fix(gateway): refuse a retained continuation at start when its operator's device was revoked (#202318)

Problem: a retained continuation restored from a tokenless (trusted-proxy) operator carried no device in its snapshot, so its own revocation capture never joined the device index; if the operator's device was revoked before the continuation started, the start check passed and the replayed send ran. On main a post-dispatch check then replaced the response with a refusal notice, so the requester was told "refused" while the target had received the send.

Fix: captureGatewayOperatorContinuation keeps the origin's device capture indexed for every hold through a revocation-only index hold (retainGatewayDeviceRevocationIndex; no caller or parent-invocation check, so a detached child stays valid after its parent chat is edited), and its start check refuses with "Operator access was revoked; start a new user turn." when that capture was revoked, then runs the restored check. The refused input settles as cancelled/revoked through the existing admission failure path. No new state or schema. Accepted edge case (maintainer): after a real restart a tokenless operator's restored authority cannot see a later device revocation; role, grant and profile revocation are still checked on restore.

Proof: source tier. Unit tests: allowed passes and revoked throws after the origin request ends, and after the parent invocation becomes obsolete with a new hold (red on the previous head). Product test server.plugin-reload-send-revocation: allowed effects 1; revoked effects 0, message requests 0, one notice, cancelled/revoked history (on main the revoked replay reached the target). 132 tests across the recovery, device-revocation and authority suites; tsgo, oxlint, format, lifetime-check, line-cap and max-lines ratchets; clean exact-head local ClawSweeper review; ocx build --dts plus packed install smoke.

Co-authored-by: Ayaan Zaidi <hi@obviy.us>