Add hosted agent runtime observability

[swombat]

Summary

• record versioned, trigger-local Chaos usage with separate ordinary input, cache creation, cache reads, output, and provider request counts
• capture fresh/resume/roll/fallback lifecycle decisions, prompt sizes, runtime identity, and content-hash identity changes in the runtime shim
• persist forward-only telemetry without fabricating historical values
• add UTC-window usage reports and an admin-only logical-session timeline with Chaos process transitions
• pin the hosted runtime image to the instrumented Chaos fork commit while upstream PR Bump puma from 6.6.0 to 7.0.0 #14 is reviewed

Why

Prompt caching is working during stable resumes, but current rows cannot explain provider cache-write costs, provider call multiplication, or session churn. This instrumentation separates HelixKit interactions, logical sessions, Chaos processes, and provider requests so future cost changes can be chosen from recorded evidence.

Tests

• bin/rails test — 2019 runs, 8759 assertions, 0 failures
• focused observability suite — 52 runs, 380 assertions, 0 failures
• yarn test:unit --run — 41 tests passed
• bin/vite build — passed
• production-shaped agent-runtime Docker build — passed
• built runtime health smoke test — HTTP 200 with pinned Chaos version
• focused RuboCop — passed
• changed Svelte files pass Prettier

Repository-wide Prettier still reports six pre-existing unrelated files.