Keep the telemetry emitter service off the modelling stack

DESIGN.md

@@ -256,8 +256,11 @@ never as namespace squatters.
 
 **Constellation versioning has a mechanism, not just an intent.** Each shard
 pins `microcosm-frame>=X,<X+1` AND asserts kernel compatibility at import (a
-cheap `frame.__version__` check) so pip's looser resolution can't silently
-assemble an incompatible set. CI builds every shard wheel once and compares
+cheap check of the installed frame version) so pip's looser resolution can't
+silently assemble an incompatible set. microcosm-build reads that version from
+the distribution metadata rather than `frame.__version__`, and loads its
+public names lazily, so importing one build submodule does not load the
+kernel's numpy and pandas stack. CI builds every shard wheel once and compares
 each archive's package contents with its source tree. Behavioral contracts run
 from the source checkout in separate environment-specific jobs; wheel validation
 does not install the artifacts or repeat behavioral tests. This validates archive

changelog.d/telemetry-emitter-startup.fixed.md

@@ -0,0 +1 @@
+Builds on a busy host no longer lose hosted telemetry because the local emitter service started slowly. `microcosm.build` now loads its public names on first use and reads the microcosm-frame version for its compatibility gate from distribution metadata, so the service no longer imports torch, pandas, numpy or the calibration stack, and the Hugging Face client loads only when the delivery worker first needs a token. On a host at load average 80–160 the service became ready in 0.4–0.8 s, down from 5.3–9.2 s. The build now waits up to 30 s for the service instead of 3 s, and the warning on giving up names the wait. The Modal environment and runner smoke probes star-import `microcosm.build`, so they still prove that the modelling stack imports.