Publish the Kubernetes vs scale-to-zero PaaS crossover analysis

[FullFran]

The honest counterpart to the rest of this milestone.

At low volume, Kubernetes is not the cheapest or simplest way to run FeedPulse.
Cloud Run, Container Apps or the existing Dokploy setup are. Claiming otherwise
would be marketing.

This issue is about finding and publishing the point where that stops being true.

Scope

• Deploy the same commit on a scale-to-zero PaaS as the comparison baseline
• Run benchmark:stages:mvp (100 / 1000 / 5000 / 10000 feeds) against both
• Measure, per stage and per platform: monthly cost, end-to-end alert latency
  percentiles, backlog drain time, and behaviour on a cold start
• Identify the crossover: the feed count and burstiness above which the
  Kubernetes deployment wins on cost, latency or operability
• Publish the numbers in the README, including the range where Kubernetes
  loses
• State the operational cost honestly: a cluster has to be upgraded, patched
  and paid for whether or not anyone uses it

Why this matters

snow-mcrt publishes the case where its own GPU backend is twelve times slower.
That honesty is what makes the rest of its numbers believable. Same principle
here: a README that only shows where the chosen architecture wins teaches the
reader nothing and convinces no one who has run infrastructure before.