compare function_call arguments semantically in codex continuation prefix check

[ItsAlbertZhang]

Fixes #118: with CCP_CODEX_PREVIOUS_RESPONSE_ID=1, the turn after a tool call falls back to a full-context request (not_append_only) whenever the model did not emit the call's arguments in serde_json's canonical form. The retained transcript keeps the raw text the model streamed (translate/reducer.rs:317), the next request re-serializes Claude Code's parsed tool_use.input with sorted keys and no whitespace (translate/request.rs:894), and input_suffix_after_prefix (continuation.rs:532-547) compared the two byte-for-byte. call_id, name, and reasoning items are built identically on both sides, so arguments is the only asymmetric field.

Change

continuation.rs only. input_suffix_after_prefix now compares items through input_items_equivalent: for two FunctionCall items, call_id and name must match and arguments must be byte-equal or parse to equal serde_json::Values; if either side fails to parse and the bytes differ, the items are not equivalent (falls back to the full resend, never to a false continuation). Every other item pair keeps the existing to_value comparison. Nothing recorded or sent upstream changes.

Tests

• function_call_arguments_with_reordered_keys_keep_continuation — records an Edit call with file_path, old_string, new_string and replays it with sorted keys plus a function_call_output; asserts previous_response_id present and a one-item delta. Fails on main with not_append_only.
• function_call_arguments_with_different_values_are_not_append_only — a real value change still yields not_append_only.
• unparseable_function_call_arguments_require_exact_text — {not json byte-identical continues; {not json vs {not json does not.
• input_items_equivalent_compares_arguments_as_json — whitespace-only difference is equivalent; different call_id/name, valid-vs-invalid JSON, and mixed item types are not; Message items compare as before.

cargo fmt --all -- --check clean; cargo test --lib continuation 20 passed; tests/codex_agent_continuation.rs and tests/smoke_cutover.rs pass. No docs/CHANGELOG change (release notes are maintained at release time).

Closes #118

🤖 Generated with Claude Code

[ItsAlbertZhang]

The Full checks failure is a single clippy::result_large_err at tests/codex_agent_continuation.rs:541 (emit_completion returning Result<(), tungstenite::Error>), a file this PR does not touch: CI's floating stable moved from 1.97.1 (last green run on main, 08-19) to 1.98.0, which newly triggers that lint. I left it out of this branch to keep it to the continuation change; happy to rebase once main is adjusted.