New models cache breaks older clients sharing CODEX_HOME

What happens

Codex 0.148 writes models_cache.json entries with canonical model_messages.instructions_template but without the deprecated top-level base_instructions. An older Codex 0.146 process sharing the same CODEX_HOME deserializes the complete model array before checking client_version, so it logs:

failed to load models cache: missing field `base_instructions`

It then falls back to a remote fetch, so functionality continues, but the warning recurs whenever the newer process refreshes the shared cache.

Source gap

ModelsResponse already uses both serialize_model_infos_with_legacy_base and deserialize_model_infos_with_legacy_base. ModelsCacheEntry.models uses only the deserializer, so the compatibility guarantee is lost specifically when persisting the file cache.

Proposed fix

Commit https://github.com/joshmouch/codex/commit/59449eb735 adds the existing legacy serializer to ModelsCacheEntry.models and a raw persisted-JSON regression.

Verification:

• cargo fmt -- --check
• cargo test -p codex-models-manager (49 passed)
• cargo clippy -p codex-models-manager --all-targets --all-features -- -D warnings

I can open the PR if a maintainer invites it under the repository contribution policy.

Start in the codex-models-manager code at ModelsCacheEntry.models and compare its deserialization with ModelsResponse's legacy serializer. Add the persisted-JSON regression described in the issue so older clients can read refreshed cache entries, then run cargo fmt -- --check, the package tests, and clippy with warnings denied.