test(doctor): assert the hanging-backend property instead of a stopwatch

[AetherAI3]

Problem

test/diagnostics.test.ts — "a hanging backend cannot stall the fast report" fails intermittently on loaded runners.

It injects clients that never resolve, sets timeoutMs: 50, and then asserts:

assert.ok(Date.now() - started < 500);

That measures the machine, not the behaviour. The 50ms timeout can fire exactly as designed and the assertion still fails because the runner was busy.

Observed:

where	elapsed	outcome
windows-latest CI	689ms	✖ fail
local, full-suite load	1818ms	✖ fail
local, in isolation	82ms	✔ pass

This is on main, so it can redden any branch that runs at the wrong moment. It has already done so once.

Contract

Assert the property the test exists to defend, not the clock: a probe fed by a hanging client must never come back claiming it verified anything.

agent.transport, auth.credential, agent.catalog and mcp.broker are named explicitly rather than filtered on an axis. A first attempt filtered on reachable !== "na" and failed on workspace.git — a local check that legitimately verifies in this fixture, because nothing about it touches the backend that is hanging.

The wall clock stays, but as a hang detector rather than a stopwatch. The failure this test guards against is an unbounded await that never returns at all; 30s separates that from a slow runner, and any finite margin would do. A busy machine is no longer a bug.

This is stronger than what it replaces, not weaker

Relaxing a timing bound looks like weakening a test, so here is the evidence it is not.

Mutation: make notChecked() return a verified axis — the exact false-green this suite exists to prevent.

	result
mutated	5 pass / 3 fail — this test among them
restored	8 / 8

The old assertion passed that mutation untouched. It never looked at a single axis. So the previous test would let a false green through while failing on a busy runner — precisely backwards.

Tests

command	result
npm run typecheck	exit 0
npm test	922 pass / 0 fail

922 is the full suite on main at 41a7e261; this PR changes no production code and adds no test, so the count is unchanged by design.

Scope

One file, one test. No production code. git diff --stat: test/diagnostics.test.ts | 27 +++++++++-.

Why this is separate

It surfaced while working an unrelated lane, where it turned windows-latest red on a PR that had not touched anything near it. Fixing it there would have buried an independent main-line defect inside an unrelated change.

Recommend landing this first, ahead of the supercluster lanes (#72, #73, #74, #75, #77, #78) — every one of them carries this flake until it does.

The same fix is currently also committed on supercluster/a5-usage-continuity because that is the branch it was blocking; that copy should be dropped once this lands.