feat(plugins): backend-engine-fact-decomposition plugin (Phase 4 #44)

[charlie83Gs]

Summary

First concrete Phase 4 provider — `LlmDefaultFactDecompositionProvider` wraps the existing `kt_facts.extraction.TextExtractor` and contributes `llm-default` via the plugin registry. Registry lookups for the id `DefaultGraphTypePlugin` declares in `composition().fact_decomposition` resolve here instead of callers importing `TextExtractor` directly.

• New `plugins/backend-engine-fact-decomposition/` package.
• Registered in `_DEFAULT_PLUGIN_TARGETS`. `is_enabled()` gated on `plugin.fact_decomposition.enabled` (default True — built-in convention).
• CI: new `plugin-fact-decomposition` entry in paths-filter + maybe_add matrix.

No new extraction logic — extractor source stays in `kt-facts` for this slice to keep blast radius to a registration adapter; a follow-up can relocate `TextExtractor` into this package once no other caller imports it directly.

Why

Closes first half of Phase 4 #44 / #45. The worker-side wire-up (`DecompositionPipeline` resolving via `ctx.config.composition.fact_decomposition` instead of constructing `TextExtractor` by hand) lands in a stacked follow-up.

Test plan

• `uv run --project plugins/backend-engine-fact-decomposition pytest plugins/backend-engine-fact-decomposition/tests/ -x -v` — 5/5 green (plugin identity, is_enabled default, llm-default contribution, factory→ABC, registry round-trip)
• `uv run --project libs/kt-config pytest libs/kt-config/tests/ -x -q` — 236/236 green
• `uv run --project libs/kt-facts pytest libs/kt-facts/tests/ --ignore=tests/integration -x -q` — 232/232 green
• CI: full workspace tests

🤖 Generated with Claude Code

[charlie83Gs]

Review items 1 + 2 addressed directly on this branch (df790b3).

Item 1 — source_id / source_url split on ABC

• DecompositionContext.source_url promoted to typed field alongside source_id. source_id reserved for opaque DB identifiers; source_url carries retrieval URL. Drops the lossy conflation where the adapter forwarded source_id onto TextExtractor.source_url.
• Plugin adapter reads ctx.source_url directly; ignores source_id (URL-only extractor doesn't need it).
• DecompositionPipeline.extract_text passes the URL kwarg to source_url= (was source_id=).
• Test updated to assert ctx.source_url + ctx.source_id is None on URL-only caller.

Item 2 — all DecompositionPipeline consumers wired

• AgentContext grows fact_decomposition_provider + fact_decomposition_settings optional fields. Propagates through create_child_context.
• Both _build_agent_context helpers (worker-bottomup.shared, worker-ingest.workflows.ingest) auto-resolve from kt_hatchet.composition when not pre-supplied. Caller-supplied values take precedence (workflow entries can pre-resolve to avoid duplicate lookups).
• All 5 bare DecompositionPipeline(ctx.model_gateway) sites in worker-nodes (building/unified ×2, nodes/pipeline ×3) now thread phase_settings + fact_decomposition_provider from ctx.
• GatherFactsPipeline._decompose_images_with_session forwards ctx.fact_decomposition_provider.
• decompose_source_task (worker-search) resolves provider at entry.

No legacy TextExtractor path remains for graphs whose composition names a registered provider. Unregistered-id path still WARN + falls back (canary tolerance preserved).

Tests: 175 green across kt-agents-core, kt-facts, worker-search, worker-nodes, worker-bottomup.

[charlie83Gs]

Addressed review feedback in 2 commits on top of #279:

Commit e7be609 — style(worker): ruff format 6 files — fixes the Backend Lint CI failure.

Commit 8bee23a — refactor(worker): drop pre-resolve at fact-decomp task entries — addresses the double-resolve risk (issue #1). Task entries no longer pre-resolve provider/settings; _build_agent_context is the single source of truth and stamps the values onto the returned AgentContext. Downstream calls that still need the settings (run_bottom_up_scope_pipeline) read from agent_ctx.fact_decomposition_settings. Also drops the now-unused kwargs from both _build_agent_context signatures.

Notes on other items:

• Issue fix: resolve cross-worker test imports and enforce unit/integration test boundaries #2 (long lines in worker-nodes) — ruff format wrapped the DecompositionPipeline(...) call sites; no 120-char violations remain in PR-touched code (the 120+ char lines pyright flagged are pre-existing section-header comments untouched by this PR).
• Issue feat: integration test pipeline with containerized services #3 (weak typing on fact_decomposition_* params) — moot: the kwargs are now removed from both _build_agent_context helpers in commit 8bee23a.

All 672 files formatted; all unit tests pass (14 libs + 7 services + 2 plugins).

[github-actions]

Thank you for your submission, we really appreciate it. Like many open-source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution. You can sign the CLA by just posting a Pull Request Comment same as the below format.

---

I have read the CLA Document and I hereby sign the CLA

---

_{You can retrigger this bot by commenting recheck in this Pull Request. Posted by the CLA Assistant Lite bot.}