refactor: consolidate block processing logic

Introduce a centralized `block-processor` and `block-scanner` to
handle markdown fenced code blocks. This replaces the previous
`inline-blocks.ts` and individual completion sources with a unified
approach that handles block stripping, directive replacement, and
metadata extraction (stats, templates, and spark tables) in a single
pass.

src/cli/completions/index.ts

@@ -4,10 +4,8 @@
 
 import { diceSource } from "./sources/dice.js";
 import { linksSource } from "./sources/links.js";
-import { sparkTablesSource } from "./sources/spark-tables.js";
-import { statsSource } from "./sources/stats.js";
-import { templatesSource } from "./sources/templates.js";
-import type { CompletionSource, CompletionsPayload } from "./types.js";
+import type { ProcessedBlocks } from "./block-processor.js";
+import type { CompletionsPayload } from "./types.js";
 
 export type {
   CompletionsPayload,
@@ -18,27 +16,22 @@ export type {
   StatTemplate,
 } from "./types.js";
 
-/** Registered sources — open for extension */
-const sources: CompletionSource[] = [
-  diceSource,
-  linksSource,
-  sparkTablesSource,
-  statsSource,
-  templatesSource,
-];
-
 /**
- * Scan the full content index and return structured completion data.
+ * Build completions from the content index and pre-collected blocks.
  * Called at server startup and on any file change.
  */
 export function scanCompletions(
   index: Record<string, string>,
+  blocks: ProcessedBlocks,
 ): CompletionsPayload {
-  const payload: Record<string, unknown[]> = {};
+  const dice = diceSource.scan(index) as CompletionsPayload["dice"];
+  const links = linksSource.scan(index) as CompletionsPayload["links"];
 
-  for (const source of sources) {
-    payload[source.key] = source.scan(index);
-  }
-
-  return payload as unknown as CompletionsPayload;
+  return {
+    dice,
+    links,
+    sparkTables: blocks.sparkTables,
+    stats: blocks.stats,
+    statTemplates: blocks.statTemplates,
+  };
 }