feat(policy): add bounded lifecycle hook bus

[PhilipJohnBasile]

Splits the provider-neutral lifecycle hook bus out of #336 as one isolated system.

The bus supports request, stream-event, response, and error phases with explicit allow, reject, rewrite, and annotate outcomes. Hook execution uses a fixed worker count, a bounded queue, per-hook timeouts, input and annotation size limits, and configurable fail-open or fail-closed behavior. No policy module is loaded from the environment and no external policy service is contacted.

Measurement on commit a6edbf9, macOS 27.0 arm64, Python 3.13.14, from scripts/bench_policy_hooks.py --requests 5000:

Hooks	p50	p95	p99	Throughput
0	3.458 us	3.750 us	5.666 us	267,004/s
1	19.208 us	30.542 us	58.042 us	46,044/s
4	62.125 us	81.208 us	148.333 us	15,207/s
16	236.083 us	389.125 us	559.667 us	3,904/s

Rewrite and reject arms each returned 5,000 correct outcomes. Under 32 concurrent callers with 2 callback workers and 2 pending slots, observed callback concurrency stayed at 2. Four calls timed out and 28 were rejected by the saturated executor. Caller p95 was 5.320 ms with a configured 5 ms hook timeout.

Validation:

• 7 passed in 0.10s
• ruff check passed
• compileall passed
• git diff --check passed

This PR does not wire the bus into the OpenAI server yet. That integration stays separate so this system can be reviewed and measured without coupling it to request capture, replay, telemetry, memory control, or dashboard work.