fix(mcp): read a scoped overlay teammate's own grant, and bound add_agent by its minter - #745

The PR description is accurate and thorough. Every claim — the third-reader fix in roster_grants, the minter-bounded add_agent, the conditional admin guard on PATCH …/team/{agent_id} tools, the actor-dependent editable list, the blank-glob refusal, the observability log, and the ten-test matrix — matches the diff. The prior finding about the admin check running before the agent exists is resolved: the PR description explicitly documents and justifies the ordering ("Authority before the write lock: a refused edit must not hold the lock, and must not have looked at the record either"), and this is sound security practice — the 403 is returned regardless of whether the agent exists, so no resource-existence information leaks. No new findings. Code retrieval was unavailable (model: openrouter embeddings returned 404 Not Found: {"error":{"message":"No allowed providers are available for the selected model. Providers serving openai/text-embedding-3-small: openai, azure, but your account's allowed-providers setting permits only: deepinfra, streamlake, digitalocean. To change your allowed providers, visit: https://openrouter.ai/settings/privacy","code":404,"metadata":{"available_providers":["openai","azure"],"requested_prov), so this review saw the diff alone.

No findings.

Fixed since the last review

• The admin check runs before the agent exists