fix(server-metrics): parse OpenMetrics exposition for the vLLM Rust frontend

[ajcasagrande]

Cherry-pick of upstream ai-dynamo#1060 (commit 8ba75f9a9e06377c0569f36f6d0a966d5307135a, authored by @majunze2001 — authorship preserved on the commit) onto cquil11/aiperf-agentx-v1.0.

What it does

vLLM's Rust frontend serves /metrics as application/openmetrics-text. The classic Prometheus parser mistypes OpenMetrics counters (family becomes an unknown named foo_total), and OpenMetrics _created samples share the _total labels so the creation timestamp overwrites the real counter value in the last-value de-dup.

• FetchResult now carries the response Content-Type; the collector routes application/openmetrics-text bodies to the OpenMetrics parser, falling back to the classic parser if the stricter parser rejects the body (e.g. missing # EOF).
• _created samples are skipped in simple-family processing so they can't clobber _total values.

Adaptations for this branch (semantic merge)

The source commit's context includes code this branch doesn't have; conflicts were resolved to take only the OpenMetrics change:

• Did not import the source branch's lazy create_tcp_connector __getattr__ machinery (unrelated circular-import fix); this branch's eager import stays.
• Dropped _warn_dropped_non_finite/_warn_construction_failures (warn-once infra that doesn't exist on this branch); kept _parse_families.
• Grafted the _created skip onto this branch's simpler None/non-finite sample filter.
• Dropped the tools/ergonomics_baseline.json hunk: data_collector.py is 427 lines here (< 500 threshold), so the file-size allowlist entry would only mask a future real violation.

Verification

• Full unit suite: 11,894 passed; the 11 failures are pre-existing weka/mmap-cache drift on the base branch (files untouched by this change).
• All pre-commit hooks passed on the commit.
• Live end-to-end check driving _fetch_metrics_text() -> _parse_metrics_to_records() against a local HTTP server: OpenMetrics content-type routes to the OpenMetrics parser (counter typed correctly, _created skipped), classic text/plain parsing unchanged.

🤖 Generated with Claude Code

[github-actions]

Try out this PR

Quick install:

pip install --upgrade --force-reinstall git+https://github.com/ai-dynamo/aiperf.git@5f55d1719db64234c0d69e4dad99d46b01a27a0d

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@5f55d1719db64234c0d69e4dad99d46b01a27a0d

---

Last updated for commit: 5f55d17 • Browse code

[ajcasagrande]

@majunze2001 Please confirm this cherry-pick is faithful to original