fix codex-core late-suite failures in remote model and search-tool flows

[dzianisv]

Summary

During package-level validation for issue #12, cargo test -p codex-core -- --nocapture did not come back clean. The failures cluster in late-suite remote-model, RMCP, search-tool, truncation, sqlite-state, and view-image paths.

Repro

• cargo test -p codex-core -- --nocapture
  • Result observed locally on 2026-03-16: 712 passed; 32 failed; 27 ignored

Representative focused repros:

• cargo test -p codex-core --test all suite::remote_models::remote_models_truncation_policy_with_tool_output_override -- --nocapture --exact
  • Fails with: timed out waiting for the remote model codex-test-truncation-override to appear
• cargo test -p codex-core --test all suite::search_tool::search_tool_selection_persists_across_turns -- --nocapture --exact
  • Fails because the mocked /responses sequence receives 0 requests instead of 3

Failure cluster observed in the full run

• suite::remote_models::*
• suite::rmcp_client::*
• suite::search_tool::*
• suite::sqlite_state::mcp_call_marks_thread_memory_mode_polluted_when_configured
• suite::truncation::*
• suite::view_image::view_image_tool_returns_unsupported_message_for_text_only_model

Why this matters

These failures showed up while validating the provider/model-switching work, but they are outside the two files changed for #12. They indicate a broader break in late-suite remote-model availability and tool-selection behavior that needs its own investigation and fix.

Acceptance criteria

• cargo test -p codex-core -- --nocapture passes cleanly
• the representative focused repros above pass cleanly
• root cause is identified and covered by regression tests

[dzianisv]

Update on issue #26:

• PR opened: test(core): fix late-suite harness regressions #30
• Branch: issue-26-core-late-suite-failures
• Commit: 18e77f05d

What was actually broken:

• TestCodexBuilder was treating the default injected test model catalog as an authoritative custom catalog. That disabled remote /models refresh in tests that should have been using the normal test ModelsManager, which is why the late-suite remote-model waiters timed out.
• core integration tests that rely on codex-rmcp-client's test_stdio_server were assuming the cross-package binary already existed in the worktree. In a fresh codex-core-only run it did not, so the turn bailed out before any /responses request and the wiremock cleanup panic masked the real error.

What changed:

• Only explicit custom model_catalog values now take the authoritative ThreadManager::new(...) path.
• The default injected core test catalog stays additive, so remote model refresh still works in the normal test harness.
• stdio_server_bin() now falls back to cargo build -p codex-rmcp-client --bin test_stdio_server under Cargo when the binary is missing from a fresh worktree.
• Added regression tests for both the default-catalog and explicit-custom-catalog behaviors.

Validation that passed:

• cargo test -p core_test_support --lib -- --nocapture
• cargo test -p codex-core --test all suite::remote_models::remote_models_truncation_policy_with_tool_output_override -- --nocapture --exact
• cargo test -p codex-core --test all suite::search_tool::search_tool_selection_persists_across_turns -- --nocapture --exact
• cargo test -p codex-core --test all suite::truncation::mcp_tool_call_output_exceeds_limit_truncated_for_model -- --nocapture --exact
• cargo test -p codex-core --test all suite::rmcp_client::stdio_server_round_trip -- --nocapture --exact
• cargo test -p codex-core --test all suite::sqlite_state::mcp_call_marks_thread_memory_mode_polluted_when_configured -- --nocapture --exact
• cargo test -p codex-core --test all suite::search_tool:: -- --nocapture
• cargo test -p codex-core --test all suite::rmcp_client:: -- --nocapture