refactor: remove as= override and render data-spark at runtime

- Drop resolveBlockAs and the as= attribute; scanDoc now switches
  directly on role and warns on unknown roles instead of silently
  stripping
- md-table derives data-spark from its loaded CSV via
  parseSparkTableCsv; delete injectSparkDirectives and the second
  registry pass so pathIndex is exactly scanDoc output
- Align spark-table dice header regex with isSparkTableHeader

src/markdown/block-attrs.ts

@@ -10,7 +10,6 @@ export interface BlockAttrs {
   lang: string;
   id?: string;
   role?: string;
-  as?: string;
   /** Any other attributes not in the standard set */
   extra: Record<string, string>;
 }
@@ -24,6 +23,8 @@ export function parseBlockAttrs(info: string): BlockAttrs {
     attrs[m[1]] = m[2].replace(/^"|"$/g, "");
   }
 
-  const { lang, id, role, as, ...extra } = attrs;
-  return { lang: lang || "", id, role, as, extra };
+  // `as` was a render-target override; roles now fully determine behavior,
+  // so it is parsed out and ignored (kept out of `extra` for directives).
+  const { lang, id, role, as: _legacyAs, ...extra } = attrs;
+  return { lang: lang || "", id, role, extra };
 }