Add agent-side rollout observability

[Glorf]

Summary

• add the opt-in ng_agent_observations contract for agent invocations, tool-call intervals, context-compaction events, model-call references, and explicit observation gaps
• add producers for Claude Code, Hermes, OpenClaw, and PinchBench, plus optional AnyTerminal sidecar transport
• preserve protocol response IDs in model-call records for exact (model_ref, response_id) joins
• propagate rollout correlation to Tau, SWE/OpenCode, GDPVal judge calls, and Stirrup
• exclude observation payloads from aggregate-metrics requests and document capability boundaries

Producer coverage

Producer	Evidence exposed
Claude Code	Transcript-derived invocation hierarchy, model-call references, artifact-derived tool intervals, and compaction markers
Hermes	Invocation hierarchy, executor tool timing, and compaction markers through the pinned Hermes integration; exact model-call ownership is unavailable
OpenClaw / PinchBench	Normalized root response items with explicit gaps for hierarchy, timing, compaction, and model-call ownership
AnyTerminal	Optional normalized sidecar bundle; otherwise an explicit availability gap

The full model request and response remain in CaptureStore. Agent observations are additive rollout evidence, not a canonical trajectory or training contract.

Compatibility

• emission remains gated by observability_enabled
• the disabled execution and output paths remain unchanged
• observation data does not replace the agent response, reward, token IDs, or log probabilities
• producer parsing failures covered here return explicit gaps instead of inferred data

Review status

This is a combined draft for cross-agent validation. Before it is ready to merge:

• preserve the existing self-HTTP /v1/responses boundary for observed Hermes and OpenClaw runs; the current observed path executes in-process to collect integration hooks
• complete real-agent validation across the target benchmark matrix
• split the shared contract and producer integrations into stacked PRs

Canonical trajectory assembly, model-call/agent-observation export joins, ATIF projection, Pi/OpenHands producers, and sandbox CPU/memory telemetry are not included.

Validation

• producer integration suites: 213 passed
• core contract, capture, rollout aggregation, GDPVal, and Stirrup suites: 194 passed
• Tau policy/user correlation: 2 passed
• SWE/OpenCode correlation and disabled serialization: 3 passed
• Ruff check and format checks passed
• staged diff and credential-pattern checks passed

Related to #1867.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[github-actions]

🌿 Preview your docs: https://nvidia-preview-feat-rollout-observability-p0.docs.buildwithfern.com/nemo/gym

Here are the markdown pages you've updated:

• nemo/gym/model-server/model-call-capture

[Glorf]

Superseded by the reviewed rollout-observability stack: #2114, #2115, #2116, #2117, #2118, #2119, and #2120.