fix(codex): filter synthetic AGENTS startup messages

[voarsh2]

Summary

Filters Codex synthetic AGENTS.md startup messages out of the Codex integration transcript reader.

Codex can persist project startup instructions as a normal response_item with role: "user" and content shaped like:

# AGENTS.md instructions for ...
<INSTRUCTIONS>...</INSTRUCTIONS>

The Codex integration previously treated that as normal user conversation, so auto-retain and recall transcript composition could include agent rules/setup context instead of only the actual user/assistant session.

Why

This can pollute retained memory with hook rules, memory policy, MCP setup instructions, and other agent-control text that the user did not actually say as part of the working conversation.

In long-running Codex sessions, this can also contribute to oversized auto-retain payloads because the retained transcript includes synthetic setup context before the real conversation starts.

Implementation

• Adds a small detector for Codex AGENTS.md startup instruction messages.
• Skips those synthetic user messages in the text transcript reader.
• Skips the same synthetic messages in the rich transcript reader used when tool calls are retained.
• Leaves normal user discussion of AGENTS.md untouched.

Validation

• pytest hindsight-integrations/codex/tests/test_content.py
• pytest hindsight-integrations/codex/tests/test_hooks.py
• Tested against a real Codex rollout transcript where the AGENTS startup block was persisted as a user message. After this change, the startup block is no longer present in parsed transcript output or retained chunk content.

[voarsh2]

Also has some relevance to my issue #1345