feat(observability): attribute token cost, budgets, and kill-switch decisions

[robinbraemer]

Outcome

Attribute tokens, cached tokens, modeled cost, budget consumption, rate limits, and kill-switch decisions to AgentOS operations without adopting content-bearing request logs.

Scope

• Reconcile Agentgateway's model-cost catalog and native token measures with the AgentOS telemetry contract.
• Export bounded aggregate metrics and protected per-operation spans for Mate/Assignment attribution.
• Distinguish gateway rejection, AgentOS budget exhaustion, provider 429, and unknown/unpriced models.
• Define pricing-version, currency, estimation/actual flags, and cache-token semantics.
• Keep authoritative budget counters outside the fail-open telemetry path.

Acceptance criteria

• Operators can answer which Assignment class, runtime, route, model family, and request kind consumed cost without metric-label cardinality leaks.
• Missing or stale pricing is explicit and never reported as zero actual cost.
• Collector/exporter outage does not reset or bypass budgets.
• Agentgateway SQLite request logs and prompt logging are unnecessary for the supported dashboards.
• Synthetic accounting reconciles requests, tokens, cache tokens, and cost within documented tolerances.

Reference

• https://agentgateway.dev/blog/2026-06-24-agentgateway-cost-tokenomics-dashboard/

[robinbraemer]

PR #156 established the authoritative non-telemetry source for #117: durable request/token/spend counters and reservations, bounded settlement fields, surgical override audit, and distinct AgentOS rate_limited/budget_exhausted envelopes. Dynamic identities are not metric labels and no credentials/provider payloads are stored. #117 remains open for bounded aggregate metrics, protected per-operation spans, modeled cost/catalog reconciliation, dashboards, alerts, and telemetry-failure conformance.