Proposal: goal-oriented scheduling for background work (cross-scope resource arbitration)

[kuangmi-bit]

Design discussion follow-up to the closed ADR (#120) — maintainer @ReganBell suggested reopening as an issue, so here it is.

The pattern. QM currently schedules work time-driven (crons, monitors, wake). If QM ever needs to arbitrate resources across concurrent scopes (interactive user sessions vs. background crons vs. long-running watchers), a goal-oriented scheduler is an alternative worth having on the radar: declare goals per scope ("interactive < 2s", "batch crons stay in budget") instead of tuning per-job priorities, run a closed-loop controller, and let importance-based preemption arbitrate contention.

Prior art. IBM z/OS WLM is the canonical implementation of this (goal mode + service classes + importance). I have a Go prototype of the semantics for Linux — deeparchi-ai/wlm — built on cgroup v2 + PSI pressure signals: 34 tests, zero production use. Honest about the prototype-to-production gap; it exists to test the semantics, not to be a dependency.

What I'm asking. Not an implementation proposal — a design discussion:

1. Does QM's scheduler have a foreseeable cross-scope contention problem (e.g. an interactive request starving behind a heavy cron)?
2. Would a goal-mode layer on top of the existing scheduler (rather than replacing it) fit QM's architecture?
3. Is there interest in a written ADR on this, given adr: goal-oriented scheduling for background work #120 was closed?

Happy to write up the ADR in QM's format if there's appetite.