fix: interceptor sync bridge + async detection, example fixes

[saurabhjain1592]

Summary

Hostile-testing sweep TEST-SDK ahead of the BukuWarung integration — tracking getaxonflow/axonflow-enterprise#2861.

• All 5 provider interceptors crashed inside a running event loop (async-adapter-bypass class). The sync wrap paths (openai, anthropic, bedrock, gemini, ollama) ran governance via loop.run_until_complete → RuntimeError: This event loop is already running in FastAPI/Jupyter/any async app driving a sync provider client. Observed live: examples/openai_integration.py crashed with coroutine 'AxonFlow.proxy_llm_call' was never awaited. New shared run_coroutine_sync bridge (mirrors SyncAxonFlow._run_sync, but on ONE persistent daemon-thread loop — a throwaway-loop bridge breaks the second call with "Event loop is closed" because the httpx pool is loop-affine) blocks the call site until the verdict is in.
• AsyncOpenAI clients were silently given the SYNC wrap path. openai>=1/anthropic decorate async create with a plain-def @required_args wrapper, so asyncio.iscoroutinefunction returns False. is_async_callable follows __wrapped__ — verified against real openai 2.44 (async detected, sync stays sync).
• Examples pass on enterprise (JWT-validating) stacks. quickstart, gateway_mode, openai_integration read AXONFLOW_USER_TOKEN instead of hardcoded non-JWT literals (they 401'd on DEPLOYMENT_MODE=enterprise). gateway_mode's block demo now uses stacked SQLi (blocked on every posture; PII defaults to redact) and exits non-zero on unexpected approval; openai_integration prints its block-probe outcome instead of ending silently.

Version 8.5.0 → 8.5.1 + CHANGELOG.

Testing

• New TestSyncWrapInsideRunningLoop (6 tests, one per provider + allowed path) — mutation-verified: fails with the pre-fix openai interceptor restored. New TestAsyncDetectionThroughDecorators.
• New runtime-e2e/interceptor_sync_bridge/ against a LIVE enterprise stack (v9.6.1): REAL openai.OpenAI client (dummy key — provider never reached) wrapped and driven from inside asyncio.run AND from plain sync code → stacked-SQLi prompt blocked with PolicyViolationError both times, no RuntimeError; AsyncOpenAI detected async, wrapped create is a coroutine function, blocked verdict awaited on the caller's loop — PASS 5/5. (The plain-sync leg caught the "Event loop is closed" pool-affinity regression during development; the persistent-loop bridge fixes it.)
• All 7 examples re-run green against the live enterprise stack (quickstart, gateway_mode, openai_integration incl. real governed OpenAI round-trip, explain_decision, indonesia_compliance, list_decisions, wcp_retry_idempotency).
• Full pytest: 1047 passed, 29 skipped. ruff clean.

Related Issues

• getaxonflow/axonflow-enterprise#2861

[saurabhjain1592]

The one red check (Live integration vs try.getaxonflow.com) is pre-existing: the same workflow has failed on main's scheduled runs since 2026-07-07 (three consecutive days), with error=LLM routing failed + SQLi arriving unblocked from the deployed community-saas service — server-side state on try.getaxonflow.com, untouched by this diff. Likely the deployed twin of getaxonflow/axonflow-enterprise#2871 (retired claude-sonnet-4-20250514 fallback); flagged there. Everything else is green (18/18).