docs: publish the cost model for a real deployment #298

Why

"Serverless and cheap, running inside your own AWS account" is the clearest differentiator Junando has against Keep, Grafana and Datadog — none of which can match it. But that claim is currently nowhere in the docs, so nobody evaluating the project can see it.

Content

A cost page with two worked scenarios (small app, large noisy app), broken down per component: Lambda, SQS, dedup store, LLM tokens. Show the free-tier headroom explicitly, and compare model choices (Nova Lite vs Claude Haiku).

Prerequisite — do not skip

The current estimates assume ~3,000 input tokens and ~500 output tokens per analysis, and a 3:1 alert-to-cluster correlation ratio. These are assumptions, not measurements. Before publishing:

• Instrument actual token counts from a real analysis with the current prompt
• Measure the real correlation ratio on a realistic alert stream

Publishing unverified numbers on the main differentiator would be worse than publishing nothing.

Done when

The docs site has a cost page backed by measured token counts, not estimates.