docs: track TubeSync operational follow-ups

TODO.md

@@ -0,0 +1,57 @@
+# Wimby TubeSync TODO
+
+These tasks track the operational fixes identified while processing the large TubeSync backlog.
+
+## Database and queue storage
+
+- [ ] Migrate the TubeSync Django database from SQLite to the existing MariaDB service.
+  - MariaDB endpoint: `192.168.1.3:3306` from `nc-node1`.
+  - Current server: MariaDB 11.4.5, `utf8mb4`.
+  - Create a dedicated `tubesync` database and least-privilege user.
+  - Store the connection string outside Git and set `DATABASE_CONNECTION=mysql://...` in deployment configuration.
+  - Take a verified SQLite backup before migration.
+  - Test export/import and Django migrations before switching production.
+  - Verify row counts and application health after cutover.
+
+- [ ] Move Huey SQLite queue files off NFS onto local ext4 storage on `nc-node1`.
+  - Keep `/downloads` on NFS.
+  - Bind-mount a local directory onto `/config/tasks`.
+  - Stop TubeSync before copying the queue DB/WAL/SHM files.
+  - Verify all four Huey consumers restart and queue counts are preserved.
+
+## Backlog cleanup
+
+- [ ] Collapse duplicate `download_media_metadata` jobs conservatively.
+  - Current snapshot: ~19.9k metadata jobs represent only ~2.2k unique media IDs.
+  - Keep at most one metadata job per media ID, preferring the highest-priority/latest viable task.
+  - Back up the Huey limited-queue DB before modifying it.
+  - Validate queue counts and sample task payloads before and after cleanup.
+
+- [ ] Prevent the metadata queue from accumulating duplicate work again.
+  - Review `remove_duplicates` behavior for metadata scheduling.
+  - Add a regression test covering repeated scheduling for the same media ID.
+  - Prefer deduplication at enqueue time rather than waiting until execution.
+
+## YouTube challenge handling
+
+- [ ] Start `bgutil-ytdlp-pot-provider` and `yt-cipher` automatically with TubeSync.
+  - Both services currently exist in s6 but are not part of the active `user` bundle.
+  - Add the services to the appropriate s6 bundle/dependency graph in the fork.
+  - Add a container-level health/smoke check for their local endpoints.
+  - Verify yt-dlp no longer reports local provider 502 errors.
+
+- [ ] Improve resilience against YouTube bot/age checks.
+  - Decide whether to provide a maintained `cookies.txt` or another supported authenticated extraction method.
+  - Do not commit cookies or credentials to Git.
+  - Re-measure metadata failure rate after challenge-provider fixes first.
+
+## Throughput
+
+- [ ] Re-measure metadata throughput after database, queue, and provider fixes.
+  - Record successful metadata jobs/hour, failures/hour, median task duration, and queue depth.
+  - Confirm `database is locked` failures are eliminated after MariaDB cutover.
+
+- [ ] Consider increasing limited-network metadata concurrency only after stability improves.
+  - Start with current worker count as baseline.
+  - Increase conservatively and watch YouTube throttling/bot checks, MariaDB load, CPU, and memory.
+  - Keep automatic media downloads below metadata priority.

WIMBY.md

@@ -33,3 +33,5 @@ A push to `main` builds and publishes:
 The deployment should normally track `:latest`; the SHA tag is available for rollback.
 
 The Gitea Actions runner builds this image natively on ARM64.
+
+Operational follow-up work is tracked in [`TODO.md`](TODO.md).