Make coding-agent review loop proof deterministic

[bmdhodl]

Summary

• clear the coding-agent review-loop trace before each example run
• add regression coverage proving reruns do not append stale guard events
• keep the dogfood incident report aligned to a single proof run

Validation

• python -m pytest sdk/tests/test_example_starters.py -q
• python examples/coding_agent_review_loop.py
• python examples/coding_agent_review_loop.py
• python -m agentguard.cli incident coding_agent_review_loop_traces.jsonl
• rg -n "guard\.(budget_exceeded|retry_limit_exceeded)" coding_agent_review_loop_traces.jsonl

[Copilot]

Pull request overview

This PR makes the coding-agent review-loop example produce deterministic proof output by clearing its JSONL trace before each run and adds regression coverage for reruns.

Changes:

• Clears coding_agent_review_loop_traces.jsonl before the example writes new trace events.
• Adds a regression test that runs the example twice and verifies guard events are not duplicated.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated no comments.

File	Description
examples/coding_agent_review_loop.py	Removes any prior trace file before creating the JSONL sink.
sdk/tests/test_example_starters.py	Adds rerun coverage for the coding-agent review-loop trace output.