docs: explain what Cortex shows and why it is worth running — interpreting the metrics and the value case

[huang195]

Feature Description

The docs will say how to install Cortex and how to drive it, but not what to do with what you
see. Someone looking at a session's token counts, its cache-read versus cache-write split, a cost
figure and latency percentiles has no page telling them what any of it means, which numbers
matter, or why leaving Cortex running is worth it.

This backs exit criterion 5 of the v0.9.0 release: a new user can self-serve on a typical
workstation, using the docs alone.

Proposed Solution

• What Cortex is for, readable before installing: what it shows you that you cannot
  otherwise see
• The value case in concrete terms — what decisions the numbers support: spotting an
  expensive prompt, seeing whether caching is working, catching a runaway session, deciding
  whether pruning is worth enabling
• How to read a session: what each token category means, and why cache-read and cache-write
  are priced differently rather than being one number
• How to read cost: what is included, what is estimated, and where the rates come from
• How to read latency: time to first token versus total response time, and what the
  percentiles actually tell you
• How to read the tool-pruning savings figure, including what a zero means for an agent with
  no tool inventory
• A worked example on a real session, with real output rather than paraphrase
• Verified by someone outside the team reading only the docs

Additional Context

Part of the Cortex v0.9.0 release. Complements #959 (the first-run path — how to get there) and
#960 (troubleshooting — what to do when it breaks); this issue is what the results mean once you
are there, and why they are worth having.

Depends on the metrics work landing: #950, #951, #952.