Inline datasets: CSV carried in the dashboard YAML

[dankor]

A chart's dataset: name now resolves two ways — a managed dataset in the store, or an inline one defined in the dashboard's own optional datasets: block:

datasets:
  order_data: |
    id,status,amount
    1,paid,42

The point is prototyping without leaving the editor: invent a table, chart it, iterate, with no upload step and nothing to clean up. A dashboard that carries its own data is also a single file you can hand to someone, and the assistant's prompt now teaches the block, so "mock me up some sales data and chart it" is one turn.

The block holds CSV and nothing else. A path or a header-only line is rejected with a message saying so, because either would otherwise parse as a valid one-column, zero-row table and fail much later complaining about a missing column.

An inline name overrides a managed one: the block is part of the file being rendered, so quietly reading someone else's stored data would be the surprising choice. Anything it doesn't define falls through to the store as before.

Each block converts to Parquet once and is cached by content checksum in the temp dir, so an unchanged block costs a stat, an edit lands on a fresh path, and two dashboards with the same sample data share a file. The write is a rename, which is atomic, so concurrent renders can't read a half-written file.

Also:

• Dashboard.dataset_names() reports only managed names, so a chart on inline data doesn't pin a stored dataset against the portal's delete-guard.
• The edit modal's column dropdowns resolve inline datasets, which would otherwise be empty for such a chart.
• The starter dashboard carries its own data. files/orders.parquet is gone (a derived binary in git), the orders dataset is no longer seeded at startup in either mode, and files/inline_orders.py regenerates the block from files/orders.csv so the two can't drift.
• The route tests carry their own data too. They had leaned on that seeded dataset, which meant they were passing off a leftover file on the developer's disk and would have failed on a clean checkout; the suite now runs with no dataset store present at all.

523 tests pass.