[maestrod] Add Grafana dashboard (per-route AI token usage + RED + vision)

[olihou]

What

Adds a Grafana dashboard to the maestrod chart, mirroring the document-engine Grafana-sidecar pattern (the serviceMonitor already exists).

• dashboards/maestrod-single-namespace.json — 14 panels across four rows:
  • AI token usage — per-route token rate + range totals by route/model (nutrient.ai.tokens_total), MEAI token distribution, AI call latency p50/p95, attempts/empty/warnings. Token counts only — no cost/pricing applied.
  • Per-route HTTP RED — request rate, error ratio, p95 latency per /run/* route.
  • Vision quality & throughput — document throughput, words/zones per page, classification confidence, artifact export latency.
  • Process health — working-set memory + CPU only (no GC/JIT/threadpool noise).
• templates/monitoring/grafana-dashboard.ConfigMap.yaml — ConfigMap labeled grafana_dashboard: "1" for sidecar auto-discovery; .Files.Get + placeholder replace + indent 4.
• values.yaml / values.schema.json — observability.metrics.grafanaDashboard block (disabled by default): enabled, configMap.labels, title, tags.
• README.md — regenerated via helm-docs for the new values.
• Chart bump 0.6.2 → 0.7.0 + CHANGELOG entry.

Panel labels are written to be self-explanatory: every timeseries carries a descriptive y-axis label (tokens/s, latency (s), error ratio, requests/s, cores, count / page, …) and the per-route token table renames raw Prometheus labels to readable column headers (Route / Model / Token type / Tokens (range total)).

How to enable

1. observability.metrics.serviceMonitor.enabled=true (Prometheus scrapes /metrics).
2. observability.metrics.grafanaDashboard.enabled=true (renders the labeled ConfigMap).
3. A Grafana instance with the dashboard sidecar watching the grafana_dashboard label auto-imports it.

Validation

No helm binary on the authoring host — validated by parsing values.schema.json and simulating the ConfigMap's placeholder substitution (incl. the tags join): the rendered dashboard is valid JSON, all 14 panels reference real metric families emitted by maestrod's /metrics. README.md regenerated with helm-docs 1.14.2 so the generate CI check passes.

Notes

• The per-route token metric (nutrient.ai.tokens_total with an operation label) is added by the companion daemon PR: PSPDFKit/GdPicture#2804. The two per-route token panels stay empty until a daemon build carrying that change is deployed; the by-model MEAI token panel (gen_ai_client_token_usage) works today.
• Dashboard authored against schemaVersion 39; Grafana 12 auto-migrates on import. Not yet visually QA'd in a live Grafana — worth an import-and-tweak pass on layout if desired.

Companion

Daemon-side metric: PSPDFKit/GdPicture#2804.

🤖 Generated with Claude Code

[tomassurin]

observability.metrics.grafanaDashboard.title and tags are user-facing values, but values containing JSON-significant characters break the dashboard payload. A valid Helm value such as title: Maestrod "prod" or a tag containing " renders a ConfigMap whose JSON cannot be imported by the Grafana sidecar.

Evidence: the template inserts raw strings into JSON placeholders via replace/join, while the dashboard keeps those placeholders inside JSON string/array literals.

Fix: JSON-encode these values instead of raw replacement. For example, replace the whole tags literal with a toJson value, and render the title placeholder from an escaped/encoded string.

[lazyoldbear]

Changelog rewrites history?