Preserve code mode output history budget

[starr-openai]

Why

The latest merged main rust-ci-full run 26140492945 failed these code-mode exec parity tests on macos-aarch64 after both nextest attempts:

• code_mode_exec_explicit_max_above_default_preserves_output
• code_mode_exec_explicit_max_above_default_truncates_larger_output
• code_mode_exec_explicit_max_above_truncation_policy_preserves_output
• code_mode_exec_without_max_preserves_output_beyond_default
• code_mode_exec_without_max_preserves_output_beyond_truncation_policy

Code mode was already shaping its own model-visible output using the @exec output budget. Conversation history then truncated that already-shaped tool output a second time with the turn-wide default policy, which erased larger explicit code-mode outputs before the follow-up request.

What Changed

• Let direct tool results carry an optional history truncation policy when they are recorded back into conversation history.
• Have code-mode exec and wait reuse the same output budget they already use for code-mode result truncation when their outputs are recorded into history.
• Leave every other tool on the existing turn-wide history truncation policy.

Verification

• Source-grounded against failing CI logs from run 26140492945 / jobs 76885895905, 76885895917, and 76885895935.
• just fmt
• git diff --check
• Attempted devbox targeted Bazel repro, but local SSH-agent signing to dev failed before sync/Bazel.
• Started a local compiler-backed codex-lsp-cli validation for the changed Rust files; it is still building the workspace in the background as of PR creation.