tenant.openagents/omega

Cut 0.2.0-rc15 on an engine that keeps going when nobody is watching

Cut 0.2.0-rc15 on an engine that keeps going when nobody is watching rc13 shipped, and rc14 was cut, on omega-effectd rc.9. Driving FA-07 gate 2 (omega#26) against the engine inside installed rc13 found there was no such thing as an unattended multi-turn run: one real Codex turn (exit 0, 164 s), then eighteen minutes of nothing, including a 180-second window with zero framed calls of any kind and the turn count stuck at one. Reconciliation is the only thing that dispatches a continuation, and on the framed transport every trigger for it was a control mutation. The three methods the Omega run monitor polls every three seconds -- list_runs, get_run, decide_attention -- deliberately do not mutate, which is right of them and is exactly why none of them reconciled. rc14 could not have fixed it. It is a one-line version bump that leaves the engine pin on rc.9, whose source commit predates the fix; installing it and re-driving gate 2 would have reproduced rc13's red exactly. That is the mistake rc11 made, and repeating it would have cost another sign-and-notarize cycle to learn nothing. rc15 pins omega-effectd rc.10 (openagents e3280b4795), which reconciles on a clock instead of on a person, only while a non-terminal run exists. Proved against the released artifact rather than its source: the gate 2 driver run against rc.10's own bin/omega-effectd reports multiTurn true and turnCount 3 across a window in which the driver made no framed call at all. The same driver against rc.9 reports multiTurn false. script/prove-omega-rc-install carries the same pin, so it would otherwise reject a correctly pinned candidate. Release Notes: - N/A