refactor: remove implicit content sniffing and table conversion

- Split loadCSV into parseCSVString (content) and loadCSVFromPath
  (path); drop isCSV/looksLikeCsv heuristics
- Delete coerceSparkTables and markedTable label-header magic;
  plain markdown tables now render as plain tables
- Add explicit markdown role=spark-table fence syntax that converts
  pipe tables to CSV at scan time, with dice-header validation
- Map ESM-only github-slugger and csv-parse browser build to CJS in
  jest config; add content-registry tests

src/components/md-table.tsx

@@ -8,7 +8,7 @@ import {
   createResource,
 } from "solid-js";
 import { parseMarkdown } from "../markdown";
-import { loadCSV, CSV, processVariables } from "./utils/csv-loader";
+import { parseCSVString, CSV, processVariables } from "./utils/csv-loader";
 import { resolveContentRef } from "./utils/resolve-content";
 import {
   areAllLabelsNumeric,
@@ -59,7 +59,7 @@ customElement(
         if (content === null) {
           throw new Error(`Failed to resolve table content: "${ref}"`);
         }
-        return loadCSV(content);
+        return parseCSVString(content);
       },
     );