations.ts

@@ -30,14 +30,33 @@ function asCount(v: unknown): number | null {
 /** Upgrade a raw stored record to the current Dataset shape. */
 export function migrateDataset(raw: unknown): Dataset {
   const r = { ...(raw as Record<string, unknown>) };
+  const source = asSource(r.source);
+
+  // v1→v2: a URL dataset used to store its address in `data`; it now stores the
+  // fetched snapshot in `data` and keeps the address in `url` (+ a `fetchedAt`).
+  // A legacy record (url source, no `url` field) becomes an *unfetched* reference:
+  // move the address to `url` and clear `data`. The N/A profile carries over as-is;
+  // rendering falls back to the live URL until a Refresh fetches + profiles it.
+  const legacyUrl = source === 'url' && typeof r.url !== 'string';
+  const url = legacyUrl
+    ? typeof r.data === 'string'
+      ? r.data
+      : undefined
+    : typeof r.url === 'string'
+      ? r.url
+      : undefined;
+  const fetchedAt = legacyUrl ? null : typeof r.fetchedAt === 'string' ? r.fetchedAt : null;
+
   return {
     ...r,
     id: typeof r.id === 'number' ? r.id : Number(r.id),
     version: CURRENT_DATASET_VERSION,
     name: typeof r.name === 'string' ? r.name : 'Untitled',
-    data: r.data,
+    data: legacyUrl ? null : r.data,
     format: asFormat(r.format),
-    source: asSource(r.source),
+    source,
+    // Only URL datasets carry url/fetchedAt; inline records stay clean.
+    ...(source === 'url' ? { url, fetchedAt } : {}),
     comment: typeof r.comment === 'string' ? r.comment : '',
     rowCount: asCount(r.rowCount),
     columnCount: asCount(r.columnCount),

src/app/infrastructure/remote-data.test.ts

@@ -0,0 +1,94 @@
+import { afterEach, describe, expect, test, vi } from 'vitest';
+import { fetchRemoteData, RemoteFetchError } from './remote-data';
+
+/** A minimal Response stand-in for the bits the adapter reads. */
+function fakeResponse(opts: {
+  ok: boolean;
+  status?: number;
+  statusText?: string;
+  body?: string;
+  contentType?: string | null;
+}): Response {
+  return {
+    ok: opts.ok,
+    status: opts.status ?? (opts.ok ? 200 : 500),
+    statusText: opts.statusText ?? '',
+    text: () => Promise.resolve(opts.body ?? ''),
+    headers: {
+      get: (k: string) => (k.toLowerCase() === 'content-type' ? (opts.contentType ?? null) : null),
+    },
+  } as unknown as Response;
+}
+
+afterEach(() => vi.unstubAllGlobals());
+
+describe('fetchRemoteData', () => {
+  test('returns the body and content-type on success', async () => {
+    vi.stubGlobal(
+      'fetch',
+      vi.fn(() =>
+        Promise.resolve(fakeResponse({ ok: true, body: 'a,b\n1,2', contentType: 'text/csv' })),
+      ),
+    );
+    const result = await fetchRemoteData('https://example.com/data.csv');
+    expect(result.text).toBe('a,b\n1,2');
+    expect(result.contentType).toBe('text/csv');
+  });
+
+  test('classifies a non-2xx response as http with its status', async () => {
+    vi.stubGlobal(
+      'fetch',
+      vi.fn(() =>
+        Promise.resolve(fakeResponse({ ok: false, status: 404, statusText: 'Not Found' })),
+      ),
+    );
+    await expect(fetchRemoteData('https://example.com/missing.csv')).rejects.toMatchObject({
+      reason: 'http',
+      status: 404,
+    });
+  });
+
+  test('classifies a rejected fetch (CORS / offline) as network', async () => {
+    vi.stubGlobal(
+      'fetch',
+      vi.fn(() => Promise.reject(new TypeError('Failed to fetch'))),
+    );
+    await expect(fetchRemoteData('https://blocked.example/data.csv')).rejects.toMatchObject({
+      reason: 'network',
+    });
+  });
+
+  test('classifies an empty body as empty', async () => {
+    vi.stubGlobal(
+      'fetch',
+      vi.fn(() => Promise.resolve(fakeResponse({ ok: true, body: '   \n' }))),
+    );
+    await expect(fetchRemoteData('https://example.com/blank.csv')).rejects.toMatchObject({
+      reason: 'empty',
+    });
+  });
+
+  test('classifies an aborted fetch as timeout', async () => {
+    vi.stubGlobal(
+      'fetch',
+      vi.fn(() => {
+        const e = new Error('aborted');
+        e.name = 'AbortError';
+        return Promise.reject(e);
+      }),
+    );
+    await expect(
+      fetchRemoteData('https://slow.example/data.csv', { timeoutMs: 1 }),
+    ).rejects.toMatchObject({
+      reason: 'timeout',
+    });
+  });
+
+  test('the thrown error is a RemoteFetchError', async () => {
+    vi.stubGlobal(
+      'fetch',
+      vi.fn(() => Promise.reject(new TypeError('Failed to fetch'))),
+    );
+    await expect(fetchRemoteData('https://x.example')).rejects.toBeInstanceOf(RemoteFetchError);
+  });
+});

src/app/infrastructure/remote-data.ts

@@ -0,0 +1,99 @@
+/**
+ * Remote-data fetch adapter (snapshot model — spec §05 → URL datasets).
+ *
+ * The ONLY place Astrolabe touches the network. Adding a URL dataset fetches the
+ * resource once here; core then profiles the returned text exactly like inline
+ * data (see dataset.ts). Keeping the fetch in infrastructure preserves the rule
+ * that `src/core` stays pure and that the rest of the app never calls `fetch`
+ * directly (AGENTS → Architecture).
+ *
+ * Failures are classified into a small set of machine-readable `reason`s and
+ * thrown as `RemoteFetchError`; the user-facing wording (and the "paste inline
+ * instead" fallback) lives in the UI layer, where it is council-reviewed — the
+ * same split as `storage-errors`. A browser cross-origin block and an
+ * offline/DNS failure are indistinguishable here — both reject with a `TypeError`
+ * and no status — so they share the `network` reason and the UI hedges the cause.
+ */
+
+/** Why a remote fetch failed, in terms the UI maps to copy + a recovery path. */
+export type RemoteFetchReason = 'network' | 'http' | 'empty' | 'timeout';
+
+/** A classified remote-fetch failure. `status` is set only for `http`. */
+export class RemoteFetchError extends Error {
+  readonly reason: RemoteFetchReason;
+  readonly status?: number;
+
+  constructor(reason: RemoteFetchReason, message: string, status?: number) {
+    super(message);
+    this.name = 'RemoteFetchError';
+    this.reason = reason;
+    this.status = status;
+  }
+}
+
+/** A successful fetch: the raw body plus a weak format hint from the server. */
+export interface RemoteData {
+  /** The response body, ready for format detection + profiling. */
+  text: string;
+  /** The server's `Content-Type`, when provided — a tertiary format hint. */
+  contentType: string | null;
+}
+
+/** Default ceiling on a single fetch before it is aborted as a timeout. */
+const DEFAULT_TIMEOUT_MS = 30_000;
+
+export interface FetchRemoteDataOptions {
+  /** Abort the fetch after this many ms (default 30 s). */
+  timeoutMs?: number;
+}
+
+function describeNetworkError(err: unknown): string {
+  const reason = err instanceof Error ? err.message : String(err);
+  return `Couldn't reach the URL (${reason}).`;
+}
+
+/**
+ * Fetch a remote dataset resource once and return its raw body. Throws
+ * `RemoteFetchError` on any failure — the caller distinguishes recovery by
+ * `reason`. The body is returned verbatim; format detection and profiling happen
+ * downstream so this adapter stays free of core logic.
+ */
+export async function fetchRemoteData(
+  url: string,
+  options: FetchRemoteDataOptions = {},
+): Promise<RemoteData> {
+  const timeoutMs = options.timeoutMs ?? DEFAULT_TIMEOUT_MS;
+  const controller = new AbortController();
+  const timer = setTimeout(() => controller.abort(), timeoutMs);
+
+  let response: Response;
+  try {
+    response = await fetch(url, { redirect: 'follow', signal: controller.signal });
+  } catch (err) {
+    // A timeout aborts via the controller (AbortError); everything else here is a
+    // CORS block, offline, DNS failure, or malformed URL — all opaque TypeErrors.
+    if ((err as { name?: string }).name === 'AbortError') {
+      throw new RemoteFetchError('timeout', `The URL took longer than ${timeoutMs} ms to respond.`);
+    }
+    throw new RemoteFetchError('network', describeNetworkError(err));
+  } finally {
+    clearTimeout(timer);
+  }
+
+  if (!response.ok) {
+    throw new RemoteFetchError(
+      'http',
+      `The server responded ${response.status} ${response.statusText}.`.trim(),
+      response.status,
+    );
+  }
+
+  const text = await response.text();
+  if (text.trim() === '') {
+    throw new RemoteFetchError('empty', 'The URL returned no data.');
+  }
+
+  // TODO: guard very large responses (cap or warn) before snapshotting into
+  // IndexedDB — a multi-hundred-MB file would load fully into memory + storage.
+  return { text, contentType: response.headers.get('content-type') };
+}

src/app/services/remote-data-errors.test.ts

@@ -0,0 +1,38 @@
+import { describe, expect, test } from 'vitest';
+import { RemoteFetchError } from '../infrastructure/remote-data';
+import { remoteFetchErrorMessage } from './remote-data-errors';
+
+describe('remoteFetchErrorMessage', () => {
+  test('a network failure hedges offline vs cross-origin and offers the inline fallback', () => {
+    const msg = remoteFetchErrorMessage(new RemoteFetchError('network', 'x'));
+    expect(msg).toMatch(/offline/i);
+    expect(msg).toMatch(/cross-origin/i);
+    expect(msg).toMatch(/paste the data inline/i);
+  });
+
+  test('an HTTP 404 explains "not found" without a raw error code', () => {
+    const msg = remoteFetchErrorMessage(new RemoteFetchError('http', 'x', 404));
+    expect(msg).toMatch(/couldn't find/i);
+    expect(msg).not.toMatch(/404/);
+  });
+
+  test('an HTTP 403 explains refused access', () => {
+    expect(remoteFetchErrorMessage(new RemoteFetchError('http', 'x', 403))).toMatch(
+      /refused access/i,
+    );
+  });
+
+  test('an empty body says it returned no data', () => {
+    expect(remoteFetchErrorMessage(new RemoteFetchError('empty', 'x'))).toMatch(/no data/i);
+  });
+
+  test('the retry recovery points at refreshing, not pasting inline', () => {
+    const msg = remoteFetchErrorMessage(new RemoteFetchError('timeout', 'x'), 'retry');
+    expect(msg).toMatch(/refreshing again/i);
+    expect(msg).not.toMatch(/inline/i);
+  });
+
+  test('an unknown error still produces a sensible fallback message', () => {
+    expect(remoteFetchErrorMessage(new Error('boom'))).toMatch(/couldn't fetch this url/i);
+  });
+});

src/app/services/remote-data-errors.ts

@@ -0,0 +1,51 @@
+/**
+ * Translate a remote-data fetch failure (snapshot model — spec §05 → URL datasets)
+ * into the message shown when adding or refreshing a URL dataset.
+ *
+ * Pure and framework-free so the wording is unit-tested directly — the same split
+ * as `storage-errors`. Follows the recorded error-copy resolution (docs/architecture
+ * /10 → error copy): plain language, name what stopped, and — because every one of
+ * these is user-fixable — point at the next step. No error codes in the user-facing
+ * line (GOV.UK / NN/g #9). The recovery clause adapts to where the error shows: the
+ * create/edit form can fall back to pasting inline; Refresh can only retry.
+ */
+
+import { RemoteFetchError } from '../infrastructure/remote-data';
+
+/** Where the error surfaces, which decides the recovery the copy points at. */
+export type FetchRecovery = 'inline' | 'retry';
+
+function nextStep(recovery: FetchRecovery): string {
+  return recovery === 'inline'
+    ? 'Check the address, or paste the data inline instead.'
+    : 'Check the address and try refreshing again.';
+}
+
+function httpCause(status?: number): string {
+  if (status === 404) return "The server couldn't find anything at this URL.";
+  if (status === 401 || status === 403) return 'The server refused access to this URL.';
+  if (status && status >= 500) return 'The server hit an error serving this URL.';
+  return "The server couldn't return this URL.";
+}
+
+/**
+ * The user-facing message for a fetch failure. `recovery` defaults to `inline`
+ * (the add/edit form); pass `retry` for the Refresh action, which has no inline
+ * fallback.
+ */
+export function remoteFetchErrorMessage(err: unknown, recovery: FetchRecovery = 'inline'): string {
+  const next = nextStep(recovery);
+  if (err instanceof RemoteFetchError) {
+    switch (err.reason) {
+      case 'network':
+        return `Couldn't fetch this URL. Your device may be offline, or the host may block cross-origin requests. ${next}`;
+      case 'http':
+        return `${httpCause(err.status)} ${next}`;
+      case 'empty':
+        return `This URL returned no data. Make sure it points directly at a data file. ${next}`;
+      case 'timeout':
+        return `This URL took too long to respond. ${next}`;
+    }
+  }
+  return `Couldn't fetch this URL. ${next}`;
+}

src/app/stores/DatasetStore.test.ts

@@ -42,15 +42,82 @@ describe('save — create', () => {
     expect(ds?.rowCount).toBe(2);
   });
 
-  test('a valid URL infers format from the extension', () => {
+  test('save() does not commit a URL create — it must go through commitUrlSnapshot', () => {
+    store().startCreate();
+    store().updateForm({ name: 'Remote', source: 'url', input: 'https://example.com/data.csv' });
+    // The component fetches first; save() never touches the network, so a URL create
+    // through save() is a no-op rather than an unfetched record.
+    expect(store().save(T)).toBe(false);
+    expect(store().datasets).toHaveLength(0);
+  });
+});
+
+describe('commitUrlSnapshot — create from a fetched body', () => {
+  test('snapshots the body, infers format from content, and profiles it', () => {
     store().startCreate();
     store().updateForm({ name: 'Remote', source: 'url', input: 'https://example.com/data.csv' });
 
-    expect(store().save(T)).toBe(true);
+    expect(store().commitUrlSnapshot({ text: 'a,b\n1,2\n3,4' }, T)).toBe(true);
     const ds = selectSelectedDataset(store());
     expect(ds?.source).toBe('url');
+    expect(ds?.url).toBe('https://example.com/data.csv');
     expect(ds?.format).toBe('csv');
-    expect(ds?.rowCount).toBeNull(); // URL datasets aren't profiled
+    expect(ds?.data).toBe('a,b\n1,2\n3,4');
+    expect(ds?.rowCount).toBe(2);
+    expect(ds?.columns).toEqual(['a', 'b']);
+    expect(ds?.fetchedAt).toBe(T.toISOString());
+    expect(store().view).toBe('detail');
+  });
+
+  test('a duplicate name is rejected even with a successful fetch', () => {
+    store().add(
+      createDataset({ name: 'Dupe', data: [{ a: 1 }], format: 'json', source: 'inline', now: T }),
+    );
+    store().startCreate();
+    store().updateForm({ name: 'dupe', source: 'url', input: 'https://x/y.csv' });
+    expect(store().commitUrlSnapshot({ text: 'a\n1' }, T)).toBe(false);
+    expect(store().formError).toMatch(/already exists/i);
+  });
+});
+
+describe('URL dataset edits & refresh', () => {
+  const seedUrl = () => {
+    store().startCreate();
+    store().updateForm({ name: 'Remote', source: 'url', input: 'https://x/y.csv' });
+    store().commitUrlSnapshot({ text: 'a,b\n1,2' }, T);
+    return store().selectedId!;
+  };
+
+  test('a metadata-only edit (same URL) updates name without re-fetching', () => {
+    const id = seedUrl();
+    store().startEdit();
+    store().updateForm({ name: 'Renamed' }); // URL field stays https://x/y.csv
+    const later = new Date('2026-07-01T00:00:00Z');
+    expect(store().save(later)).toBe(true);
+    const ds = store().datasets.find((d) => d.id === id)!;
+    expect(ds.name).toBe('Renamed');
+    expect(ds.data).toBe('a,b\n1,2'); // snapshot untouched
+    expect(ds.fetchedAt).toBe(T.toISOString()); // not re-fetched
+    expect(ds.modified).toBe(later.toISOString());
+  });
+
+  test('refreshDataset re-snapshots, re-profiles, and advances fetchedAt', () => {
+    const id = seedUrl();
+    const later = new Date('2026-07-01T00:00:00Z');
+    expect(store().refreshDataset(id, { text: 'a,b\n1,2\n3,4\n5,6' }, later)).toBe(true);
+    const ds = store().datasets.find((d) => d.id === id)!;
+    expect(ds.rowCount).toBe(3);
+    expect(ds.fetchedAt).toBe(later.toISOString());
+    expect(ds.modified).toBe(later.toISOString());
+    expect(ds.name).toBe('Remote'); // preserved
+  });
+
+  test('refreshDataset is a no-op for an inline dataset', () => {
+    store().add(
+      createDataset({ name: 'Inline', data: [{ a: 1 }], format: 'json', source: 'inline', now: T }),
+    );
+    const id = store().selectedId!;
+    expect(store().refreshDataset(id, { text: 'a\n1' }, T)).toBe(false);
   });
 });
 

src/app/stores/DatasetStore.ts

@@ -19,8 +19,14 @@
  */
 
 import { create } from 'zustand';
-import { detectFormat, detectFormatFromUrl, type DataFormat } from '@core/format-detection';
+import { detectFormat, type DataFormat } from '@core/format-detection';
-import { createDataset, computeDatasetProfile, type DataSource, type Dataset } from '@core/dataset';
+import {
+  createDataset,
+  computeDatasetProfile,
+  snapshotFromText,
+  type DataSource,
+  type Dataset,
+} from '@core/dataset';
 import { isNameTaken } from '@core/naming';
 import { useSnippetStore } from './SnippetStore';
 
@@ -67,6 +73,21 @@ export interface DatasetState {
    */
   save: (now?: Date) => boolean;
 
+  /**
+   * Commit a URL dataset from an already-fetched body. The component performs the
+   * network fetch via the remote-data adapter and passes the result here, keeping
+   * this store browser-free. Validates the form, snapshots + profiles the body, and
+   * creates (view `new`) or updates (view `edit`) the dataset — including renaming
+   * referencing snippets on an edit. Returns whether it committed.
+   */
+  commitUrlSnapshot: (fetched: { text: string }, now?: Date) => boolean;
+  /**
+   * Re-snapshot an existing URL dataset from a freshly-fetched body ("Refresh"):
+   * re-detect the format, re-profile, and advance `fetchedAt`/`modified`. Name,
+   * comment, and URL are preserved. Returns false for a non-URL dataset.
+   */
+  refreshDataset: (id: number, fetched: { text: string }, now?: Date) => boolean;
+
   /**
    * Low-level: add a fully-formed dataset and select it. The id is (re)assigned
    * via `nextDatasetId`, so a `createDataset` default id (`Date.now()`) can never
@@ -144,7 +165,12 @@ function validateForm(
   return null;
 }
 
-/** Validate a form into a saveable shape, or return an error message. */
+/**
+ * Validate an **inline** form into a saveable shape, or return an error message.
+ * URL datasets never reach here — they are fetched first and committed through
+ * `commitUrlSnapshot` (which snapshots the fetched body), so this only shapes the
+ * pasted-inline payload.
+ */
 function resolveForm(
   form: DatasetForm,
   datasets: Dataset[],
@@ -153,20 +179,12 @@ function resolveForm(
   const error = validateForm(form, datasets, excludeId);
   if (error) return { error };
 
-  const name = form.name.trim();
-  const input = form.input.trim();
-
-  if (form.source === 'url') {
-    // Format is inferred from the extension; default to JSON when unknown (§05).
-    return { name, data: input, format: detectFormatFromUrl(input) ?? 'json', source: 'url' };
-  }
-
   // validateForm guaranteed a detectable format above.
   const format = detectFormat(form.input).format as DataFormat;
   // JSON/TopoJSON are stored parsed; CSV/TSV keep their raw text (data model §09B).
   const data =
     format === 'json' || format === 'topojson' ? (JSON.parse(form.input) as unknown) : form.input;
-  return { name, data, format, source: 'inline' };
+  return { name: form.name.trim(), data, format, source: 'inline' };
 }
 
 export const useDatasetStore = create<DatasetState>((set, get) => ({
@@ -201,12 +219,14 @@ export const useDatasetStore = create<DatasetState>((set, get) => ({
         name: ds.name,
         source: ds.source,
         comment: ds.comment,
-        // Re-render the stored payload as editable text: raw for csv/tsv/url,
+        // Re-render the editable text: the URL for url datasets, raw text for
-        // pretty-printed JSON for json/topojson.
+        // csv/tsv, pretty-printed JSON for json/topojson.
         input:
-          ds.source === 'url' || ds.format === 'csv' || ds.format === 'tsv'
+          ds.source === 'url'
-            ? String(ds.data)
+            ? (ds.url ?? '')
-            : JSON.stringify(ds.data, null, 2),
+            : ds.format === 'csv' || ds.format === 'tsv'
+              ? String(ds.data)
+              : JSON.stringify(ds.data, null, 2),
       },
     });
   },
@@ -225,6 +245,30 @@ export const useDatasetStore = create<DatasetState>((set, get) => ({
     const editing = view === 'edit';
     const excludeId = editing ? (selectedId ?? undefined) : undefined;
 
+    // URL datasets that need the network — a create, a URL change, or an inline→URL
+    // conversion — are routed through `commitUrlSnapshot` after the component
+    // fetches; save() never fetches. The one URL case it commits is a metadata-only
+    // edit of an existing snapshot (same URL): just update name/comment, no re-fetch.
+    if (form.source === 'url') {
+      const error = validateForm(form, datasets, excludeId);
+      if (error) {
+        set({ formError: error });
+        return false;
+      }
+      if (!editing || selectedId === null) return false;
+      const existing = datasets.find((d) => d.id === selectedId);
+      if (!existing || existing.source !== 'url' || form.input.trim() !== (existing.url ?? '')) {
+        return false;
+      }
+      const name = form.name.trim();
+      get().update(selectedId, { name, comment: form.comment }, now);
+      if (name !== existing.name) {
+        useSnippetStore.getState().renameDatasetRefs(existing.name, name, now);
+      }
+      set({ view: 'detail', form: EMPTY_FORM, formError: null });
+      return true;
+    }
+
     const resolved = resolveForm(form, datasets, excludeId);
     if ('error' in resolved) {
       set({ formError: resolved.error });
@@ -234,11 +278,12 @@ export const useDatasetStore = create<DatasetState>((set, get) => ({
     if (editing && selectedId !== null) {
       const existing = datasets.find((d) => d.id === selectedId);
       if (!existing) return false;
-      const profile = computeDatasetProfile(resolved.data, resolved.format, resolved.source);
+      const profile = computeDatasetProfile(resolved.data, resolved.format);
       // Re-profile and update the record (including any new name), then propagate
       // the rename across referencing snippets so each spec and its datasetRefs
       // stay consistent (docs/architecture/07 §6). SnippetStore never imports this
-      // store, so the direct call is cycle-free.
+      // store, so the direct call is cycle-free. Clear any url/fetchedAt left over
+      // from a URL→inline conversion so the record carries no stale remote origin.
       get().update(
         selectedId,
         {
@@ -246,6 +291,13 @@ export const useDatasetStore = create<DatasetState>((set, get) => ({
           data: resolved.data,
           format: resolved.format,
           source: resolved.source,
+          // TODO: `update` shallow-merges, so these set the keys to `undefined`
+          // rather than removing them — a URL→inline record keeps `url`/`fetchedAt`
+          // present-but-undefined. Invisible today (rendering/profiling key off
+          // `source`/`data == null`, and JSON export drops undefined), but it breaks
+          // the "inline records carry no url/fetchedAt keys" invariant on this path.
+          url: undefined,
+          fetchedAt: undefined,
           comment: form.comment,
           ...profile,
         },
@@ -275,6 +327,76 @@ export const useDatasetStore = create<DatasetState>((set, get) => ({
     return true;
   },
 
+  commitUrlSnapshot: (fetched, now) => {
+    const { view, form, datasets, selectedId } = get();
+    const editing = view === 'edit';
+    const excludeId = editing ? (selectedId ?? undefined) : undefined;
+
+    const error = validateForm(form, datasets, excludeId);
+    if (error) {
+      set({ formError: error });
+      return false;
+    }
+
+    const name = form.name.trim();
+    const url = form.input.trim();
+    const iso = (now ?? new Date()).toISOString();
+    // The fetched body snapshots + profiles exactly like inline data (snapshot
+    // model): detect format from content, shape, and profile through core.
+    const { data, format } = snapshotFromText(fetched.text, url);
+
+    if (editing && selectedId !== null) {
+      const existing = datasets.find((d) => d.id === selectedId);
+      if (!existing) return false;
+      const profile = computeDatasetProfile(data, format);
+      get().update(
+        selectedId,
+        {
+          name,
+          data,
+          format,
+          source: 'url',
+          url,
+          fetchedAt: iso,
+          comment: form.comment,
+          ...profile,
+        },
+        now,
+      );
+      if (name !== existing.name) {
+        useSnippetStore.getState().renameDatasetRefs(existing.name, name, now);
+      }
+      set({ view: 'detail', form: EMPTY_FORM, formError: null });
+      return true;
+    }
+
+    const dataset = createDataset({
+      name,
+      data,
+      format,
+      source: 'url',
+      url,
+      fetchedAt: iso,
+      comment: form.comment,
+      now,
+    });
+    get().add(dataset);
+    set({ view: 'detail', form: EMPTY_FORM, formError: null });
+    return true;
+  },
+
+  refreshDataset: (id, fetched, now) => {
+    const dataset = get().datasets.find((d) => d.id === id);
+    if (!dataset || dataset.source !== 'url' || !dataset.url) return false;
+    const iso = (now ?? new Date()).toISOString();
+    const { data, format } = snapshotFromText(fetched.text, dataset.url);
+    const profile = computeDatasetProfile(data, format);
+    // Name, comment, and url are preserved; only the snapshot + its profile and the
+    // fetch time change. `update` advances `modified`.
+    get().update(id, { data, format, fetchedAt: iso, ...profile }, now);
+    return true;
+  },
+
   add: (dataset) =>
     set((s) => {
       const withId = { ...dataset, id: nextDatasetId(s.datasets) };

src/core/dataset.test.ts

@@ -5,6 +5,8 @@ import {
   createDataset,
   datasetReference,
   parseDelimited,
+  snapshotFromText,
+  tabularRows,
 } from './dataset';
 
 describe('datasetReference', () => {
@@ -65,51 +67,125 @@ describe('parseDelimited', () => {
 });
 
 describe('computeDatasetProfile', () => {
-  test('inline JSON array-of-objects is profiled', () => {
+  test('JSON array-of-objects is profiled', () => {
     const data = [
       { a: 1, b: 'x' },
       { a: 2, b: 'y' },
     ];
-    const profile = computeDatasetProfile(data, 'json', 'inline');
+    const profile = computeDatasetProfile(data, 'json');
     expect(profile.rowCount).toBe(2);
     expect(profile.columns).toEqual(['a', 'b']);
     expect(profile.size).toBe(new TextEncoder().encode(JSON.stringify(data)).length);
   });
 
-  test('inline JSON that is not an array is N/A but sized', () => {
+  test('JSON that is not an array is N/A but sized', () => {
-    const profile = computeDatasetProfile({ a: 1 }, 'json', 'inline');
+    const profile = computeDatasetProfile({ a: 1 }, 'json');
     expect(profile.rowCount).toBeNull();
     expect(profile.size).toBeGreaterThan(0);
   });
 
-  test('inline CSV is parsed and profiled; size is the raw text byte length', () => {
+  test('CSV is parsed and profiled; size is the raw text byte length', () => {
     const text = 'a,b\n1,2';
-    const profile = computeDatasetProfile(text, 'csv', 'inline');
+    const profile = computeDatasetProfile(text, 'csv');
     expect(profile.rowCount).toBe(1);
     expect(profile.columns).toEqual(['a', 'b']);
     expect(profile.size).toBe(new TextEncoder().encode(text).length);
   });
 
-  test('inline TSV is parsed and profiled', () => {
+  test('TSV is parsed and profiled', () => {
-    const profile = computeDatasetProfile('a\tb\n1\t2', 'tsv', 'inline');
+    const profile = computeDatasetProfile('a\tb\n1\t2', 'tsv');
     expect(profile.rowCount).toBe(1);
     expect(profile.columns).toEqual(['a', 'b']);
   });
 
-  test('inline TopoJSON is N/A (non-tabular) but sized', () => {
+  test('TopoJSON is N/A (non-tabular) but sized', () => {
     const topo = { type: 'Topology', objects: {} };
-    const profile = computeDatasetProfile(topo, 'topojson', 'inline');
+    const profile = computeDatasetProfile(topo, 'topojson');
     expect(profile.rowCount).toBeNull();
     expect(profile.columnCount).toBeNull();
     expect(profile.size).toBe(new TextEncoder().encode(JSON.stringify(topo)).length);
   });
 
-  test('URL is N/A; size is the byte length of the URL string', () => {
+  test('a fetched URL snapshot is profiled exactly like inline data', () => {
-    const url = 'https://example.com/data.csv';
+    // The snapshot model: once fetched, a URL dataset carries its payload in `data`
-    const profile = computeDatasetProfile(url, 'csv', 'url');
+    // and profiles through the same path as inline (no source distinction here).
+    const profile = computeDatasetProfile('a,b\n1,2\n3,4', 'csv');
+    expect(profile.rowCount).toBe(2);
+    expect(profile.columns).toEqual(['a', 'b']);
+  });
+
+  test('an unfetched URL reference (null data) is N/A with size 0', () => {
+    const profile = computeDatasetProfile(null, 'csv');
     expect(profile.rowCount).toBeNull();
     expect(profile.columnCount).toBeNull();
-    expect(profile.size).toBe(new TextEncoder().encode(url).length);
+    expect(profile.columns).toEqual([]);
+    expect(profile.size).toBe(0);
+  });
+});
+
+describe('tabularRows', () => {
+  test('CSV parses to rows (same parsing as profiling)', () => {
+    expect(tabularRows('a,b\n1,2\n3,4', 'csv')).toEqual([
+      { a: '1', b: '2' },
+      { a: '3', b: '4' },
+    ]);
+  });
+
+  test('TSV parses to rows', () => {
+    expect(tabularRows('x\ty\n1\t2', 'tsv')).toEqual([{ x: '1', y: '2' }]);
+  });
+
+  test('a JSON array of objects is returned as-is', () => {
+    const data = [{ a: 1 }, { a: 2 }];
+    expect(tabularRows(data, 'json')).toEqual(data);
+  });
+
+  test('limit returns only the head', () => {
+    const rows = tabularRows('n\n1\n2\n3\n4', 'csv', 2);
+    expect(rows).toEqual([{ n: '1' }, { n: '2' }]);
+  });
+
+  test('non-tabular payloads return null (single object, topojson, unfetched)', () => {
+    expect(tabularRows({ a: 1 }, 'json')).toBeNull();
+    expect(tabularRows({ type: 'Topology', objects: {} }, 'topojson')).toBeNull();
+    expect(tabularRows(null, 'csv')).toBeNull();
+    expect(tabularRows('a,b', 'csv')).toBeNull(); // header only — no rows
+  });
+});
+
+describe('snapshotFromText', () => {
+  test('JSON content is parsed and typed json (content beats extension)', () => {
+    const { data, format } = snapshotFromText('[{"a":1}]', 'https://x/data.txt');
+    expect(format).toBe('json');
+    expect(data).toEqual([{ a: 1 }]);
+  });
+
+  test('CSV content keeps its raw text and is typed csv', () => {
+    const { data, format } = snapshotFromText('a,b\n1,2', 'https://x/data.csv');
+    expect(format).toBe('csv');
+    expect(data).toBe('a,b\n1,2');
+  });
+
+  test('TSV content is typed tsv', () => {
+    expect(snapshotFromText('a\tb\n1\t2', 'https://x/d').format).toBe('tsv');
+  });
+
+  test('a Topology body is typed topojson and parsed', () => {
+    const { data, format } = snapshotFromText('{"type":"Topology","objects":{}}', 'https://x/d');
+    expect(format).toBe('topojson');
+    expect(data).toEqual({ type: 'Topology', objects: {} });
+  });
+
+  test('unrecognized content falls back to the URL extension', () => {
+    // A single token: not JSON, not delimited. The .json extension decides the
+    // format; the unparseable body is kept verbatim rather than throwing.
+    const { data, format } = snapshotFromText('not-data', 'https://x/data.json');
+    expect(format).toBe('json');
+    expect(data).toBe('not-data');
+  });
+
+  test('unrecognized content with no useful extension defaults to json', () => {
+    expect(snapshotFromText('not-data', 'https://x/feed').format).toBe('json');
   });
 });
 
@@ -157,20 +233,45 @@ describe('createDataset', () => {
     expect(d.columns).toEqual(['a', 'b']);
   });
 
-  test('URL dataset gets an N/A profile with a size', () => {
+  test('URL dataset carries url + fetchedAt and profiles the fetched snapshot', () => {
     const d = createDataset({
       id: 2,
       name: 'Remote',
-      data: 'https://example.com/x.json',
+      data: 'a,b\n1,2',
-      format: 'json',
+      url: 'https://example.com/x.csv',
+      fetchedAt: '2026-06-10T00:00:00.000Z',
+      format: 'csv',
       source: 'url',
       comment: 'remote source',
     });
-    expect(d.rowCount).toBeNull();
+    expect(d.source).toBe('url');
-    expect(d.columnCount).toBeNull();
+    expect(d.url).toBe('https://example.com/x.csv');
-    expect(d.columns).toEqual([]);
+    expect(d.fetchedAt).toBe('2026-06-10T00:00:00.000Z');
+    expect(d.rowCount).toBe(1);
+    expect(d.columns).toEqual(['a', 'b']);
     expect(d.comment).toBe('remote source');
-    expect(d.size).toBeGreaterThan(0);
+  });
+
+  test('an unfetched URL dataset (null data) gets an N/A profile and null fetchedAt', () => {
+    const d = createDataset({
+      id: 3,
+      name: 'Unfetched',
+      data: null,
+      url: 'https://example.com/x.csv',
+      format: 'csv',
+      source: 'url',
+    });
+    expect(d.url).toBe('https://example.com/x.csv');
+    expect(d.fetchedAt).toBeNull();
+    expect(d.rowCount).toBeNull();
+    expect(d.columns).toEqual([]);
+    expect(d.size).toBe(0);
+  });
+
+  test('inline dataset carries no url/fetchedAt keys', () => {
+    const d = createDataset({ id: 4, name: 'I', data: [], format: 'json', source: 'inline' });
+    expect('url' in d).toBe(false);
+    expect('fetchedAt' in d).toBe(false);
   });
 
   test('defaults the id when not injected', () => {

src/core/dataset.ts

@@ -7,22 +7,29 @@
  * a simple delimited-text parser, the profiling orchestration, and a factory that
  * stamps timestamps/version and fills the derived summary fields.
  *
- * A dataset has one of two **sources** — `inline` (data stored in the record) or
+ * A dataset has one of two **sources** — `inline` (data pasted into the record) or
- * `url` (only the link is stored, fetched on demand at render time) — and one of
+ * `url` (fetched once from a remote address and **snapshotted** into the record) —
- * four **formats** (reused from format-detection: `json`/`csv`/`tsv`/`topojson`).
+ * and one of four **formats** (reused from format-detection: `json`/`csv`/`tsv`/
- * The `data` field's shape follows source/format: a URL string for `url`; raw
+ * `topojson`). Either way `data` holds the actual payload, shaped by format: raw
- * text for inline CSV/TSV; a parsed value for inline JSON/TopoJSON.
+ * text for CSV/TSV, a parsed value for JSON/TopoJSON. A `url` dataset additionally
+ * keeps its source `url` (so it can be re-fetched) and a `fetchedAt` timestamp;
+ * until its first successful fetch `data` is `null` (an unfetched reference).
  *
- * Only tabular inline data (JSON array-of-objects, CSV, TSV) is profiled; URL and
+ * Any tabular payload (JSON array-of-objects, CSV, TSV) is profiled — including a
- * non-tabular data get an N/A profile but are still sized (see profile.ts).
+ * fetched URL snapshot; non-tabular or not-yet-fetched data gets an N/A profile but
+ * is still sized (see profile.ts).
  */
 
-import type { DataFormat } from './format-detection';
+import { detectFormat, detectFormatFromUrl, type DataFormat } from './format-detection';
 import { profileData, type ColumnStats, type DatasetProfile } from './profile';
 import type { ColumnType } from './type-inference';
 
-/** Current schema version for a Dataset record (read-time migration target). */
+/**
-export const CURRENT_DATASET_VERSION = 1;
+ * Current schema version for a Dataset record (read-time migration target).
+ * v2 moved a URL dataset's address out of `data` into its own `url` field and made
+ * `data` hold the fetched snapshot (see `migrateDataset`).
+ */
+export const CURRENT_DATASET_VERSION = 2;
 
 /** Where a dataset's data lives: embedded in the record, or fetched from a URL. */
 export type DataSource = 'inline' | 'url';
@@ -35,14 +42,25 @@ export interface Dataset {
   /** Unique, human-readable name; the key snippets reference via `datasetRefs`. */
   name: string;
   /**
-   * The payload. For `source = url`: the URL string. For `source = inline`: the
+   * The payload, shaped by format: raw CSV/TSV text, or the parsed JSON/TopoJSON
-   * raw CSV/TSV text, or the parsed JSON/TopoJSON value.
+   * value. For `source = url` this is the fetched snapshot, or `null` before the
+   * first successful fetch.
    */
   data: unknown;
   /** One of `json`, `csv`, `tsv`, `topojson`. */
   format: DataFormat;
   /** One of `inline` or `url`. */
   source: DataSource;
+  /**
+   * For `source = url`: the remote address the snapshot was fetched from, retained
+   * so the dataset can be re-fetched ("Refresh"). Absent for inline datasets.
+   */
+  url?: string;
+  /**
+   * For `source = url`: ISO timestamp of the last successful fetch, or `null` when
+   * it has never been fetched. Absent for inline datasets.
+   */
+  fetchedAt?: string | null;
   /** Free-form user note about the dataset. */
   comment: string;
   /** Data rows, or `null` when N/A (URL / non-tabular). */
@@ -187,26 +205,52 @@ function asObjectRows(value: unknown): Array<Record<string, unknown>> | null {
 }
 
 /**
- * Orchestrate profiling for a dataset payload: compute `size` (always), decide
+ * The tabular rows of a dataset payload for a table preview, or `null` when the
- * tabular vs N/A by source/format, and delegate to `profileData`.
+ * payload isn't tabular (a single JSON object, TopoJSON, or an unfetched URL). Uses
- *
+ * the **same** parsing as profiling — `parseDelimited` for CSV/TSV, `asObjectRows`
- *   - `url` (any format)        → N/A profile; size = byte length of the URL string.
+ * for JSON — so the previewed rows agree exactly with the profiled `columns`. A
- *   - inline `json`             → rows when a non-empty array of objects, else N/A.
+ * positive `limit` returns only the head (a preview needs a sample, not the whole
- *   - inline `topojson`         → N/A (non-tabular).
+ * payload). Returns non-null on precisely the inputs `computeDatasetProfile` counts
- *   - inline `csv` / `tsv`      → `parseDelimited` rows.
+ * as rows, so a caller can gate "table vs. raw text" on this alone.
- *
- * `size` is the UTF-8 byte length of the raw string for csv/tsv/url, or of
- * `JSON.stringify(data)` for json/topojson.
  */
-export function computeDatasetProfile(
+export function tabularRows(
   data: unknown,
   format: DataFormat,
-  source: DataSource,
+  limit?: number,
-): DatasetProfile {
+): Array<Record<string, unknown>> | null {
-  if (source === 'url') {
+  if (data == null) return null;
-    const url = typeof data === 'string' ? data : (JSON.stringify(data) ?? '');
+  let rows: Array<Record<string, unknown>> | null;
-    return profileData(null, byteLength(url));
+  switch (format) {
+    case 'csv':
+    case 'tsv':
+      rows = parseDelimited(typeof data === 'string' ? data : '', format);
+      break;
+    case 'json':
+      rows = asObjectRows(data);
+      break;
+    default:
+      rows = null;
   }
+  if (!rows || rows.length === 0) return null;
+  return limit != null && limit >= 0 && rows.length > limit ? rows.slice(0, limit) : rows;
+}
+
+/**
+ * Orchestrate profiling for a dataset payload: compute `size` (always), decide
+ * tabular vs N/A by format, and delegate to `profileData`. Identical for inline
+ * data and for a fetched URL snapshot — both carry the payload in `data`.
+ *
+ *   - `null` data (unfetched URL) → N/A profile, size 0.
+ *   - `json`                      → rows when a non-empty array of objects, else N/A.
+ *   - `topojson`                  → N/A (non-tabular).
+ *   - `csv` / `tsv`               → `parseDelimited` rows.
+ *
+ * `size` is the UTF-8 byte length of the raw string for csv/tsv, or of
+ * `JSON.stringify(data)` for json/topojson.
+ */
+export function computeDatasetProfile(data: unknown, format: DataFormat): DatasetProfile {
+  // An unfetched URL reference (or a genuinely absent payload): nothing to profile.
+  if (data == null) return profileData(null, 0);
 
   switch (format) {
     case 'csv':
@@ -226,6 +270,29 @@ export function computeDatasetProfile(
   }
 }
 
+/**
+ * Shape a freshly-fetched URL body into the `{ data, format }` a snapshot stores
+ * (spec §05 → URL datasets, snapshot model). Format is sniffed from the **content**
+ * first — authoritative, since `detectFormat` only reports `json` when the body
+ * actually parses — falling back to the URL's file extension, then JSON.
+ * JSON/TopoJSON are stored parsed; CSV/TSV keep their raw text — the same
+ * per-format shaping inline data uses, so a fetched dataset profiles and renders
+ * identically to an inline one (see `computeDatasetProfile`, rendering.ts).
+ */
+export function snapshotFromText(text: string, url: string): { data: unknown; format: DataFormat } {
+  const format = detectFormat(text).format ?? detectFormatFromUrl(url) ?? 'json';
+  if (format === 'json' || format === 'topojson') {
+    try {
+      return { data: JSON.parse(text) as unknown, format };
+    } catch {
+      // The extension promised JSON but the body isn't — keep the raw text so the
+      // render surfaces a readable error instead of us throwing mid-commit.
+      return { data: text, format };
+    }
+  }
+  return { data: text, format };
+}
+
 export interface CreateDatasetOptions {
   /** The dataset name (uniqueness is enforced upstream — see naming.ts). */
   name: string;
@@ -235,6 +302,10 @@ export interface CreateDatasetOptions {
   format: DataFormat;
   /** One of `inline` or `url`. */
   source: DataSource;
+  /** For `source = url`: the remote address (retained for Refresh). */
+  url?: string;
+  /** For `source = url`: ISO timestamp of the fetch that produced `data`. */
+  fetchedAt?: string | null;
   /** Optional free-form note. */
   comment?: string;
   /** Clock injection for deterministic tests; defaults to the current time. */
@@ -253,7 +324,7 @@ export interface CreateDatasetOptions {
 export function createDataset(options: CreateDatasetOptions): Dataset {
   const now = options.now ?? new Date();
   const iso = now.toISOString();
-  const profile = computeDatasetProfile(options.data, options.format, options.source);
+  const profile = computeDatasetProfile(options.data, options.format);
 
   return {
     id: options.id ?? Date.now(),
@@ -262,6 +333,8 @@ export function createDataset(options: CreateDatasetOptions): Dataset {
     data: options.data,
     format: options.format,
     source: options.source,
+    // URL datasets carry their address + fetch time; inline records stay clean.
+    ...(options.source === 'url' ? { url: options.url, fetchedAt: options.fetchedAt ?? null } : {}),
     comment: options.comment ?? '',
     rowCount: profile.rowCount,
     columnCount: profile.columnCount,

src/core/import-normalize.test.ts

@@ -233,6 +233,42 @@ describe('normalizeImport — dataset normalization', () => {
     expect(d.version).toBe(CURRENT_DATASET_VERSION);
   });
 
+  it('v1→v2: a legacy URL dataset (address in data) imports as an unfetched reference', () => {
+    const parsed = {
+      version: '1.0',
+      snippets: [],
+      datasets: [{ id: 1, name: 'Remote', source: 'url', format: 'csv', data: 'https://x/y.csv' }],
+    };
+    const d = normalizeImport(parsed, { now: FIXED_NOW }).datasets[0];
+    expect(d.source).toBe('url');
+    expect(d.url).toBe('https://x/y.csv');
+    expect(d.data).toBeNull();
+    expect(d.fetchedAt).toBeNull();
+    expect(d.version).toBe(CURRENT_DATASET_VERSION);
+  });
+
+  it('a v2 URL snapshot imports with its data, url, and fetchedAt intact', () => {
+    const parsed = {
+      version: '1.0',
+      snippets: [],
+      datasets: [
+        {
+          id: 1,
+          name: 'Remote',
+          source: 'url',
+          format: 'csv',
+          data: 'a,b\n1,2',
+          url: 'https://x/y.csv',
+          fetchedAt: '2026-06-10T00:00:00.000Z',
+        },
+      ],
+    };
+    const d = normalizeImport(parsed, { now: FIXED_NOW }).datasets[0];
+    expect(d.data).toBe('a,b\n1,2');
+    expect(d.url).toBe('https://x/y.csv');
+    expect(d.fetchedAt).toBe('2026-06-10T00:00:00.000Z');
+  });
+
   it('fills gaps with defaults and coerces id to a number', () => {
     const parsed = { version: '1.0', snippets: [], datasets: [{ id: '42', name: 'D' }] };
     const result = normalizeImport(parsed, { now: FIXED_NOW });

src/core/import-normalize.ts

@@ -151,6 +151,11 @@ function normalizeSnippet(raw: unknown, nowIso: string, makeId: () => string): S
  * datasets already carry their derived summary fields (rowCount, columns, …); we
  * preserve those and only fill gaps. We do NOT re-profile here — that needs the
  * profiling pipeline and would be wasteful for already-summarized records.
+ *
+ * Applies the v1→v2 URL-snapshot shaping (mirrors `migrateDataset`): a pre-v2 export
+ * stored a URL dataset's address in `data`, so we move it into `url` and clear `data`
+ * to `null` — importing such a record yields an unfetched reference, not a record
+ * with a URL string masquerading as its snapshot.
  */
 function normalizeDataset(raw: unknown, nowIso: string): Dataset {
   const r = isPlainObject(raw) ? raw : {};
@@ -158,13 +163,25 @@ function normalizeDataset(raw: unknown, nowIso: string): Dataset {
   const created = asNonEmptyString(r.created) ?? nowIso;
   const modified = asNonEmptyString(r.modified) ?? created;
 
+  const source = (typeof r.source === 'string' ? r.source : 'inline') as DataSource;
+  const legacyUrl = source === 'url' && typeof r.url !== 'string';
+  const url = legacyUrl
+    ? typeof r.data === 'string'
+      ? r.data
+      : undefined
+    : typeof r.url === 'string'
+      ? r.url
+      : undefined;
+  const fetchedAt = legacyUrl ? null : typeof r.fetchedAt === 'string' ? r.fetchedAt : null;
+
   return {
     id: typeof r.id === 'number' ? r.id : Number(r.id) || 0,
     version: CURRENT_DATASET_VERSION,
     name: typeof r.name === 'string' ? r.name : 'Untitled',
-    data: r.data,
+    data: legacyUrl ? null : r.data,
     format: (typeof r.format === 'string' ? r.format : 'json') as DataFormat,
-    source: (typeof r.source === 'string' ? r.source : 'inline') as DataSource,
+    source,
+    ...(source === 'url' ? { url, fetchedAt } : {}),
     comment: typeof r.comment === 'string' ? r.comment : '',
     rowCount: typeof r.rowCount === 'number' ? r.rowCount : null,
     columnCount: typeof r.columnCount === 'number' ? r.columnCount : null,

src/core/rendering.test.ts

@@ -111,7 +111,16 @@ describe('prepareSpecForRender — dataset resolution (spec §04 Rendering Contr
       format: 'topojson',
       source: 'inline',
     },
-    { name: 'UrlDs', data: 'https://x/y.csv', format: 'csv', source: 'url' },
+    // A fetched URL snapshot carries its payload in `data` (like inline); an
+    // unfetched reference has `data: null` and only its `url`.
+    {
+      name: 'UrlFetchedDs',
+      data: 'a,b\n1,2',
+      url: 'https://x/y.csv',
+      format: 'csv',
+      source: 'url',
+    },
+    { name: 'UrlUnfetchedDs', data: null, url: 'https://x/y.csv', format: 'csv', source: 'url' },
   ];
 
   test('inline JSON → values inlined', () => {
@@ -140,8 +149,13 @@ describe('prepareSpecForRender — dataset resolution (spec §04 Rendering Contr
     });
   });
 
-  test('URL → url reference tagged with the dataset format', () => {
+  test('fetched URL snapshot → its payload inlined, tagged with the format (like inline)', () => {
-    const out = prepareSpecForRender({ data: { name: 'UrlDs' } }, { datasets });
+    const out = prepareSpecForRender({ data: { name: 'UrlFetchedDs' } }, { datasets });
+    expect(out.data).toEqual({ values: 'a,b\n1,2', format: { type: 'csv' } });
+  });
+
+  test('unfetched URL reference → live url fallback tagged with the format', () => {
+    const out = prepareSpecForRender({ data: { name: 'UrlUnfetchedDs' } }, { datasets });
     expect(out.data).toEqual({ url: 'https://x/y.csv', format: { type: 'csv' } });
   });
 

src/core/rendering.ts

@@ -42,6 +42,8 @@ export interface ResolvableDataset {
   format: DataFormat;
   /** One of `inline` or `url`. */
   source: DataSource;
+  /** For `source = url`: the remote address, used only as the unfetched fallback. */
+  url?: string;
 }
 
 /** Thrown when a spec references a library dataset name that does not exist. */
@@ -129,13 +131,20 @@ function selfDefinedDatasetNames(spec: unknown): Set<string> {
  */
 function resolvedData(dataset: ResolvableDataset, rest: SpecNode): SpecNode {
   const restFormat = isSpecNode(rest.format) ? rest.format : {};
-  if (dataset.source === 'url') {
+
+  // An unfetched URL reference (a legacy record, or one whose snapshot fetch never
+  // succeeded) has no local data — fall back to a live Vega-Lite URL fetch so it
+  // still renders until a Refresh snapshots it (snapshot model; spec §04 step 1).
+  if (dataset.source === 'url' && dataset.data == null) {
     return {
       ...rest,
-      url: typeof dataset.data === 'string' ? dataset.data : (JSON.stringify(dataset.data) ?? ''),
+      url: typeof dataset.url === 'string' ? dataset.url : '',
       format: { ...restFormat, type: dataset.format },
     };
   }
+
+  // Inline data and fetched URL snapshots resolve identically: the payload is in
+  // `data`, shaped by format.
   switch (dataset.format) {
     case 'json':
       return { ...rest, values: dataset.data };