Phase 4 follow-ups: source_id ABC debt + remaining DecompositionPipeline consumer wire-ups

[charlie83Gs]

Follow-ups flagged during PR #279/#280 review. Neither is a blocker for Phase 4 close but should land before Phase 4 is marked fully done.

1. `DecompositionContext.source_id → source_url` mapping is lossy

`kt_core_engine_api.decomposition.DecompositionContext.source_id` and the built-in `LlmDefaultFactDecompositionProvider` conflate "DB id" with "URL" — the adapter forwards `ctx.source_id` onto `TextExtractor`'s `source_url` kwarg because that's the closest analogue on the existing extractor.

Options:

• Rename: `source_id → source_url` on the ABC. Matches existing extractor semantics; forces callers threading a bare DB id to be explicit.
• Split: add both `source_id` and `source_url` typed fields. Providers pick whichever they need.

Prefer split — URL-less sources (uploaded files, future connectors) still have a DB id.

Files:

• `libs/kt-core-engine-api/src/kt_core_engine_api/decomposition/provider.py`
• `plugins/backend-engine-fact-decomposition/src/kt_plugin_be_fact_decomposition/provider.py`
• `libs/kt-facts/src/kt_facts/pipeline.py` (extract_text → DecompositionContext construction)

2. Remaining `DecompositionPipeline` consumer wire-ups

#280 wired only `decompose_chunk_task` (worker-search). Other call sites still run legacy `TextExtractor`:

• `services/worker-search/src/kt_worker_search/workflows/decompose.py` — `decompose_source_task`
• `services/worker-nodes/src/kt_worker_nodes/pipelines/gathering/pipeline.py:532` — `_decompose_images_with_session`
• `services/worker-nodes/src/kt_worker_nodes/pipelines/building/unified.py:137,287` — refresh + expand paths
• `services/worker-nodes/src/kt_worker_nodes/pipelines/nodes/pipeline.py:97,464,751`

Pattern per call site:

1. Resolve provider via `kt_hatchet.composition.resolve_fact_decomposition_provider(state, graph_id, gateway)` at task entry.
2. Pass as `fact_decomposition_provider=` to `DecompositionPipeline`.
3. Worker-nodes sites inside `AgentContext`-wrapped pipelines need `graph_id` threaded through — either extend `AgentContext` or thread as explicit kwarg like Phase 2 fix(helm): use version_variables for Chart.yaml version bumping #33 did.

One PR per call site or one bundled — either works. Use caveman canary test runs to catch regression (`validate_all_graph_types` should stay quiet for the default graph type).

Related

• feat(plugins): backend-engine-fact-decomposition plugin (Phase 4 #44) #279 (plugin + consumer-half merged)
• feat(kt-facts): DecompositionPipeline routes through registry-resolved provider (Phase 4 #44) #280 (pipeline provider-routing merged via feat(plugins): backend-engine-fact-decomposition plugin (Phase 4 #44) #279 squash)

No blockers. Both items are cleanup landed on Phase 4 foundations.

[charlie83Gs]

Addressing directly, not tracking separately.