PR #17885 takeover verification
Head 85bf1cf0572264e490f50c08672b152cf89bc671
Base 1451e7cb23034d29774338608368a15d8845df36
PR-owned tests 224 -> 171 added lines across existing files.
Six suites: 415 pass.
Exact-main negative control: four expected failures (wake hold, SQL readiness, memory readiness, Codex message_text delivery); Stop and missing-worktree controls pass.
Additional MCP toolkit: 2 pass. Focused cross-thread caller ownership: 1 pass.
Server-scoped typecheck, targeted lint and format pass; existing warnings only.

Native service integration: actual Claude SDK0.3.276/CLI2.1.303, production adapter, SQL event/projection store, continuation service, orchestrator and effect executor.
A native background Agent runs while root settles. The user queue hold is seeded through EventSink, rather than an actual server restart. The effect worker is manually drained to control checkpoint timing.
ROOT_WAITING_HELD_QUEUE_SEEDED: root ordinal1 waiting, user ordinal2 queued queueHeld=true.
NATIVE_WAKE_QUEUED: root waiting, user queued held, wake ordinal3 queued without hold.
Checkpoint completes root and wake starts.
HELD_USER_PRESERVED: queued, queueHeld=true.
NATIVE_WAKE_HAS_ACTIVE_T3_RUN: ordinal3 running, queueHeld=false.
Native child tools execute in the CLI; Anthropic model responses are a deterministic loopback stub.
No live cross-thread MCP invocation was made from that wake; separate service/toolkit tests and source ownership trace support that boundary.
No full app restart, browser or other-platform claim. Stop, missing-worktree, SQL/memory and Codex classification are test-level proof, not live-provider proof.
Human scheduler review remains required.