doc: add ADR with preview feature

Author: itallix

docs/geti_tune/adr/adr0004_preview.md

@@ -0,0 +1,120 @@
+# ADR-0004: Source Preview Feature
+
+- Status: `Proposed`
+- Deciders: @itallix @leoll2
+- Date: 2025-08-28
+
+Technical Story: #4595
+
+## Context and Problem Statement
+
+Users need to preview the source of a pipeline (e.g., video file, camera stream) without running the entire pipeline.
+This allows validation that the source is correctly configured and accessible before enabling processing and inference
+through the full pipeline.
+
+## Decision Drivers
+
+- Preview feature should be designed around existing concept of pipelines.
+- Only one pipeline can be active at any time (either preview or full).
+- The preview pipeline outputs to a dedicated WebRTC stream.
+
+## Considered Options
+
+1. **Preview thread:**
+
+   Reuse existing pipeline infrastructure. Introduces a thread that links the frame queue to the WebRTC preview stream.
+
+   - Good:
+
+     - Uses established IPC components (queues and event).
+     - Minimal additional code outside thread management (relaxed validation for pipeline).
+     - No changes to existing worker routines or API.
+
+   - Bad:
+
+     - Idle thread consumes some resources when preview is not active due to preemptive multitasking.
+     - More threads mean slightly increased debugging and maintenance complexity.
+
+2. **Dummy configs:**
+
+   Run pipeline with "dummy" sink/model configuration, so inference and dispatch routines are started but don't perform work.
+
+   - Good:
+
+     - No new threading/IPC.
+     - Relies on full pipeline logic for preview state transitions.
+
+   - Bad:
+
+     - Introduces more complexity: workers logic should be extended, new configuration types.
+     - New orchestration phases should consider dummy types.
+     - Unintended data flow (it will still write to the queues that are not needed for preview flow).
+
+3. **Preview source only:**
+
+   Create a separate preview abstraction around the source component, bypassing pipeline logic.
+
+   - Good:
+
+     - Better fits conceptually as it's totally separated from pipeline logic.
+
+   - Bad:
+
+     - Introduces more complexity: requires DB schema update, new APIs, new lifecycle management.
+     - Does not meat decision drivers.
+
+## Decision Outcome
+
+**Chosen option:** Preview thread
+
+- Meets all decision drivers.
+- Requires minimal changes and keeps pipeline abstraction intact.
+- Outputs to existing WebRTC stream.
+- No effects on the main pipeline workers or APIs.
+
+### Positive Consequences
+
+- No changes to core workers routines.
+- No additional IPC primitives required: queues and events reused.
+- Pipeline abstraction remains consistent.
+- No new API endpoints needed; existing logic covers preview functionality.
+
+### Negative Consequences
+
+- Idle preview thread may use some CPU/memory when no preview is active.
+- Multithreading complexity increases.
+
+### Implementation Details
+
+The Scheduler is responsible for creating and cleaning up processes, threads, IPC resources, and related components.
+The Scheduler should create a new thread called "Preview":
+
+![Preview Thread Diagram](../media/adr0004_scheduler.jpg)
+
+**Preview Thread Behavior:**
+
+The Preview thread runs only when the pipeline is in preview mode.
+It reads frames from the source frame queue and pushes them to the existing WebRTC queue.
+
+**Preview Mode Conditions:**
+
+- The pipeline has a source configured.
+- The pipeline does not have a model configured.
+- The pipeline does not have a sink configured.
+- The pipeline is in the running state.
+
+![Preview Flow Diagram](../media/adr0004_preview_flow.jpg)
+
+**Note:**
+The Preview thread operates only alongside the "Stream loader" process, as other workers are inactive in preview mode.
+
+### Open Questions?
+
+- Is there any benefit to having a separate WebRTC queue for preview mode, or should we reuse the existing queue?
+- Should the Preview thread be created on demand when the pipeline enters preview mode, rather than always running after
+  being created by the Scheduler during server startup?
+
+## Links <!-- optional -->
+
+- [MADR project template](https://github.com/joelparkerhenderson/architecture-decision-record/tree/main/locales/en/templates/decision-record-template-of-the-madr-project)
+- Implementation: see [Scheduler](../../../backend/app/core/scheduler.py), [Pipeline](../../../backend/app/schemas/pipeline.py)