Measure provider latency through response body consumption

[SamCT86]

Problem

The OpenAI adapter captured latencyMs immediately after fetch() returned, before response.json() consumed the response body. That could under-report provider latency even though the runtime later uses the measured value as an acceptance bound.

Failure-first evidence

A deterministic regression advanced the injected clock by 75 ms while consuming the response body. On current main, the test failed with 0 !== 75.

Change

• measure provider latency after the response body has been consumed
• add a focused regression test for delayed body consumption
• keep the existing provider timeout and cost accounting boundaries unchanged

Verification

Exact remote branch head was cloned and verified on STRATO:

• npm test: 28/28 passed
• npm run eval: 5/5 passed
• git diff --check: passed
• diff from current main: only src/runtime-gate.mjs plus the focused latency regression test

Boundary

This fixes provider-response latency accounting. It does not claim an end-to-end application SLO or distributed execution guarantee.