Searching the whole platform for a metric found none:

grep -rn "new Meter(" src/                  →  0 results
grep -rln "System.Diagnostics.Metrics" src/ →  0 results

Every number any investigation had ever used was a .NET runtime built-in standing in for what was actually wanted.

Why that was worse than inconvenient

One investigation asked why roughly 6,900 synchronisation hubs were holding 2.7 GB of a 3.5 GB heap. Nothing reported how many hubs were alive, how they split between kinds, or how many were still running versus already dead. The only way to find out was a heap dump against a live replica — and that suspends every thread for about 106 seconds, against a liveness budget of 90.

So the replica gets restarted, and the restart destroys the state being investigated. Worse, the freeze scales with heap size: the population only exists on a replica fat enough that measuring it guarantees losing it. A question that cannot be asked without destroying its own answer is not an expensive question — it is an unanswerable one.

It also meant a fix could not be checked. One change released about 1,500 dead hubs, and its author asked readers to expect the curve to flatten, not vanish — while nobody could see the curve at all. A merged fix whose effect cannot be observed is indistinguishable from one that did nothing.

What there is now

A meter, MeshWeaver.Platform, reporting the hubs alive in the mesh — tagged by the kind of address and by run level. Collectors pick it up automatically.

Two design choices worth stating, because they are what make the number trustworthy:

• It is measured, not counted. A counter has to be incremented and decremented at every place a hub appears or disappears; miss one and the number drifts from reality while still looking like a measurement. This walks the live tree when the collector asks, so it cannot disagree with what is actually there.
• The tag is the kind of address, never the address. One series per hub would mean thousands of them — the very population under investigation — and the meter would become the second thing nobody can afford to collect.

It is deliberately a first version: hub counts and nothing else. Streams, queues and pending callbacks each want their own owner and their own cost, and a meter that tries to answer everything at once is how instrumentation becomes the thing nobody trusts.