Host ROLES: the matrix beast vs the native-only arm oracles

• Type: feature (Track T topology) — Track T
• Opened: 2026-08-01, recording an intended topology that is currently undocumented.

Intended topology

The point of the rPis is native vs QEMU differential: xeon already runs the arm targets under emulation (CLASS_WEIGHT has a qemu class), and a real ARM box says whether a QEMU result is the truth or an emulator artifact. Their job is not to grind full suites all day — they are slow, and a full matrix on a Pi would take longer than the answer is worth.

What already exists (do not rebuild it)

• trackt setup has a first-run role-profile wizard with three profiles — dedicated / limited / restricted (tools/trackt.py:588-660) — writing max_cores and max_mem_mb into the clone's twatch.conf.
• Per-host conf already carries tier and fast_tier, and the full phase accepts none, so "fast phase only" is expressible today.
• Per-host tstate state and reports are already keyed by host.

Naming note: the launcher is trackt (repo root) + tools/trackt.py. There is no trackt.sh — worth stating because it gets referred to that way.

What is missing

1. Profiles express RESOURCE CEILINGS, not PURPOSE. restricted means "use half the cores and a fraction of RAM" — an rPi set to restricted still works through a full tier, just slowly, which is precisely the behaviour not wanted. A role needs to say which job classes this box exists to run (e.g. "native aarch64 jobs only; never corpus, never fuzz, never opt"), not merely how hard to push. Suggested shape: a jobs/classes allowlist in the conf, plus a native-oracle profile that sets it.

2. Nothing compares two hosts. tstate is per-host, and a verdict is published per host+sha. The entire value of a native ARM box is host A native vs host B QEMU on the same sha+job — that comparison does not exist. Without it a Pi just re-reports what xeon already said, and the differential has to be done by eye.

3. The topology is undocumented, which is why this ticket exists. It belongs in devdocs/dev/ alongside the Track T notes once settled.

Landmine this topology walks straight into

The rPis will run a different distro again (Raspberry Pi OS vs Ubuntu 26.04 vs the dev box's 24.04). Cross-distro assertion coupling has already produced one permanent phantom RED that cost several mis-diagnoses — see [[bug-t-host-dependent-test-assertions-cross-distro]]. Adding two more distros multiplies that surface, so that ticket should land before the Pis are enrolled, not after.

Gate

An arm64 rPi enrolled with a native-oracle role runs only its native job set, never a full tier, and its results are comparable against xeon's QEMU verdicts for the same sha without hand work.

Log

• 2026-08-03 — moved to rainy-day: the arm oracles are aspirational with no date and no hardware on hand (user, [[decide-t-queue-scope-2026-08-03]]). Kept as the design record for the topology — it is the thing to read when the boxes do arrive, and [[bug-t-host-dependent-test-assertions-cross-distro]] already installed the triage rule that makes a third distro cheap. Anything in here that pays off on xeon ALONE was split out rather than parked with it.