The metric registry says what each instrument measures, and the
observability guide says how telemetry is collected, but neither takes an
operator from a symptom to a cause. A user reporting a slow request or a
503 had no path through the data.

These three guides each start at a symptom and end at a component. Every
step gives a query, says how to read the result, and names the next step
or the guide to switch to. The shared index holds what all three need:
the two spellings of the instrument names and which ingest path produces
each, the four reasons a query returns nothing, and the pointer to the
subsystems that emit no metrics at all.

Every query was run against a live cluster, and the readings drove the
text: the resume label is only meaningful next to the outcome, a fleet
gauge that every replica reports needs max rather than sum, and a phase
histogram has to exclude failures before its quantile means anything.

Review found decision-tree arrows that send an operator to the wrong
component, and queries that contradict the paragraph beside them.

Routing. The router maps a shed request and an exhausted park budget onto
the same no_capacity outcome, so that outcome can no longer send the
reader straight to the capacity guide; step 3 now splits the two on
parking.rejected, which is the only signal that separates them. The
timeout row claimed an ended park budget, which never classifies that
way: an ended budget surfaces the condition that blocked the resume. The
capacity guide sent the common path to "make spec.replicas larger"
without passing the step that asks whether the workers are doing any
work.

Hidden capacity has two causes the guides did not name: a paused actor is
pinned to the node holding its snapshot and reports required_nodes rather
than selector, and the idle worker count includes draining workers that
the scheduler will never pick.

Queries. The CPU rate summed across the stats source that the text below
it says never to add together. The snapshot-size quantile dropped the
file name that the next line says to compare on. The step 1 restore
quantile counted failed phases, which hold their timers, while step 2
excludes them. The desired-minus-ready subtraction used a bare on() join
that fails once a second replica reports the same pool.

A paused restore was also documented as never downloading, which is
untrue for the data_on_golden scope: the golden files still come from
object storage.