fix(codex): preserve UTF-8 in stop summaries

[valbarko]

Summary

• replace byte-based truncation in the Codex stop hook with Unicode-safe character truncation
• sanitize fallback summary output before appending to memory markdown
• add a regression test that runs the stop worker with long Cyrillic text and reads the resulting memory file as UTF-8

Verification

• bash -n plugins/codex/hooks/stop.sh
• uv run pytest tests/test_codex_stop_hook_utf8.py tests/test_codex_parse_rollout.py
• uv run ruff check tests/test_codex_stop_hook_utf8.py
• git diff --check

[googs1025]

I hit this issue locally with the Codex hook as well. The byte-oriented truncation path can leave invalid UTF-8 in the markdown memory file, and then later
memsearch index .memsearch/memory reports a UnicodeDecodeError when reading that file.

[googs1025]

This PR looks like the right source-side fix for Codex:

• character-based truncation avoids splitting multi-byte UTF-8 sequences
• the regression test exercises the worker fallback path with non-ASCII text