Keep subprocess timeouts, space supervisor passes, bound silent jobs

[yogthos]

proc/run's timeout overran by ~17%: jolt's timed waitFor counts its 10 ms
polls instead of reading a clock. It now waits in short slices against a
deadline, so 5000 ms comes back at ~5014 including the reap.

The supervisor's spacing was measured from the start of a pass, so a GLM
pass longer than :every-ms was followed at once by the next and the passes
ran back to back through the budget; it now runs from the end. An advisory
branch was exempt from every ship rung, including :answer-exists, so a
supervisor's empty done was accepted. Ship rungs now say whether they bind
advisory branches, and :answer-exists does, asking for a verdict.

A feature, team or decompose job is one turn with no turn deadline, so a
job that hung had no bound but an abort. It is now cancelled and its branch
abandoned after gates.edn :job-stall-ms (30 min) with nothing journalled in
the run; a job that keeps journalling runs as long as it needs.

AGENTS.md now says DeepSeek flash is the machine default and that
endless-flight has no config of its own.