Give `clear_dark_triggers()` a dedicated
`MsgStream.subscribe()` handle with `raise_on_lag=False` instead
of mutating the private lag policy on whichever root or cached
child sits in `Flume.stream`.
Also,
- factor the hot quote loop into `_clear_dark_triggers()` so the
public task owns the child receiver for its full lifetime.
- add an EMS regression proving the lag policy, child lifetime,
and source-stream state remain isolated.
(this patch was generated in some part by `opencode` using `gpt-5.6-sol` (`openai`))
Gud Boi