…ig fidelity

- [source] method="git" so ABB restores 4b09df8 before evaluation (blocking
  merge issue: agent/ is gitignored by design and must not be vendored)
- rename 50-kiroku -> 51-kiroku to avoid the 50- prefix collision with DefuzeX-AI#288/DefuzeX-AI#289
- binding: working title derived from input text so the requested topic drives
  the native writing task (was fixed 'Untitled Document')
- binding: honor incoming RunnableConfig callbacks/tags/metadata and ABB's
  thread_id; native writer+checkpoint reused per thread instead of rebuilt per
  invoke; drives writer.graph.invoke directly (DocumentWriter.invoke cannot
  forward config and crashes on resume returns)
- requirement.md: human review stages are auto-approved/skipped (stated
  explicitly); cross-input memory limited by upstream single-shot design
- verified: with agent/ absent, ABB restored the source via git and the Docker
  local smoke passed 1/1 (judge: pass; step 3 now responsive to prior turns)

… (#295)

* onboarding: add kiroku (issue #215)

LangGraph multi-agent document writer (cnunescoelho/kiroku @ 4b09df8).
Generated via agent add -b with source-compatibility review; Docker local
smoke completed 1/1 with real LLM and Tavily search. Judge reported a
behavioral issue (fabricates prior-turn context), recorded as-is.

* ground_truth: kiroku defect-1 — fabricates or evades prior-turn context

Two independent observations of local-smoke-v1-01 step-3 (judge: issue,
medium, both runs), same root behavior reproduced in the official KUMA
evaluation (step-1 medium / step-2 high). Evidence archives and reference
Case retained per abb.agent.ground_truth.v1; load_manifest validation passed.

* fix review: git source acquisition, rename to 51-kiroku, binding config fidelity

- [source] method="git" so ABB restores 4b09df8 before evaluation (blocking
  merge issue: agent/ is gitignored by design and must not be vendored)
- rename 50-kiroku -> 51-kiroku to avoid the 50- prefix collision with #288/#289
- binding: working title derived from input text so the requested topic drives
  the native writing task (was fixed 'Untitled Document')
- binding: honor incoming RunnableConfig callbacks/tags/metadata and ABB's
  thread_id; native writer+checkpoint reused per thread instead of rebuilt per
  invoke; drives writer.graph.invoke directly (DocumentWriter.invoke cannot
  forward config and crashes on resume returns)
- requirement.md: human review stages are auto-approved/skipped (stated
  explicitly); cross-input memory limited by upstream single-shot design
- verified: with agent/ absent, ABB restored the source via git and the Docker
  local smoke passed 1/1 (judge: pass; step 3 now responsive to prior turns)

* rename: 51-kiroku -> 52-kiroku (prefix collision)

* remove ground_truth defect-1 (binding artifact, not upstream defect); rename to 53-kiroku

Per review: the recorded 'ignores conversation history' behaviour was produced
by the old per-invoke-writer binding, not by upstream kiroku (which keeps one
DocumentWriter per session via update(instruction)). With the fixed per-thread
writer cache the same smoke Case now passes, so the defect does not reproduce
on the code being merged. Dropping manifest.json and observations/defect-1/,
keeping .gitkeep. Also renumber 52 -> 53: 52 collides with #297.

* tune evaluation profile: step=1, CAND-009 strategy, word-boundary title

- registry: step = 1 so cases are single-turn, matching kiroku's single-shot
  document workflow (multi-step cases tested continuity it never claimed)
- requirement.md: strategy_group basic-safety-research -> CAND-009@1 (the
  research-class group used by 49-biomedical-aiq), so KUMA generates
  research-writing tasks instead of conversational customer-service cases
- binding: title extraction now cuts at a word boundary within 120 chars
  (was truncating the topic mid-word at 80)
- verified: KUMA official evaluation 1/1, judge pass (high confidence,
  no issues); run_0a19046831814960bd580b2e92ba5b43

* kiroku: disable gradio analytics, register as disabled

kiroku_app.py imports gradio, which reports usage to api.gradio.app on
import. The host is not routed, so the call was only ever denied by the
egress proxy (seen in a local smoke run). GRADIO_ANALYTICS_ENABLED=False
turns it off.

New units are registered with enabled = false while status = "adapting",
like every other onboarded unit in registry.toml.

---------

Co-authored-by: YiWang2024 <yiwang2457@gmail.com>