feat: add full-response ITL and output throughput

[cquil11]

Summary

• add full_decode_duration, measured from first parsed content through explicit HTTP response completion
• add full_response_inter_token_latency = full_decode_duration / (OSL - 1)
• add full_response_output_token_throughput_per_user = 1 / full_response_inter_token_latency
• export and document both metrics while preserving the existing parsed-content metrics

Why

A structured-output parser can stop emitting content before the engine finishes generation. With server token counting enabled, the existing ITL can then divide a parsed-content duration by the server raw-token count, mixing two different intervals.

The full-response pair keeps the client-observed duration aligned with the server-reported token count. In the motivating AgentX request, 26,571 completion tokens and a 145.861 second request lifecycle produce about 182.82 tokens per second instead of an impossible 326,009 tokens per second.

These are client-observed full-response metrics, not exact engine token timestamps. Aggregate latency and throughput percentiles also retain their normal reciprocal-distribution semantics: 1 / p75(ITL) describes the slow tail, while p75(TPS) describes the fast side.

Validation

• targeted metric, registry, and JSON exporter tests: 50 passed
• full metrics/property/exporter suite from the preceding commit: 833 passed
• uv run pre-commit run --all-files: passed

Follow-up and merge order

1. This PR
2. feat(agentx): aggregate full-response ITL and interactivity InferenceX#2504
3. fix(agentx): use full-response interactivity InferenceX-app#677

[github-actions]

Try out this PR

Quick install:

pip install --upgrade --force-reinstall git+https://github.com/ai-dynamo/aiperf.git@8128d08e7cfa31c36254f3c3d8c167488d51dde0

Recommended with virtual environment (using uv):

uv venv --python 3.12 && source .venv/bin/activate
uv pip install --upgrade --force-reinstall git+https://github.com/ai-dynamo/aiperf.git@8128d08e7cfa31c36254f3c3d8c167488d51dde0

---

Last updated for commit: 8128d08 • Browse code

[github-actions]

Fern Docs Preview: generation failed — see the Actions log for details. This does not block merge; ask a maintainer to retry if needed.