`FileSystemRepository.close()` never reaches its event broker — a parked `watch()` iterator stays parked after shutdown

[os-zhuang]

Found while implementing #11021 (SysMetadataRepository.close() cannot drain a filtered or numeric-since watcher). Same defect class, different implementation and different package, so it is filed rather than fixed in that PR.

The mechanism

FileSystemRepository.close() (packages/metadata-fs/src/repository.ts:198) does exactly two things:

async close(): Promise<void> {
 this.stopResync();
 if (this.watcher) {
 await this.watcher.close(); // chokidar
 this.watcher = null;
 }
 this.started = false;
}

It never touches this.broker (packages/metadata-fs/src/repository.ts:117, createBroker in packages/metadata-fs/src/sync.ts:49). The broker has no teardown of its own — subscribe/unsubscribe add to and delete from a plain Set, and nothing else empties it.

Each watch() iterator (packages/metadata-fs/src/watch-iterable.ts) parks its pending next() on a waiter callback that only two things can settle: a broker push, or the iterator's own local close() — which runs from iterator.return() / iterator.throw() and from nowhere else. So after repo.close():

• the chokidar source is gone, so no further push can arrive;
• the subscriber is still registered, and nothing calls its terminator.

A consumer holding a for await over fsRepo.watch(...) at shutdown therefore never sees the loop end. Unlike the #11021 shape this is not filter-dependent — there is no drain attempt at all, so it applies to every subscription shape including watch({}).

Confidence

Code read, not a runtime probe. #11021's scope is packages/metadata-protocol, and building/booting metadata-fs to confirm was out of that card's surface. The read is unambiguous (there is no path from close() to the broker or to any iterator terminator), but the matrix that #11021 carries for its own repository has not been produced here. Whoever picks this up should measure it first.

Related contract statement

#11021 adds invariant 8 to the MetadataRepository invariant table in packages/metadata-core/src/repository.ts — "shutdown terminates; it does not emit" — stating what a repository-level close() owes a pending iterator. That row names this gap as the one measured non-conformance among today's implementations. Fixing it is the same one-line shape the #11021 repair uses: hold each subscription's terminator alongside its event sink, and run every terminator from close().

InMemoryRepository is unaffected: it exposes no repository-level shutdown at all, so its iterators end only through return().

---

Generated by Claude Code

[os-zhuang]

Triage: pm:queue, domain:engine, Bug. FileSystemRepository.close() never reaches its event broker, so a parked watch() iterator survives shutdown — lifecycle invariant restore, no contract change.

---

Generated by Claude Code