A delivery could reach a bad end without the engine recording why, and
a retrying delivery could fail to reach an end at all.
An unknown target type marked the delivery failed and wrote no
DeliveryResult, so the event log showed "failed" with no attempts and
the only account of why was one line in the server log. It now records
a result naming the type before failing the delivery.
A deleted target left its retrying deliveries stranded. Both recovery
and the sweep began with a scoped loadTarget, which cannot see a soft
deleted row, so both logged and returned: the delivery stayed retrying
for the life of the database while the sweep repeated the same error
every minute. Both now terminalise it with a recorded reason.
Deleting a target also did not stop deliveries to it. A scheduled
retry is a time.AfterFunc holding the target's configuration from when
the chain began, and nothing on that path read the target row, so the
timer kept firing and kept sending to the destination the operator had
removed for the rest of the backoff chain; terminalising in recovery
and the sweep alone would only have caught it after a restart.
processRetryTask now confirms the target still exists before it
attempts, and abandons the chain when it does not.
Only a target confirmed gone stops anything. A lookup that fails for
any other reason is the main database being unreadable, which is
transient, and every path leaves the delivery exactly as it was rather
than failing it.
The reason text comes from one Unscoped lookup confined to these
terminal paths, because a soft deleted row is what distinguishes a
target the operator deleted from an id that never named one. The
engine's normal target loading stays scoped, or deleting a target
would stop nothing.
Terminal writes reached from recovery keep going through the existing
retainIdle ownership gate; the retry path writes directly, as a
target's own Deliver does, because the worker already holds that
delivery. Nothing was added to either sweep dispatch arm.
Retry fixtures that drove processRetryTask for a target id with no row
in the main database now create one. That state is not reachable in
service: the handler reads the target to build the task.
GORM's association upsert copied whole targets rows -- plaintext
credential-bearing config -- into the per-webhook event databases with an
empty webhook_id. The leak was in updateDeliveryStatus, not the create
path: Update leaves Statement.Model pointing at a Delivery whose Target
the engine populated, so save_before_associations upserts it.
A connection-level callback now appends clause.Associations to
Statement.Omits on the create and update chains of every per-webhook
connection, so every write path is covered rather than one call site.
Existing files are swept on first open: the leaked rows are deleted and
the file is VACUUMed, because DELETE alone only unlinks the pages and
leaves the credential recoverable in the file's free space. The sweep is
recorded in PRAGMA user_version only after the VACUUM returns, so a sweep
that fails or is interrupted fails the open and is retried on the next
one, rather than being marked done.
Encryption of target config at rest is deliberately out of scope and
deferred to #212.
