Automated-test documents no longer run as written

[DutchJaFO]

Description

Five automated-test documents no longer run as written, found while running them for #370's T2 pass on 2026-09-18 and #409's on 2026-09-19:

• A file left awaiting review raises an alert, and resolving it retires the alert (import-and-staged-actions/20-pending-review-alert.md) — step 6 expects three alerts, one of them obsolete, after two reseeds. Since Reseed should only import the designated files, not delete data first #372 a reseed no longer truncates a batch, and since A Source-level Modify conflict stages a new Pending action on every reseed instead of staying stable #376 an already-reported conflict is not re-staged, so no alert can become obsolete: the run shows two alerts, one active and one resolved. Step 8's "No longer applicable" half is unreachable for the same reason. Step 7 also destroys qt-review-20 before steps 8 and 9 drive it in the browser, so the document cannot run in its written order.
• An already-reported conflict does not stage a duplicate on every reseed (import-and-staged-actions/25-an-already-reported-conflict-does-not-accumulate.md) — steps 4 and 5 print empty counts instead of 1. Its Unresolved function returns .Count on a filtered result that is a single object, which PowerShell 5.1 unrolls, so .Count is empty — the pitfall the pending-review alert document itself warns about. The behaviour under test is correct: step 6 reads AlreadyReported=2 and Modify=1.
• Bulk-deciding a staged batch via file export and re-import, in both wire formats (import-and-staged-actions/13-bulk-decide-via-file-export-import.md) — step 2 imports the curated file under review and expects 202. Since An import that re-states identical content reports it as modified #373 an already-stored file stages nothing, so it answers 200 and the document never reaches the bulk-decide round trip it exists for — the reason The staged review → decide → apply workflow moved to the conflict fixture.
• The changelog is served from its own on-disk database, not the JSON fallback (notifications-and-changelog/07-changelog-served-from-its-own-database.md) — step 5 reads the log as soon as health answers after a restart, but the changelog import runs after that: its refreshed line appeared 716 ms later. The step then reports no import.
• Reset wipes the entire database and does not reseed (database-lifecycle/03-reset-is-a-full-wipe.md) — the cleanup removes smoke156-before.db but not the -wal/-shm files DbInspector creates beside it when step 7 reads it.

Running them for this issue on 2026-09-19 found four more ways a run reports something the test did not cause:

• The pending-review alert document writes into its bind folder without clearing it. A folder an earlier run left behind still holds that run's database, so step 6 counted four alerts.
• The already-reported conflict document reseeds straight after docker restart, before the application answers; the reseed fails with the connection closed.
• Every browser-driven document logs a CryptographicException / AntiforgeryValidationException pair it did not cause. Each test container starts with its own DataProtection key ring, and the browser pane keeps a cookie from an earlier container — it does not separate cookies by port, nor by *.localhost host name. Measured: 2 to 7 such lines on a first visit, and 0 when both containers read one shared key ring.
• Exception counts included the application's own shutdown. Stopping the application logs a SocketException per listening port, and an OperationCanceledException per connected page; a log read after a stop counts them with the test's own.

Reproduction steps

Run each document exactly as written against a current quotinator:local.

Expected behaviour

1. The pending-review alert document's step 6 asserts what a reseed now produces, and step 8 reaches both dismiss reasons through a path that still exists; if none does, the obsolete half moves to a test that constructs the state. Steps run in their written order.
2. The already-reported conflict document's steps 4 and 5 print 1 -> 1 -> 1.
3. The bulk-decide document stages a batch that has something to decide, and runs its round trip.
4. The changelog database document's step 5 reads the import line once it has been written.
5. The reset document leaves nothing behind in .claude/temp.
6. The pending-review alert document runs against a fresh database, whatever an earlier run left behind.
7. The already-reported conflict document reseeds only once the restarted application answers.
8. A test logs the antiforgery pair only when it sets out to provoke it. Test containers share one key ring by default; a test can opt out, and a read-only data directory never shares it. A test that provokes the pair ends by restoring a cookie the shared ring can read, and proves the next visit logs nothing.
9. A container's log is read for exceptions before the application stops, except in a test about shutdown.

Actual behaviour

• Pending-review alert, step 6: cc95b26e isDismissed=False, 9818586e isDismissed=True reason=resolved, active alerts = 1 — two alerts, none obsolete.
• Already-reported conflict, steps 4–5: after cold start: and -> ->.
• Bulk decide, step 2: < 200 OK, Expected 202, got 200.
• Changelog database, step 5: after the restart only [Changelog - Init] schema is up to date (v1); the [Changelog - Import] refreshed line follows later.
• Reset: smoke156-before.db-shm and smoke156-before.db-wal remain in .claude/temp.
• Pending-review alert, over a folder an earlier run left: step 6 lists four alerts, two of them resolved, created before the container started.
• Already-reported conflict, step 3: The underlying connection was closed: The connection was closed unexpectedly.
• Notifications list, dismiss, render, and drive their action: CryptographicException and AntiforgeryValidationException before its first stop, and 7 more lines logged by the stop itself.

Failing tests

Test class	Test method	Status before fix
A file left awaiting review raises an alert, and resolving it retires the alert	Steps 6 and 8 as written	❌
An already-reported conflict does not stage a duplicate on every reseed	Steps 4 and 5 as written	❌
Bulk-deciding a staged batch via file export and re-import, in both wire formats	Step 2 as written	❌
The changelog is served from its own on-disk database, not the JSON fallback	Step 5 as written	❌
Reset wipes the entire database and does not reseed	Cleanup as written	❌
A file left awaiting review raises an alert, and resolving it retires the alert	Step 1, over a folder an earlier run left	❌
An already-reported conflict does not stage a duplicate on every reseed	Step 3 as written	❌
Notifications list, dismiss, render, and drive their action	No [Runtime - Exception] line before its first stop	❌
A file left awaiting review raises an alert, and resolving it retires the alert	No [Runtime - Exception] line before step 9	❌
A new document provoking the antiforgery pair	The pair is logged once; after the restore, the next visit logs nothing	❌

Definition of done

• Failing test(s) listed above are red before the fix is written
• Fix implemented
• All listed tests pass (green)
• No regression in related tests
• Findings summarised in a closing comment