fix(bench): dispose WASM worker pool and keep progress off stdout (#1009)

* fix(bench): dispose WASM worker pool and keep progress off stdout

v3.9.5 moved WASM tree-sitter parsing into a node:worker_threads Worker
(via a shared singleton pool in src/domain/wasm-worker-pool.ts). The
benchmark worker scripts never dispose that pool, so the worker thread
kept the event loop alive after JSON was written to stdout. The parent
fork-engine then SIGKILLed the hung child at 600s and discarded the
valid stdout as "Both engines failed."

Also fix the embedding benchmark: src/domain/search/models.ts wrote
"Embedded N/M\r" progress directly to process.stdout, bypassing the
worker's console.log->stderr redirect and corrupting the JSON payload
with "Unexpected token 'E'..." errors.

- benchmark/query/incremental/resolution scripts: import disposeParsers
  from the installed package (try/catch for older releases) and call it
  before cleanup(), then process.exit(0) to guarantee termination.
- fork-engine: on SIGKILL, try to salvage captured stdout as JSON before
  giving up — defence-in-depth against future event-loop-keepalive bugs.
- search/models.ts: write progress to stderr so it can't pollute stdout
  in JSON-consuming benchmark workers.

docs check acknowledged — internal benchmark infra fix, no user-facing
docs affected.

Impact: 7 functions changed, 16 affected

* fix(bench): move process.exit(0) outside finally to preserve exception exit codes (#1009)

Addresses Greptile P1: placing process.exit(0) inside the finally block
would swallow exceptions from the try body (missing fixture, DB open failure,
import error) and exit with code 0, masking the failure from CI and the
parent forkWorker. Mirror the pattern in benchmark.ts/incremental-benchmark.ts/
query-benchmark.ts: exit only after the success path completes.

Author: carlos-alm

scripts/benchmark.ts

@@ -82,6 +82,13 @@ const { buildGraph } = await import(srcImport(srcDir, 'domain/graph/builder.js')
 const { fnDepsData, fnImpactData, pathData, rolesData, statsData } = await import(
 	srcImport(srcDir, 'domain/queries.js')
 );
+// v3.9.5+ parses WASM in a worker_thread that keeps the event loop alive until
+// disposed. Older releases don't export disposeParsers — fall back to a no-op.
+let disposeParsers = async () => {};
+try {
+	const parser = await import(srcImport(srcDir, 'domain/parser.js'));
+	if (typeof parser.disposeParsers === 'function') disposeParsers = parser.disposeParsers;
+} catch { /* older release — no worker pool to dispose */ }
 
 const INCREMENTAL_RUNS = 3;
 const QUERY_RUNS = 5;
@@ -226,4 +233,6 @@ const workerResult = {
 
 console.log(JSON.stringify(workerResult));
 
+await disposeParsers();
 cleanup();
+process.exit(0);

scripts/incremental-benchmark.ts

@@ -143,6 +143,13 @@ const { srcDir, cleanup } = await resolveBenchmarkSource();
 const dbPath = path.join(root, '.codegraph', 'graph.db');
 
 const { buildGraph } = await import(srcImport(srcDir, 'domain/graph/builder.js'));
+// v3.9.5+ parses WASM in a worker_thread that keeps the event loop alive until
+// disposed. Older releases don't export disposeParsers — fall back to a no-op.
+let disposeParsers = async () => {};
+try {
+	const parser = await import(srcImport(srcDir, 'domain/parser.js'));
+	if (typeof parser.disposeParsers === 'function') disposeParsers = parser.disposeParsers;
+} catch { /* older release — no worker pool to dispose */ }
 
 // Redirect console.log to stderr so only JSON goes to stdout
 const origLog = console.log;
@@ -218,4 +225,6 @@ console.log = origLog;
 const workerResult = { fullBuildMs, noopRebuildMs, oneFileRebuildMs, oneFilePhases };
 console.log(JSON.stringify(workerResult));
 
+await disposeParsers();
 cleanup();
+process.exit(0);

scripts/lib/fork-engine.ts

@@ -93,7 +93,18 @@ export function forkWorker(scriptPath, envKey, workerName, argv = [], timeoutMs
 
 			if (signal) {
 				console.error(`[fork] ${workerName} worker killed by signal ${signal}`);
-				settle(null);
+				// A worker can finish its measurements and write valid JSON to
+				// stdout before hanging on a non-daemonized handle (e.g. an
+				// un-disposed worker_thread pool). SIGKILL from our timeout
+				// then discards the result. Try to parse captured stdout first
+				// so we don't lose real data to a lingering event-loop keepalive.
+				try {
+					const parsed = JSON.parse(stdout);
+					console.error(`[fork] ${workerName} worker produced results before being killed — salvaging`);
+					settle(parsed);
+				} catch {
+					settle(null);
+				}
 				return;
 			}
 

scripts/query-benchmark.ts

@@ -94,6 +94,13 @@ const { buildGraph } = await import(srcImport(srcDir, 'domain/graph/builder.js')
 const { fnDepsData, fnImpactData, diffImpactData } = await import(
 	srcImport(srcDir, 'domain/queries.js')
 );
+// v3.9.5+ parses WASM in a worker_thread that keeps the event loop alive until
+// disposed. Older releases don't export disposeParsers — fall back to a no-op.
+let disposeParsers = async () => {};
+try {
+	const parser = await import(srcImport(srcDir, 'domain/parser.js'));
+	if (typeof parser.disposeParsers === 'function') disposeParsers = parser.disposeParsers;
+} catch { /* older release — no worker pool to dispose */ }
 
 // Redirect console.log to stderr so only JSON goes to stdout
 const origLog = console.log;
@@ -259,4 +266,6 @@ console.log = origLog;
 const workerResult = { targets, fnDeps, fnImpact, diffImpact };
 console.log(JSON.stringify(workerResult));
 
+await disposeParsers();
 cleanup();
+process.exit(0);

scripts/resolution-benchmark.ts

@@ -219,6 +219,14 @@ console.log = (...args) => console.error(...args);
 
 const { srcDir, cleanup } = await resolveBenchmarkSource();
 
+// v3.9.5+ parses WASM in a worker_thread that keeps the event loop alive until
+// disposed. Older releases don't export disposeParsers — fall back to a no-op.
+let disposeParsers = async () => {};
+try {
+	const parser = await import(srcImport(srcDir, 'domain/parser.js'));
+	if (typeof parser.disposeParsers === 'function') disposeParsers = parser.disposeParsers;
+} catch { /* older release — no worker pool to dispose */ }
+
 try {
 	const { buildGraph } = await import(srcImport(srcDir, 'domain/graph/builder.js'));
 	const { openReadonlyOrFail } = await import(srcImport(srcDir, 'db/index.js'));
@@ -296,5 +304,7 @@ try {
 	console.log(JSON.stringify(results, null, 2));
 } finally {
 	console.log = origLog;
+	await disposeParsers();
 	cleanup();
 }
+process.exit(0);

src/domain/search/models.ts

@@ -253,7 +253,7 @@ export async function embed(
     }
 
     if (texts.length > batchSize) {
-      process.stdout.write(`  Embedded ${Math.min(i + batchSize, texts.length)}/${texts.length}\r`);
+      process.stderr.write(`  Embedded ${Math.min(i + batchSize, texts.length)}/${texts.length}\r`);
     }
   }