[realmeylisdev]

Problem

Tests (shard 2/4) keeps timing out on pull requests that have nothing to do with caching. Two tests in test/services/cache_recovery_service_test.dart — "preserves pending uploads while clearing disposable Hive caches" and "preserves durable notification preferences while clearing caches" — each wait out the full 10-minute test timeout, so the shard loses 20 minutes and needs a re-run. #9022 records it on #8540, #8994 and #9018 within four hours, each under a different shuffle seed.

The hang is caused by a different test. CI runs every untagged test file in one merged process, and Hive tracks boxes by name for the whole process. The test navigating to /hashtag/rust leaves the shell entirely in test/router/app_shell_integration_test.dart (re-enabled by #8950) mounts the real HashtagFeedScreen. That screen's provider chain builds HashtagCacheService, whose initialize() calls Hive.openBox('hashtag_stats'). Inside the fake-async zone testWidgets runs in, that open never completes. Hive then keeps hashtag_stats in its private registry of in-flight opens for the rest of the process. Hive.isBoxOpen cannot see that registry, so neither the victim's own cleanup nor the root Hive harness in flutter_test_config.dart can clear it. Any later test that opens hashtag_stats waits on the stranded open until it times out, and both hanging tests open hashtag_stats. Whether the victim runs after the leaker depends on the shuffle seed, which is why only some seeds hang.

Fix

app_shell_integration_test.dart now overrides hashtagServiceProvider with a mock stubbed for the two calls the screen makes when it opens (getVideosByHashtags and subscribeToHashtagVideos). The five hashtag_feed_* screen tests already use this pattern. The test still asserts that the route lands on HashtagFeedScreen outside the shell; it just no longer touches Hive.

The unmount() doc comment cited HashtagService's 60-second periodic timer as the reason the container has to be disposed inside the test body. With the service mocked, no provider timer is left pending in this file, so the comment keeps the general rule and drops that example.

What this leaves alone

• No production code changes. HashtagCacheService opens its box normally in the app; the hang exists only because a widget test started real Hive I/O inside fake async.
• test(cache): cache_recovery_service_test hangs on a pending Hive box open under one shuffle seed #9022's suggested direction is not attempted here: making the harness able to see, and then await or fail, a pending openBox. This PR removes the test behind the current occurrences. The harness gap that lets a future leaker hide the same way stays open.
• The other test files that mention hashtag routes were checked. The five hashtag_feed_* screen tests already override the provider. The router tests that navigate to a hashtag path never mount a widget tree, so none of them reach Hive.

Verification

A two-file bundle reproduces the merged-process order locally. It runs the leaker first and then the victim in one process, calls Hive.init in setUpAll the way an earlier suite does in CI, passes --dart-define=DIVINE_STRICT_CHANNELS=true as CI does, and uses --timeout=120s so the hang stops after two minutes instead of ten:

import 'router/app_shell_integration_test.dart' as leaker;
import 'services/cache_recovery_service_test.dart' as victim;

void main() {
  setUpAll(() {
    Hive.init(Directory.systemTemp.createTempSync('repro_9022').path);
  });
  group('router/app_shell_integration_test.dart', leaker.main);
  group('services/cache_recovery_service_test.dart', victim.main);
}

Run	Before	After
Full bundle	17 passed, 2 failed: both #9022 tests hit TimeoutException after 0:02:00 (4m04s)	19 passed (4s)
Only the /hashtag/rust test, plus the victim file	The same two tests time out	—
Every other test in the leaker file, plus the victim file	18 passed, no hang	—

The two narrowed runs pin the leak to that single test.

On the rebased head:

• app_shell_integration_test.dart (8/8) and cache_recovery_service_test.dart (11/11) pass on their own.
• The logged type 'Null' is not a subtype lines, which come from unrelated mocks, are the same set with the same counts before and after. The new mock swallows no errors of its own.
• A probe copy of the leaker file with c.dispose() removed passed 8/8, which confirms no provider timer is left pending in this file.
• flutter analyze lib test integration_test and dart format are clean.
• These ratchet checks pass: check_shared_setup_stubs, check_placeholder_tests, check_unpinned_unchanged_assertions, check_ungrouped_tests, check_skip_ceiling, check_l10n_delegates_ceiling, check_nostr_id_log_truncation, check_future_delayed_ceiling, check_process_global_mutations, check_shared_channel_overrides and check_http_overrides_isolation.

Closes #9022

Follow-up: #9053 tracks the pending-openBox harness gap left outside this point fix.

[github-actions]

Mobile PR Preview

Preview refreshed for 04995d6

Last refresh: 04995d6 at 2026-09-10 21:17:27 UTC (preview run)

Property	Value
Preview URL	https://b7ec7533.openvine-app.pages.dev
Pages project	openvine-app
Preview branch	pr-9050
PR branch	fix/9022-hashtag-stats-stranded-open
Commit	04995d6

[NotThatKindOfDrLiz]

The point fix is correct and fully verified on 04995d6f125de84916589a7694f210e001fa9a3f. #9053 now tracks the pending-openBox harness gap left outside this PR.