test(server): add tool-state isolation smoke harness

[kira-ariaki]

Summary

• add a standard-library tool-call state smoke harness for sequential shared-prefix and concurrent requests
• validate exact function names, complete JSON argument objects, sentinel identity, and finish reasons
• add engine-free unit/integration tests with a local fake HTTP server
• make the stronger check opt-in from the manual Server smoke workflow
• run benchmark-tool unit tests in CI
• update the manual workflow to the current lowercase release asset and basert-serve CLI names

Why

A one-shot chat completion cannot catch cross-request tool argument contamination. This harness sends distinct sentinels through a repeated tool schema and reports exact mismatches as structured FAIL records.

Failures are consistent with request-state leakage, but the harness remains diagnostic and does not claim an engine root cause by itself.

Refs #26.

Verification

• python3 -m unittest discover -s benchmarks/tests -v — 17/17 passed
• ruff check benchmarks/scripts/tool_state_smoke.py benchmarks/tests/test_tool_state_smoke.py
• Python compile checks and workflow YAML parsing passed
• live against BaseRT 0.1.7 + Qwen3.6-35B-A3B: correctly exits nonzero and identifies cross-request sentinel corruption in sequential, N=2, and N=4 cases
• exact-argument tests reject missing or leaked extra fields; CLI tests reject malformed URLs and non-finite timeouts