I Built a Delete Button That Refuses to Delete My Old Music Pack.
My music library had 40 candidates and a delete button to build. The awkward part wasn’t removing a row. It was proving which files belonged only to that row.
Each song had A/B candidates. There were source files, delivery files, comparison versions, preferences, and possible links to a game chart. A candidate disappearing from the list didn’t mean its source had stopped being useful elsewhere.
I chose a limit that still exists in the October 12, 2026 verification record: the old music pack couldn’t be permanently deleted through the workbench. I hadn’t proved all its references. The cost was disk space I couldn’t safely reclaim from that interface.
That’s less satisfying than a tidy trash can. It was also the result I could defend.
I split hiding a candidate from removing bytes
The workbench’s ordinary delete action moved a candidate to the trash. It retained preferences, notes, and source links. Restoring it brought that state back.
Permanent deletion was a separate operation. It checked whether a file was shared, approved, used in the game, or referenced by a chart. The implementation was AI-assisted; I checked these behaviors with temporary audio and an isolated database.
I didn’t use my real library as a deletion test fixture. I’ve already lost manually collected data to a test cleanup. I didn’t need a second example for the same rule.
| October 12 check | Result recorded |
|---|---|
| Trash, restart, and restore | Notes and candidate state preserved |
| Approved or game-used candidate | Permanent deletion blocked |
| Shared file hash | File protected |
| Chart reference | Deletion blocked |
| Database deletion failure | File returned to its original location |
| Successful isolated deletion | Managed file removed; reimport prevention recorded |
| Original library after checks | All 40 MP3 hashes unchanged |
Those are distinct checks. “The delete request returned success” wouldn’t substitute for any of them.
A shared file isn’t owned by its first visible row
The interface encouraged a simple mental model: one row, one song, one file. My storage didn’t follow it.
A delivery file could be one representation of a candidate with a separate source. A comparison could still depend on a file I was about to remove. A game chart could refer to an approved version even after I stopped wanting it in the current browsing list.
I needed a reference check before removing a managed file. If I couldn’t establish that the old pack was unreferenced, I blocked deletion rather than guessing from the current view.
That doesn’t mean every file was shared. It means I couldn’t prove that every relevant connection had been checked. “Unknown” wasn’t a synonym for “unused.”
The database and filesystem could disagree
Deleting metadata and deleting a file are two different changes. A database rollback doesn’t automatically put a removed file back.
The isolated failure check mattered because it exercised that boundary. When the database deletion failed, the file was restored to its original location. I verified that behavior with temporary material, not with the only copy of a finished candidate.
I won’t call this protection against every possible power loss. The recorded test covers an injected database failure. It doesn’t establish crash consistency at every instruction or prove that a damaged disk can recover itself.
That limit belongs beside the success result. Otherwise a useful rollback test starts sounding like a backup system.
A missing drive wasn’t an empty library
The audio lived on external storage. I also needed to know what happened when that storage wasn’t available.
The isolated check kept the catalog and notes. It didn’t create a replacement source folder and pretend the old files had vanished. The interface could report that the storage needed connecting without destroying the metadata I’d already accumulated.
That was a separate boundary from deletion, but the same ownership question applied. A path I can’t read right now doesn’t authorize me to replace what’s supposed to be there.
Don’t promise that delete means every copy is gone
For newly generated candidates, permanent deletion covered workbench-managed files. It didn’t automatically remove the studio’s remote output copy.
I wanted that limitation stated plainly. Otherwise a button named “permanent delete” would imply a broader operation than the system performed.
I also wouldn’t turn a favorite star into final approval. The library started with four preferred candidates, but preference wasn’t game inclusion. Those states had different consequences and needed different controls.
The original 40 candidates survived the checks with identical MP3 hashes. That verifies preservation for this run; it isn’t proof that every future deletion is safe.
What I accepted
I accepted a trash can that didn’t instantly free disk space. I accepted a permanent-delete action that refused some requests. I accepted that old comparison references needed investigation before I could reclaim their files.
The next improvement wasn’t a more forceful button. It was better reference information.
My library became easier to manage once it could tell me why deletion wasn’t allowed. I didn’t need it to guess on my behalf, especially when guessing could erase the evidence I’d need to undo the mistake.