[epic] Rollout observability & reliability fast-follows

[sephmard]

Program role

This epic owns rollout observability, correlation, capture fidelity, live operational inspection, and exported telemetry. It observes runtime contracts defined by reliability program #2750; it does not own retry, masking, cleanup, cohort, or lifecycle policy.

Ownership

• Rollout, invocation, model-call, and diagnostic environment-session correlation.
• Capture fidelity and completeness.
• Shared observability conventions and exported signals.
• feat: live visibility into in-progress rollouts #2859 — user-facing live rollout inspection and stalled-work presentation.
• feat(observability): standardize error-recovery telemetry #3182 — error-recovery metrics, spans, and minimum structured-log convention.

Relationship to #2750

#2750 is the reliability program and defines policy-neutral runtime facts. This epic consumes stable fields as they land.

Neither live visibility nor comprehensive exported telemetry blocks #2750 correctness or completion. Basic bounded logs required to operate each fix remain in the producing issue; shared metric and trace conventions live here.

Priority

• Incrementally instrument already-stable behavior.
• Add feat: live visibility into in-progress rollouts #2859 after feat(core): define typed rollout outcomes and recovery #2135 provides stable attempt identity and feat(core): own startup, shutdown, and cancellation #3160 provides progress state.
• Add producer-specific feat(observability): standardize error-recovery telemetry #3182 telemetry as each Make rollout failures bounded, visible, and recoverable #2750 contract stabilizes.

Already landed

#2192 completed capture cleanup and Anthropic SSE capture. #2646, #2647, and #2790 added OpenTelemetry foundations, cross-process spans, rollout and verify timing, active-server gauges, and experiment coverage metrics.

Completion criteria

• Runtime policy remains owned by Make rollout failures bounded, visible, and recoverable #2750 rather than duplicated here.
• High-cardinality identifiers are available in logs/traces but never metric labels.
• feat: live visibility into in-progress rollouts #2859 distinguishes active, slow, retrying, stalled, cancelled, and terminal work through bounded redacted output.
• feat(observability): standardize error-recovery telemetry #3182 defines low-cardinality error-recovery telemetry and a useful structured-log fallback.
• Instrumentation can land incrementally without turning observability into a reliability merge gate.