Add dashboard diarize action

[PsychOsmosis]

Summary:

• Add a dashboard action to trigger diarization for existing transcriptions.
• Add the API route/handler and service flow needed to run diarization from the dashboard.
• Add persistent diarization worker controls and idle CUDA VRAM reservation.
• When a task requests the non-resident model, swap workers with the same reservation and restore the original model, including after task failure.
• Add persisted global System settings for single-user versus multi-user deployment policy and process-wide startup model selection.
• Restrict global settings to the first registered account and resident-model controls to that owner in multi-user mode.
• Keep profile Hugging Face tokens scoped to model-load requests; do not copy profile credentials into the process-wide HF_TOKEN environment variable.
• In multi-user mode, startup PyAnnote requires a server-level HF_TOKEN.
• Migrate the existing per-user startup model preference into the new singleton global setting once.

Testing:

• env GOCACHE=/tmp/scriberr-gocache GOMODCACHE=/tmp/scriberr-gomodcache /usr/local/go/bin/go test ./...
• env GOCACHE=/tmp/scriberr-gocache GOMODCACHE=/tmp/scriberr-gomodcache /usr/local/go/bin/go test -race ./cmd/server ./internal/api ./internal/repository ./internal/transcription/adapters
• npm run build
• npm run lint

[PsychOsmosis]

I rebased this branch with your recent changes.
I find it useful to be able to rediarize with Sortformer when Pyannote did a bad job without having to retranscribe everytime.

[PsychOsmosis]

@abn What do you think ?

[PsychOsmosis]

@abn I also added a way to keep the diarization model always loaded on VRAM.
That was my last remaining pain point with Scriberr and now it's fixed.
You can choose to auto-load a model on startup or not, and load and unload it on the fly with a clickable button.

[abn]

Awesome @PsychOsmosis Let me try get to this over the weekend (that magical weekend 😄 )

[Copilot]

Pull request overview

This PR adds a “diarize” action for already-transcribed jobs in the dashboard UI and introduces a backend “diarization-only” flow, including support for a persistent (resident) diarization worker that can be loaded/unloaded (and optionally loaded at server startup).

Changes:

• Frontend: adds a diarize action in the transcription list (desktop + swipe actions) and reuses the existing profile dialog to run diarization-only with optional forced diarization settings.
• Backend: adds /api/v1/transcription/:id/diarize plus DiarizationOnly job handling to load an existing transcript and merge new diarization results without re-transcribing.
• Adds a persistent diarization worker manager (PyAnnote + Sortformer) with new API routes and UI controls, plus a user setting to request loading a diarization model at startup.

Reviewed changes

Copilot reviewed 17 out of 17 changed files in this pull request and generated 4 comments.

Show a summary per file

File	Description
web/frontend/src/features/transcription/components/AudioFilesTable.tsx	Adds a dashboard “Diarize” action and submits diarization-only jobs against completed transcripts.
web/frontend/src/features/settings/components/ProfileSettings.tsx	Adds a user setting to choose a diarization model to keep loaded at startup.
web/frontend/src/components/ui/swipeable-item.tsx	Adds an optional “Diarize” swipe action and dynamically sizes the action area.
web/frontend/src/components/transcription/TranscriptionConfigDialog.tsx	Extends WhisperXParams typing to include diarization_only.
web/frontend/src/components/TranscribeDDialog.tsx	Generalizes dialog copy/labels and adds forceDiarization to show diarization settings even if profile disables them.
web/frontend/src/components/Header.tsx	Adds UI + polling + load/unload dialog for the persistent diarization worker.
internal/transcription/unified_service.go	Implements diarization-only processing by loading an existing transcript and optionally running diarization separately.
internal/transcription/queue_integration.go	Exposes persistent diarization worker operations via the unified job processor.
internal/transcription/adapters/sortformer_adapter.go	Adds persistent-worker execution path and embeds/copies the worker script.
internal/transcription/adapters/pyannote_adapter.go	Adds persistent-worker execution path and relaxes HF token requirement when the resident model is loaded.
internal/transcription/adapters/py/pyannote/pyannote_worker.py	Adds a long-lived PyAnnote diarization worker process (stdin/stdout JSON protocol).
internal/transcription/adapters/py/nvidia/sortformer_worker.py	Adds a long-lived Sortformer diarization worker process (stdin/stdout JSON protocol).
internal/transcription/adapters/persistent_diarization_manager.go	Adds the Go manager that owns the single resident worker process and request protocol.
internal/models/transcription.go	Adds diarization_only to job parameters and startup_diarization_model to users.
internal/api/router.go	Registers the new transcription diarize route and diarization-worker management routes.
internal/api/handlers.go	Implements StartDiarization and persistent diarization worker load/unload/status endpoints + user setting validation.
cmd/server/main.go	Loads a persistent diarization model at startup based on user settings and resolves HF token from a profile.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[PsychOsmosis]

@abn Please run a new Copilot review, I just fixed all the issues that were raised.