fix(gateway): resume a root whose yield source was lost instead of failing it (#202044)

Problem: at Gateway startup, every yielded session whose wake source could not survive the restart was closed with a visible failure. A sub-agent is then settled by the registry and its requester wakes (#202000), but an ordinary root has no registry row, so its work stayed failed until a human wrote to it.

Fix: resolveSessionYieldWait (the registry read owner) reports a root as waiting while a child still owes it a settlement wake, an unfinished completion, a yield or a pending/in-progress delivery; that registry completion is the root's only wake. A root still unreachable (nothing will wake it) is admitted through the ordinary restart-resume input instead of failed. Sub-agents keep the lost-source failure plus registry settlement. No ordering or handshake: the registry is shared across physical agents and hydrated before restart resume. A failed delivery is terminal and does not count as owed.

Proof: source tier (ocx test): red on unchanged main for both roots (marked failed); green with one restart input across two boots, or one subagent_settle wake with no restart input. 230 tests across 19 files; tsgo, test types, oxlint, format, lifetime-check, line-cap and session-identity ratchets; clean exact-head local ClawSweeper review; ocx build --dts plus packed install smoke on the head. Real-Gateway restart proof (isolated Gateway, real model, natural SIGKILL source loss): root without owed child gets restart x1 and ends done; root with settled lost-source child gets subagent_settle x1, restart x0, ends done; unchanged main ends failed.

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