bench: add hosted MCP workload proof

[lbliii]

Summary

Add the first mergeable slice of #229: an artifact-ready, real-Milo MCP workload plus Pounce-owned transport proof.

• extend the fixed-rate driver to rotate typed request variants with distinct bodies and headers
• add a modern tools/call workload alternating a pure-Python CPU tool and a blocking tool
• run the same CLI.asgi_app() under Pounce, Uvicorn, Hypercorn, or Granian through the existing Epic: Benchmark evidence program — sustained load, p99, published artifacts #228 artifact pipeline
• require the coordinated-omission-safe --rate path instead of silently sending a different request through wrk/hey
• prove real Pounce workers preserve MCP routing/auth header values and complete multi-event SSE framing
• preserve existing --workload all behavior unless the fixed-rate driver is selected

Progresses #229. This deliberately does not close the epic.

Dependency and deployment blockers

The workload uses Milo's public CLI.asgi_app() from milo-cli#127. Until that draft stack merges and releases, install that branch in the benchmark environment; the benchmark app fails with an actionable message on older Milo releases.

The existing Railway pounce-railway-smoke service is healthy and remains untouched. #229's Milo MCP deployment should be a separate service rather than repurposing the production canary; provisioning that service is intentionally deferred pending explicit infrastructure/cost approval and the Milo release.

Milo #107 still owns the end-user Railway/Cloud Run/Fly recipes. Pounce #248 already supplies the free-threaded Railway $PORT/healthcheck base those recipes can copy.

Proof

• hosted PR matrix: every required lane passed, including Linux/macOS 3.14t, coverage, process workers, framework compatibility, Railway recipe container, types/lint, and issue-closure acceptance
• full suite: 2,425 passed, 7 skipped, 10 benchmark tests deselected
• focused issue suite: 44 passed
• Ruff lint and format: clean
• silent-exception, raise-message, and import-boundary gates: clean
• ty check src/pounce/: clean
• Python 3.14t with PYTHON_GIL=0: the local Milo Investigate and add a measured sendfile minimum-size threshold so tiny static files skip transport-detach overhead #127 app completed the alternating workload under both Pounce and Uvicorn with no request errors
• emitted smoke artifacts validated against benchmarks/artifact-schema.json; no local number is promoted as a public claim

Performance note

No Pounce parser, worker, scheduler, framing, or request-pipeline hot path changed. The implementation is benchmark-driver and fixture code only, so no before/after runtime benchmark is required. Short local runs validated executability and artifact shape; publishable comparison numbers still require the repeated hosted 3.14/process and 3.14t/thread matrix described by #229.

Steward notes

• Performance evidence: alternating bodies/headers retain fixed-rate scheduling, p50/p99/p999, raw output, telemetry, and schema-compatible artifacts.
• ASGI/protocol: real-worker tests verify header values and ordered multi-event SSE bytes without changing the HTTP implementation.
• Security/exposure: the benchmark uses an explicit public fixture token and exact Origin allowlist; no real credential is stored.
• Free-threading: the CPU handler is pure Python and the blocking handler models synchronous external work; the same app/request plan is used for every server comparison.
• Public claims: README names the upstream draft dependency and forbids treating pre-release/local runs as product evidence.
• Collateral: benchmark docs, issue-linked unit/integration proof, and changelog fragment move together; no runtime API or config collateral is needed.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.