perf(kernel): direct-to-store decode — buffers flow to the store worker, main thread never materializes nodes

Kernel-routed files ship their flat tables from the parse worker to the
store worker as buffers (tryKernelExtractRaw → kernelBuffers on the
result → KernelStoreBundle); the store worker decodes and finalizes
(finalizeStoreBundle shared with the object path so filter semantics
can never drift). Files with applicable framework extract() hooks keep
the decoded path; the no-writer fallback materializes via
materializeKernelResult. Byte-identical dumps re-verified on dubbo,
excalidraw, express, gson; full suite green (2,467).

Measurement (plan §4d): dubbo's parse-loop wall is 94% store-writer
busy time — the many-core fresh-index wall is single-writer SQLite
ingest, not extraction or main-thread work. d2s improves the writer
lane ~11% (structured-clone deserialization avoided on the writer) and
frees the main thread; the remaining many-core gap is a
store-architecture arc (deferred index builds, multi-file
transactions, buffer→bind), out of the kernel project's scope.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

docs/design/rust-kernel-migration-plan.md

@@ -264,6 +264,31 @@ Default routing: `DEFAULT_ROUTED = {typescript, tsx, javascript, jsx}` in
   Windows VM still deferred (same fallback rationale as §4b).
 - Default routing now includes `java`.
 
+### 4d. Direct-to-store decode (2026-07-16) — and where the wall ACTUALLY is
+
+Kernel-routed files now ship their flat buffers from the parse worker all the way
+to the STORE WORKER, which decodes + finalizes them there (`tryKernelExtractRaw` →
+`ExtractionResult.kernelBuffers` → `KernelStoreBundle` → `decodeKernelBundle`;
+filter semantics shared via `finalizeStoreBundle`). The main thread's per-file work
+drops to O(1) + the content hash — it never materializes per-node objects, and both
+postMessage hops move flat bytes instead of object graphs. Files whose applicable
+frameworks carry an `extract()` hook keep the decoded path (hooks merge into decoded
+results); non-writer paths (main-thread store, tests) materialize via
+`materializeKernelResult`. Byte-identical dumps re-verified on dubbo, excalidraw,
+express, gson.
+
+**Measurement that closes the §4c question:** with the store worker instrumented,
+dubbo's parse-loop wall is **94% store-writer busy time** (4,202ms of 4,493ms on the
+kernel arm). The many-core fresh-index wall is the single-writer SQLite ingest —
+not extraction, not main-thread work. d2s still improves the writer lane ~11%
+(4,726→4,202ms: buffers skip structured-clone deserialization ON the writer) and
+frees the main thread, but the remaining cbm gap on many-core medium repos is a
+STORE-ARCHITECTURE question (their RAM-first design defers all durability). Next
+levers there (a separate perf arc, not this project): deferred/bulk index builds
+during the parse phase, multi-file write transactions, buffer→bind without object
+materialization. Note the #1320-arc post-mortem already measured statement batching
+and sorted inserts as ~zero on this path — B-tree maintenance is the floor.
+
 ## 4. Per-language tracker
 
 Tiers: **T1** = mostly `.scm` + mapping config. **T2** = needs bespoke pre/post passes kept

src/extraction/index.ts

@@ -23,7 +23,8 @@ import {
 import { QueryBuilder } from '../db/queries';
 import { extractFromSource } from './tree-sitter';
 import { ParseWorkerPool, resolveParsePoolSize, resolveParseTimeoutMs } from './parse-pool';
-import { StoreWriter, StoreBundle } from './store-writer';
+import { StoreWriter, StoreBundle, finalizeStoreBundle } from './store-writer';
+import { materializeKernelResult } from './kernel';
 import { detectLanguage, isSourceFile, isLanguageSupported, isFileLevelOnlyLanguage, initGrammars, loadGrammarsForLanguages, readGrammarWasmBytes } from './grammars';
 import { loadExtensionOverrides, loadIncludeIgnoredPatterns, loadExcludePatterns, loadIncludePatterns } from '../project-config';
 import { isCodeGraphDataDir } from '../directory';
@@ -1706,16 +1707,34 @@ export class ExtractionOrchestrator {
       const bp = walBackpressure?.();
       if (bp) await bp;
 
+      // Kernel deferred-decode results carry table sizes in kernelCounts
+      // (their object arrays are empty — decode happens at the store).
+      const nodeCount = result.kernelCounts?.nodes ?? result.nodes.length;
+      const edgeCount = result.kernelCounts?.edges ?? result.edges.length;
+
       // Store: on the writer thread when active (fresh DB — bundles applied
       // in the same file order this chain dispatches them), else on the main
       // thread (SQLite connections are per-thread).
-      if (result.nodes.length > 0 || result.errors.length === 0) {
+      if (nodeCount > 0 || result.errors.length === 0) {
         const language = detectLanguage(filePath, content, overrides);
         if (storeWriter) {
+          if (result.kernelBuffers) {
+            // Buffers go to the writer as-is; the worker decodes + finalizes.
+            // The main thread's only per-file work stays O(1) + the content hash.
+            storeWriter.send({
+              kernel: true,
+              filePath,
+              language,
+              buffers: result.kernelBuffers,
+              file: this.buildFileRecord(filePath, content, language, stats, nodeCount, result.errors),
+            });
+          } else {
             storeWriter.send(this.buildFreshStoreBundle(filePath, content, language, stats, result));
+          }
           await storeWriter.waitBelow(STORE_WRITER_WINDOW);
         } else {
-          await this.storeExtractionResult(filePath, content, language, stats, result, commitYield);
+          const materialized = materializeKernelResult(result, filePath, language);
+          await this.storeExtractionResult(filePath, content, language, stats, materialized, commitYield);
         }
       }
 
@@ -1726,10 +1745,10 @@ export class ExtractionOrchestrator {
         errors.push(...result.errors);
       }
 
-      if (result.nodes.length > 0) {
+      if (nodeCount > 0) {
         filesIndexed++;
-        totalNodes += result.nodes.length;
+        totalNodes += nodeCount;
-        totalEdges += result.edges.length;
+        totalEdges += edgeCount;
       } else if (result.errors.some((e) => e.severity === 'error')) {
         filesErrored++;
       } else {
@@ -2383,6 +2402,27 @@ export class ExtractionOrchestrator {
    * check, no cross-file edge snapshot (both are re-index concerns — a fresh
    * database has neither). Filters mirror storeExtractionResult exactly.
    */
+  /** The FileRecord for a fresh-index store (nodeCount is the PRE-filter count). */
+  private buildFileRecord(
+    filePath: string,
+    content: string,
+    language: Language,
+    stats: fs.Stats,
+    nodeCount: number,
+    resultErrors: ExtractionResult['errors']
+  ): FileRecord {
+    return {
+      path: filePath,
+      contentHash: hashContent(content),
+      language,
+      size: stats.size,
+      modifiedAt: stats.mtimeMs,
+      indexedAt: Date.now(),
+      nodeCount,
+      errors: resultErrors.length > 0 ? resultErrors : undefined,
+    };
+  }
+
   private buildFreshStoreBundle(
     filePath: string,
     content: string,
@@ -2390,33 +2430,12 @@ export class ExtractionOrchestrator {
     stats: fs.Stats,
     result: ExtractionResult
   ): StoreBundle {
-    const validNodes = result.nodes.filter((n) => n.id && n.kind && n.name && n.filePath && n.language);
+    return finalizeStoreBundle(
-    const insertedIds = new Set(validNodes.map((n) => n.id));
+      result,
-    const validEdges = result.edges.filter(
+      filePath,
-      (e) => insertedIds.has(e.source) && insertedIds.has(e.target)
-    );
-    const validRefs = result.unresolvedReferences
-      .filter((ref) => insertedIds.has(ref.fromNodeId))
-      .map((ref) => ({
-        ...ref,
-        filePath: ref.filePath ?? filePath,
-        language: ref.language ?? language,
-      }));
-    return {
-      nodes: validNodes,
-      edges: validEdges,
-      refs: validRefs,
-      file: {
-        path: filePath,
-        contentHash: hashContent(content),
       language,
-        size: stats.size,
+      this.buildFileRecord(filePath, content, language, stats, result.nodes.length, result.errors)
-        modifiedAt: stats.mtimeMs,
+    );
-        indexedAt: Date.now(),
-        nodeCount: result.nodes.length,
-        errors: result.errors.length > 0 ? result.errors : undefined,
-      },
-    };
   }
 
   /**

src/extraction/kernel/index.ts

@@ -17,6 +17,11 @@
 import type { ExtractionResult, Language } from '../../types';
 import { getKernel, kernelSupports } from './loader';
 import { decodeExtractBuffers } from './decode';
+import {
+  KERNEL_ABI_VERSION as LAYOUT_ABI,
+  META as LAYOUT_META,
+  NONE as LAYOUT_NONE,
+} from './layout';
 
 export { getKernel, kernelSupports, resetKernelForTests } from './loader';
 export { decodeExtractBuffers } from './decode';
@@ -66,6 +71,84 @@ export function kernelRoutes(language: Language): boolean {
 /** Warned-once registry so a broken language logs a single line, not one per file. */
 const warned = new Set<string>();
 
+/** The raw table buffers + the cheap facts the orchestrator needs pre-decode. */
+export interface KernelRawResult {
+  buffers: NonNullable<ExtractionResult['kernelBuffers']>;
+  counts: { nodes: number; edges: number; refs: number };
+  errors: ExtractionResult['errors'];
+}
+
+/**
+ * Extract via the kernel WITHOUT decoding — the bulk-index fast path. The
+ * tables ride to the store boundary as buffers (decoded on the store worker),
+ * so the main thread never materializes per-node objects. Returns null under
+ * exactly the conditions tryKernelExtract does, PLUS when the language has a
+ * registered post() pass (post passes operate on decoded results, so those
+ * languages keep the decoded path).
+ */
+export function tryKernelExtractRaw(
+  filePath: string,
+  source: string,
+  language: Language
+): KernelRawResult | null {
+  if (!kernelRoutes(language) || POST_PASSES[language]) return null;
+  const kernel = getKernel();
+  if (!kernel) return null;
+  try {
+    const buffers = kernel.extractFile(filePath, source, language);
+    const meta = buffers.meta;
+    if (meta.readUInt8(LAYOUT_META.version) !== LAYOUT_ABI) {
+      throw new Error(`kernel buffer ABI ${meta.readUInt8(0)} != expected ${LAYOUT_ABI}`);
+    }
+    const counts = {
+      nodes: meta.readUInt32LE(LAYOUT_META.nodeCount),
+      edges: meta.readUInt32LE(LAYOUT_META.edgeCount),
+      refs: meta.readUInt32LE(LAYOUT_META.refCount),
+    };
+    let errors: ExtractionResult['errors'] = [];
+    const errorsOff = meta.readUInt32LE(LAYOUT_META.errorsOff);
+    if (errorsOff !== LAYOUT_NONE) {
+      const errorsLen = meta.readUInt32LE(LAYOUT_META.errorsLen);
+      errors = JSON.parse(
+        buffers.arena.toString('utf8', errorsOff, errorsOff + errorsLen)
+      ) as ExtractionResult['errors'];
+    }
+    return { buffers, counts, errors };
+  } catch (err) {
+    const message = err instanceof Error ? err.message : String(err);
+    if (message.includes('defer:')) return null;
+    if (!warned.has(language)) {
+      warned.add(language);
+      process.stderr.write(
+        `[codegraph-kernel] ${language} extraction failed (${message}) — falling back to the wasm path\n`
+      );
+    }
+    return null;
+  }
+}
+
+/**
+ * Decode a buffer-carrying result (see ExtractionResult.kernelBuffers) into a
+ * plain, fully-materialized ExtractionResult — the fallback for store paths
+ * that need objects (main-thread store, tests).
+ */
+export function materializeKernelResult(
+  result: ExtractionResult,
+  filePath: string,
+  language: Language
+): ExtractionResult {
+  if (!result.kernelBuffers) return result;
+  const b = result.kernelBuffers;
+  const asBuf = (u: Uint8Array) => Buffer.from(u.buffer, u.byteOffset, u.byteLength);
+  const decoded = decodeExtractBuffers(
+    { meta: asBuf(b.meta), nodes: asBuf(b.nodes), edges: asBuf(b.edges), refs: asBuf(b.refs), arena: asBuf(b.arena) },
+    filePath,
+    language
+  );
+  decoded.durationMs = result.durationMs;
+  return decoded;
+}
+
 /**
  * Extract via the native kernel. Returns null when the kernel doesn't apply
  * (not routed / not available / kill switch) — the caller falls back to the

src/extraction/parse-worker.ts

@@ -16,6 +16,8 @@ try {
 import { parentPort } from 'worker_threads';
 import { extractFromSource } from './tree-sitter';
 import { detectLanguage, loadGrammarsForLanguages, resetParser } from './grammars';
+import { tryKernelExtractRaw } from './kernel';
+import { getAllFrameworkResolvers, getApplicableFrameworks } from '../resolution/frameworks';
 import type { Language, ExtractionResult } from '../types';
 
 // Emscripten prints `Aborted()` (and a follow-up RuntimeError diag
@@ -80,7 +82,36 @@ parentPort!.on('message', async (msg: { type: string; id?: number; filePath?: st
       // codegraph.json extension overrides) and sends it; fall back to detection
       // for older callers / safety.
       const language = msg.language ?? detectLanguage(filePath!, content);
-      const result: ExtractionResult = extractFromSource(filePath!, content!, language, frameworkNames);
+
+      // Kernel deferred-decode fast path: ship the file's tables as flat
+      // buffers and decode at the STORE boundary, so the main thread never
+      // materializes per-node objects (nor pays their structured-clone cost —
+      // buffer clone is a flat memcpy). Only when no applicable framework has
+      // an extract() hook: those merge extra nodes/refs into the DECODED
+      // result inside extractFromSource, so such files keep the decoded path.
+      let result: ExtractionResult | undefined;
+      const frameworksNeedDecode =
+        frameworkNames && frameworkNames.length > 0
+          ? getApplicableFrameworks(
+              getAllFrameworkResolvers().filter((r) => frameworkNames.includes(r.name)),
+              language
+            ).some((fw) => !!fw.extract)
+          : false;
+      if (!frameworksNeedDecode) {
+        const raw = tryKernelExtractRaw(filePath!, content!, language);
+        if (raw) {
+          result = {
+            nodes: [],
+            edges: [],
+            unresolvedReferences: [],
+            errors: raw.errors,
+            durationMs: 0,
+            kernelBuffers: raw.buffers,
+            kernelCounts: raw.counts,
+          };
+        }
+      }
+      result ??= extractFromSource(filePath!, content!, language, frameworkNames);
 
       // Periodic parser reset to reclaim WASM heap memory
       const count = (parseCounts.get(language) ?? 0) + 1;

src/extraction/store-worker.ts

@@ -30,7 +30,8 @@ try {
 import { parentPort } from 'worker_threads';
 import { QueryBuilder } from '../db/queries';
 import { createDatabase, SqliteDatabase } from '../db/sqlite-adapter';
-import type { StoreBundle } from './store-writer';
+import { finalizeStoreBundle, type KernelStoreBundle, type StoreBundle } from './store-writer';
+import { decodeExtractBuffers } from './kernel/decode';
 
 if (!parentPort) {
   throw new Error('store-worker must be run as a worker thread');
@@ -42,10 +43,27 @@ let queries: QueryBuilder | null = null;
 
 type InMessage =
   | { type: 'open'; dbPath: string; fastInit: boolean }
-  | { type: 'bundle'; bundle: StoreBundle }
+  | { type: 'bundle'; bundle: StoreBundle | KernelStoreBundle }
   | { type: 'drain'; id: number }
   | { type: 'close' };
 
+/** Decode a kernel bundle's buffers into the standard pre-filtered StoreBundle. */
+function decodeKernelBundle(bundle: KernelStoreBundle): StoreBundle {
+  const asBuf = (u: Uint8Array) => Buffer.from(u.buffer, u.byteOffset, u.byteLength);
+  const decoded = decodeExtractBuffers(
+    {
+      meta: asBuf(bundle.buffers.meta),
+      nodes: asBuf(bundle.buffers.nodes),
+      edges: asBuf(bundle.buffers.edges),
+      refs: asBuf(bundle.buffers.refs),
+      arena: asBuf(bundle.buffers.arena),
+    },
+    bundle.filePath,
+    bundle.language
+  );
+  return finalizeStoreBundle(decoded, bundle.filePath, bundle.language, bundle.file);
+}
+
 port.on('message', (msg: InMessage) => {
   try {
     switch (msg.type) {
@@ -70,7 +88,8 @@ port.on('message', (msg: InMessage) => {
       }
       case 'bundle': {
         if (!queries) throw new Error('store-worker: bundle before open');
-        queries.storeFileBundle(msg.bundle);
+        const bundle = 'kernel' in msg.bundle ? decodeKernelBundle(msg.bundle) : msg.bundle;
+        queries.storeFileBundle(bundle);
         port.postMessage({ type: 'ack' });
         break;
       }

src/extraction/store-writer.ts

@@ -8,7 +8,7 @@
  */
 
 import { Worker } from 'worker_threads';
-import { Node, Edge, UnresolvedReference, FileRecord } from '../types';
+import { ExtractionResult, Language, Node, Edge, UnresolvedReference, FileRecord } from '../types';
 
 /** One file's complete store payload (pre-filtered — see storeFileBundle). */
 export interface StoreBundle {
@@ -18,6 +18,50 @@ export interface StoreBundle {
   file: FileRecord;
 }
 
+/**
+ * A kernel deferred-decode payload: the file's raw table buffers plus the
+ * FileRecord the main thread built from meta counts. The store WORKER decodes
+ * and finalizes (same filters as the object path), so per-node objects never
+ * exist on the main thread.
+ */
+export interface KernelStoreBundle {
+  kernel: true;
+  filePath: string;
+  language: Language;
+  buffers: NonNullable<ExtractionResult['kernelBuffers']>;
+  file: FileRecord;
+}
+
+/**
+ * The validation/denormalization every bundle gets before storeFileBundle —
+ * shared by the orchestrator's object path and the store worker's kernel
+ * decode path so the two can never drift:
+ *   - nodes missing identity fields are dropped (#42-class safety),
+ *   - edges must connect inserted nodes (FK integrity),
+ *   - refs must originate from inserted nodes and carry the denormalized
+ *     filePath/language the resolver reads.
+ */
+export function finalizeStoreBundle(
+  result: Pick<ExtractionResult, 'nodes' | 'edges' | 'unresolvedReferences'>,
+  filePath: string,
+  language: Language,
+  file: FileRecord
+): StoreBundle {
+  const validNodes = result.nodes.filter((n) => n.id && n.kind && n.name && n.filePath && n.language);
+  const insertedIds = new Set(validNodes.map((n) => n.id));
+  const validEdges = result.edges.filter(
+    (e) => insertedIds.has(e.source) && insertedIds.has(e.target)
+  );
+  const validRefs = result.unresolvedReferences
+    .filter((ref) => insertedIds.has(ref.fromNodeId))
+    .map((ref) => ({
+      ...ref,
+      filePath: ref.filePath ?? filePath,
+      language: ref.language ?? language,
+    }));
+  return { nodes: validNodes, edges: validEdges, refs: validRefs, file };
+}
+
 export class StoreWriter {
   private worker: Worker;
   private readyPromise: Promise<void>;
@@ -102,7 +146,7 @@ export class StoreWriter {
   }
 
   /** Post one file's bundle. Throws immediately if the writer already failed. */
-  send(bundle: StoreBundle): void {
+  send(bundle: StoreBundle | KernelStoreBundle): void {
     if (this.firstError) throw this.firstError;
     if (this.exited) throw new Error('store worker already exited');
     this.outstanding++;

src/types.ts

@@ -276,6 +276,23 @@ export interface ExtractionResult {
 
   /** Extraction duration in milliseconds */
   durationMs: number;
+
+  /**
+   * Deferred-decode transport (native kernel, bulk-index path): when present,
+   * `nodes`/`edges`/`unresolvedReferences` are EMPTY and the file's tables
+   * ride as flat buffers to be decoded at the store boundary (the store
+   * worker), so the MAIN thread never materializes per-node objects.
+   * `kernelCounts` carries the table sizes for bookkeeping. Decode into a
+   * plain result with `materializeKernelResult` (src/extraction/kernel).
+   */
+  kernelBuffers?: {
+    meta: Uint8Array;
+    nodes: Uint8Array;
+    edges: Uint8Array;
+    refs: Uint8Array;
+    arena: Uint8Array;
+  };
+  kernelCounts?: { nodes: number; edges: number; refs: number };
 }
 
 /**