permission, not the queue

proved the mono lander idle every time its queue was read, and separated that from the thing actually broken
- every read came back empty with origin/main static because the last task landed and its worktree was already swept
- the lander log repeated one workload key refusal about a hundred and ninety times, so the server will not mint the authority this machine needs and every future landing sits behind it
- the stall classifier attributed the queue to work that was never there: its own reads failed as malformed responses under a load average near two hundred
recorded the result as needs-you instead of touching anything, and found the outcome ledger only accepts a write when the agent session id is exported into the call

traced the desktop connection failures to two separate faults rather than one
- a backend hosting twelve agents answered even trivial requests in seven or eight seconds under that same load
- the app window itself sat at 213 percent CPU, and the repeated errors came from one transcript that threw while being packed for context handoff
reported both instead of restarting the app, because a restart would have erased the transcript that broke

found a session left over from an inadvertent reboot cannot be resumed from outside the app: the control socket turns away every caller that was not started inside it
- nothing was nudged and no pane was written to

git identity remains unfinished: an account exists and one commit already carries it, while every other commit lands under another name, and the run that was meant to prove otherwise died twice under load