Weaver 0.16.4 release notes

Highlights

Bulk actions over large selections. Cancel, pause, resume, and edit jobs
in bounded groups instead of rejecting the whole selection as too complex.
Every group is accounted for: successful jobs leave the selection, refused
jobs stay selected, and the result reports partial success.
Reliable cancellation. A cancelled job leaves the queue only after its
history record commits. If that write fails, the job stays live and its
repair and completion work can resume instead of disappearing until a
restart. GraphQL adds cancelJobs with a result for each job.
Restore archive write ordering. Revert the delayed disk-write behavior
introduced in 0.16.3 so archive readers see decoded bytes only after their
write completes. This restores reliable direct unpack; slow disk writes can
again briefly delay pipeline and queue updates.

Recover unexplained download stalls. After two minutes with queued work
but no active connections and no known reason to wait, reset the affected
jobs' dispatch holds and try again. If a job makes no byte progress for ten
minutes after that reset, treat its remaining queued articles as missing
and let normal repair and completion checks decide the outcome. Other jobs'
retry deadlines and legitimate waits remain protected.
Accurate download status. A job that has started remains Downloading
between retries or connection assignments. NZBGet-compatible clients no
longer see it revert to queued or incorrectly report server standby.
Keep server rotation flexible. A preference to retry an article on
another server no longer blocks the original server when no alternative
can serve it.
Wait for queue scripts before moving files. Completion waits for all
outstanding queue-event scripts, including FILE_DOWNLOADED runs when no
NZB_DOWNLOADED subscriber is configured.
Fail lost handback writes. A failed disk write while falling back from
direct extraction fails the job instead of leaving an unreported hole.

Responsive bulk cleanup. History deletion and cancellation use batches,
and bandwidth usage saves run through the database writer queue. History
delete batches are limited to 50 jobs to contain slow scripts and errors.
Keep history views synchronized. Announce every deleted history row,
even when another item is refused or filesystem cleanup fails afterward.
Cleanup failures still appear in the operation's result.
Stall recovery diagnostics. Count resets and releases in
weaver_pipeline_download_dispatch_stall_recoveries_total. Set
WEAVER_DOWNLOAD_STALL_RECOVERY=0, false, or off to disable recovery.
Build and test improvements. Optimize additional backup cryptography,
SQLite, and compression crates in test builds. Intel macOS packages use
ThinLTO.