feat(core): v2 union-find with overlap window and deferred summarization

Architectural changes:
- Forest.union() is now synchronous — structural merge only, no LLM calls
- ContextWindow.append() is now synchronous — no blocking
- ContextWindow.render() returns cached summaries (synchronous)
- New resolveDirty() method — async fire-and-forget batch summarization
- Overlap window (graduateAt/evictAt) eliminates staleness

Key improvements over v1:
- O(n) cost instead of O(n²) — each message summarized at most once
- Zero blocking on append or render
- Background resolveDirty runs during main LLM call wait

89/89 tests pass, 0 type errors.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Author: kimjune01

packages/core/src/context/chatCompressionService.ts

@@ -162,13 +162,11 @@ async function truncateHistoryToBudget(
           } else if (responseObj && typeof responseObj === 'object') {
             if (
               'output' in responseObj &&
-               
               typeof responseObj['output'] === 'string'
             ) {
               contentStr = responseObj['output'];
             } else if (
               'content' in responseObj &&
-               
               typeof responseObj['content'] === 'string'
             ) {
               contentStr = responseObj['content'];
@@ -237,8 +235,11 @@ async function truncateHistoryToBudget(
   return truncatedHistory;
 }
 
-/** Number of messages kept verbatim in the hot zone. */
-const UNION_FIND_HOT_SIZE = 30;
+/** Start graduating messages when hot zone exceeds this count. */
+const UNION_FIND_GRADUATE_AT = 26;
+
+/** Evict oldest from hot zone when it exceeds this count. */
+const UNION_FIND_EVICT_AT = 30;
 
 /** Maximum number of clusters in the cold zone. */
 const UNION_FIND_MAX_COLD_CLUSTERS = 10;
@@ -610,7 +611,8 @@ export class ChatCompressionService {
       modelStringToModelConfigAlias(model),
     );
     const contextWindow = new ContextWindow(embedder, summarizer, {
-      hotSize: UNION_FIND_HOT_SIZE,
+      graduateAt: UNION_FIND_GRADUATE_AT,
+      evictAt: UNION_FIND_EVICT_AT,
       maxColdClusters: UNION_FIND_MAX_COLD_CLUSTERS,
       mergeThreshold: UNION_FIND_MERGE_THRESHOLD,
     });
@@ -628,17 +630,22 @@ export class ChatCompressionService {
         .join(' ')
         .trim();
       if (text) {
-        await contextWindow.append(text);
+        contextWindow.append(text);
       }
     }
 
-    // Render the compacted context
+    // Render the compacted context (synchronous — uses cached summaries)
     const rendered = contextWindow.render(
       null,
       UNION_FIND_RETRIEVE_K,
       UNION_FIND_RETRIEVE_MIN_SIM,
     );
 
+    // Fire-and-forget: resolve dirty clusters in background
+    // In production, this runs during the main LLM call wait.
+    // Here we await it since there's no concurrent main call.
+    await contextWindow.resolveDirty();
+
     // Build new history: cold summaries as a single user message, then hot messages
     const coldSummaries = rendered.slice(
       0,