fix(part20): replace fabricated config blocks with real Hermes schemas

[OnlyTerp]

Summary

Closes #13. The reporter flagged that the model_routing: / intent: / complexity: config block in part20-observability.md ("Rule 1: Route by Task Complexity") doesn't exist in Hermes. They're right — and on audit, that's the tip of the iceberg. The previous rev of part20 invented a stack of YAML blocks the Hermes loader does not read.

Verified against Hermes v0.14.0 source (/home/terp/.hermes/hermes-agent, commit cea87d913), the upstream cli-config.yaml.example, and the bundled plugins/observability/langfuse plugin.

Fabricated keys removed → real ones used

Old (invented)	New (real)
model_routing: with routes / intent / complexity / match / tokens_gte	auxiliary: per-task models (vision, web_extract, compression, summarization) + provider_routing: for OpenRouter + hermes fallback chain
prompt_caching: with enabled / providers / cache_system_prompt / cache_skills / cache_memory_digest / min_cache_tokens	One-key prompt_caching.cache_ttl ("5m" or "1h"), plus the separate openrouter.response_cache layer
telemetry: spans: browser_cdp: true / computer_use: true	Grep hermes logs tail -f --level WARNING or use a Langfuse alert on the existing browser_use / computer_use tool spans
fast_mode: with per-gateway defaults + user_override	/fast on|off|status slash command — it's a runtime toggle, no YAML block
compression.auto.at_tokens / preserve.tool_results_matching / topics_from	Real keys: compression.{enabled, threshold, target_ratio, protect_last_n}
alerts: cost_spike: / token_anomaly: block	Alerts live in the tracing backend (Langfuse / Helicone / Phoenix) or in a small log-tailer that fires through hermes send telegram
observability: langfuse: block in config.yaml	HERMES_LANGFUSE_PUBLIC_KEY / HERMES_LANGFUSE_SECRET_KEY / HERMES_LANGFUSE_BASE_URL / HERMES_LANGFUSE_SAMPLE_RATE / etc. in ~/.hermes/.env, matching the bundled plugin's README and __init__.py
observability: otel: block with endpoint / protocol / headers / attributes	Standard OTEL_EXPORTER_OTLP_* env vars in ~/.hermes/.env
hermes evals init / dataset / run / compare subcommands	Langfuse Datasets & Experiments workflow (Hermes has no built-in evals subcommand)

Heads-up callout for smart_model_routing

Added a paragraph at the top of the "Cost Routing Playbook" noting that the older smart_model_routing: block (cheap-vs-strong by message length) was removed upstream in commit 424e9f36b ("refactor: remove smart_model_routing feature", #12732, 2026-04-19). Readers who saw it referenced in older Hermes docs or upstream issue #16211 won't try to bring it back.

Verification

• Hermes v0.14.0 installed locally; ~/.hermes/hermes-agent/hermes_cli/config.py is the loader of record.
• grep -rn smart_model_routing /home/terp/.hermes/hermes-agent --include="*.py" returns zero hits in current source (only the removal commit in git log).
• ~/.hermes/hermes-agent/plugins/observability/langfuse/{plugin.yaml,README.md,__init__.py} confirms the env-var contract.
• cli-config.yaml.example confirms the real prompt_caching, compression, openrouter, and provider_routing shapes.
• git diff --check clean; only part20-observability.md (rewrite of Rules 1/2/2B/3/4/5 + Langfuse + OTel + eval sections) and CHANGELOG.md (dated entry) changed.

Fixes #13.

---

[devin-ai-integration]

✅ Devin Review: No Issues Found

Devin Review analyzed this PR and found no potential bugs to report.

View in Devin Review to see 3 additional findings.