[xyi74976-del]

Summary

Fixes #2467. CLI creation persists experts through a separate embedded runtime. The running Dashboard keeps its earlier harness registry, so a valid new team can return an empty agent_list and fail peer resolution. Updating an avatar merely triggers a reload; the avatar values are unrelated.

Load missing eligible roster experts before the team host starts or reloads. Keep existing instances and recheck the current host, member, roster, state and ownership/sharing permissions inside the member lifecycle lock. Explicitly stopped, disabled, deleted, private or nested-team entries remain unloaded; two null owner IDs do not establish ownership.

Related #2471 changes first-invocation loading. This regression occurs during peer discovery before a member can be invoked.

Target branch and change type

• Base is develop
• Bug fix; changelog and verification notes updated
• Added regression tests

Verification

• On unchanged upstream, the three initial discovery/reload cases failed. Separate deterministic races exposed stale state after waiting for the member lifecycle lock.
• All 10 new scenarios pass: real SQLite and two real harness registries, null/populated profiles, shared/admin/private and null ownership, existing-instance preservation, and stop/disable/delete/unshare/roster-removal races. Independent AI review found and corrected the null-owner policy case; its independent rerun passed all 10 cases.
• Related agent/team suite: 216 passed, 1 skipped.
• make all PYTEST_JOBS=2 with --dist=loadfile: 5,313 passed, 35 skipped; Ruff and strict mypy passed. The backend/test blobs were unchanged through the subsequent upstream CSS-only synchronization to 44c22434.
• Normal pre-commit hook and dashboard production build passed at the published head, without bypass.

Native Windows, live providers, a real operator database and an interactive Dashboard end-to-end run were not tested. Existing dependency deprecation warnings remain. Hosted CI is separate and is not claimed as passed.

Evidence and review

Evidence boundaries. The screenshot renders actual local test summaries; it is not a product UI capture or human-review attestation.

Codex assisted investigation, implementation, tests, independent AI review and this description. Contributor review confirmed for 8f5551469740a17f65376cb22499f2312f054210 on 2026-10-29.

Hosted checks

At current head 8f5551469740a17f65376cb22499f2312f054210, the CodeCC scan passed. The upstream CI workflow is awaiting maintainer approval (action_required); its test jobs have not executed. Local results above are separate from hosted CI. The contributor has confirmed review of this exact revision.