[yuechao-qin]

What

The status-transition histogram carries only the two statuses, so every wait looks the
same. These five labels separate the ones that are ours from the ones that are not:

• a GPU family is scarce right now, versus the orchestrator being slow for everyone
• one tenant's cluster is not giving machines back, versus our queue being deep
• one namespace is queueing, versus all of them

Why

An alert that cannot tell those apart either pages for someone else's capacity problem
or stays silent through ours.

	Label on the histogram	Values
today	execution.status.from	the status the execution left
today	execution.status.to	the status it reached
today	execution.run_attempt	first, repeat, none
new	execution.cloud_provider	google, nebius, davies, none
new	execution.cluster	us-ce1-lt3, b300-cluster, us-ce1-eo9, none
new	execution.accelerator_type	NVIDIA-B300-SXM6-PC, B300-SXM6-PC, none, unparseable
new	execution.accelerator_count	0, 1, 3, 4, 8
new	k8s.namespace.name	about 42 distinct in a day, and the set grows

Changes

Every label is resolved, not copied, and nothing is ever omitted.

Label	Where the value comes from
execution.cloud_provider	the registered target map, so an alias is not a second platform
execution.cluster	the same map; the annotation names a target, not a cluster
execution.accelerator_type	the parsed annotation — 8 raw strings in production are 2 models
execution.accelerator_count	the same parse; a malformed annotation degrades the labels, not the measurement
k8s.namespace.name	the task's own service account, else the cluster's namespace

An absent or unroutable value is labelled "none" rather than left off. An attribute
that disappears forms a second series that no filter on the first one covers, which is
the failure these labels exist to prevent.

Target gains a namespace, so a task without its own service account still reports
the namespace it actually ran in — about half of all traffic.

Nothing queries. Every value is read off task_spec, a loaded column, because this runs
inside before_commit where a lazy load would put a SELECT in the commit path.

Tests

• each label resolved from a registered map, and from an alias, to the same series
• the accelerator annotation parsed from all 7 shapes production writes, including the
  two that differ by one space and the two that cannot be parsed at all
• a namespace from a service account, from a cluster, and from neither
• the exhaustive attribute assertion lists all eight, so a new label cannot be added
  without a test naming it