Merge main and move the Codex compatibility work onto main's tests/e2e/codex
suite, replacing the parallel tests/e2e/client_apps suite and its harness tests.
A makereport hook in codex/compat_results.py appends one JSON line per decided
cell when CODEX_COMPAT_RESULTS_PATH is set, and codex/matrix_builder.py builds a
table in the Claude Code matrix shape and exits 3 when a cell that passed no
longer does. The row id comes from the test file name and the label from the
class name, so a new row needs no other edit.
The suite now pins @openai/codex 0.166.0 in its own package.json and lockfile,
adds Gemini on Vertex AI as a setup, and adds prompt caching, reasoning effort,
vision, structured output and web search rows. Web search carries strict xfails
where Codex or LiteLLM leaves the search tool out, and auto-compaction now also
expects Vertex AI to fail, for the same bridge gap Claude on Bedrock has