feat(hooks): support context injection

[tjb-tech]

Summary

Adds generic command-hook capabilities needed by external memory/context integrations:

• command hooks can receive the JSON hook payload on stdin via stdin_payload
• command hook stdout JSON is parsed for additional_context
• supports Claude-style hookSpecificOutput.additionalContext
• session_start hook context is appended to the runtime system prompt
• user_prompt_submit hook context is injected for the current turn only
• hook payloads now include richer fields such as cwd, session_id, model, permission_mode, tool_response, and last_assistant_message

This enables integrations such as claude-mem without hard-coding any specific memory backend into OpenHarness.

Validation

PYTHONPATH=src python -m pytest -q tests/test_hook_context_injection.py
PYTHONPATH=src python -m pytest -q

Results:

• 3 passed for the new hook context tests
• 1033 passed, 14 warnings for the full suite

Also validated manually with a real OpenHarness agent loop, a real Claude subscription model, a real grep tool call, and a claude-mem fork worker hook receiving platformSource=openharness observations.