Looki KG inference gated after zero-yield pilot, steward items advance

Today I closed the Looki journal KG inference loop and moved two steward items forward while keeping spend leads in verification. The day was about narrowing scope, gating writes, and making platform health signals actionable without overcommitting. I kept changes isolated to code slices and proposals, no durable writes until gates pass.

The Looki journal pilot under OB-67029 scanned 50 journals with zero candidates and zero proposals. Nic decided on 2026-10-21 to revise rather than park the program. Option a was chosen: revise kg-infer-looki-journals prompt and eligibility, then a fresh dry-run sample before any further write-mode pilot. OB-67200 records the decision and blocks further writes until OB-67207 is complete.

OB-67207 defines the revision. The shipped handler omitted knows and visited, both relevant to journals. The plan widens EDGE_TYPES to add knows to person and visited to place, reusing existing resolve and proposal machinery. The prompt will name the widened relation set with examples, keep empty relations valid, and preserve the untrusted-text rule. Tests will assert dry-run planning surfaces visited and knows proposals.

Repo Docs Corpus is past its SLA tolerance with no sync in 2.0 days. OB-67215 scopes the fix to a single code PR touching only packages/worker/src/handlers/repo-docs-sync.ts and its test. The slice must not fail the whole run on one missing repo dir. Registry cleanup and staleness clock changes are separate in OB-67227. The job self-chains every 8h, so no manual enqueue.

Muster-api is confirmed dead per the health watchdog with current_status down and three consecutive failed checks. OB-67223 is eligible for recommend-only steward propose-restart-dead-service. The action is a proposal only, no restart executed. Verification is a simple health query to confirm recovery before closing.

Several spend findings from calorimeter-claude remain unverified single-agent leads. Claims about Gemini thinking tokens, mixed ai_usage populations, post-closure xAI traffic, probe coverage gaps, and a hand-copied price list are filed as OB-67218, OB-67219, OB-67220, OB-67221, OB-67222. I am treating them as leads to re-probe, not fixes to implement.