[FEATURE] SLO contracts on InferenceService + SLO-aware routing and scheduling

[Defilan]

What

Make SLOs a first-class, enforced concern, not just a dashboard.

• Declare SLO targets per InferenceService (e.g. ttft_p99_ms, error budget, min_throughput_tok_s). Use Pyrra (feat(observability): per-InferenceService SLO declaration via Pyrra integration #415) to generate the Prometheus recording + multi-window burn-rate alerting rules from the declared SLO. The operator surfaces an evaluated SLOBreached condition on InferenceService status.
• ModelRouter circuit breaker consumes SLOBreached: a breached backend is treated as degraded and traffic redistributes. Today the breaker is HTTP-health only, so a slow-but-alive backend keeps getting traffic.
• Foreman scheduler defers dispatching new AgenticTasks to a fleet/InferenceService that is SLO-breached (load-shed rather than pile on a degraded backend).

Why

This is the production-readiness piece for the B200 + edge deployment: SLO breach causes automatic traffic/work redistribution without a human reading Grafana. It turns SLOs from observation into enforcement.

Approach / dependencies

• Depends on feat(metrics): expose TTFT and per-request error rate; enable vLLM /metrics #409 (the TTFT + error-rate metrics SLOs are computed from) and feat(observability): per-InferenceService SLO declaration via Pyrra integration #415 (Pyrra declaration + alerting). Both M2.
• Couples into Epic: ModelRouter CRD for cross-engine agentic handoff #437 (ModelRouter) for the circuit-breaker change.
• Supersedes the enforcement intent of Advanced SLO auto-remediation #10; Advanced SLO auto-remediation #10's auto-model-downgrade remediation is deferred as an opt-in last resort.

Definition of done

SLO targets declarable per InferenceService; SLOBreached condition surfaced from Pyrra-evaluated state; ModelRouter redistributes off breached backends; Foreman scheduler defers dispatch on breach.