fix(codex): preserve valid tool inputs and result pairing

[NishantJoshi00]

Issue: #59

What changed and why?

This follows #60. Raw ApplyPatch arguments can still export as a JSON object even though Codex requires custom_tool_call.input to be a string. Serialize those arguments as text, while unwrapping native patch envelopes unchanged.

Keep foreign WebSearch calls in their existing function-call form. Converting their arguments directly to a native search action can omit the required action type, while leaving a function_call_output without its matching function call. Native search conversion needs to handle both sides together.

Track pending patch calls in message order so a completed ID can be reused by a shell command, edit, write, or raw patch. Scope reader-side duplicate suppression to each function/custom call occurrence; a later canonical result must not erase an earlier call's only result.

Add regressions for reuse in both directions and duplicate shell/patch result records in either order, including structured error output. Exercise overlapping calls, out-of-order completion, and ID reuse through a saved and reloaded Codex session. Extend coverage for shell working directories and call IDs, multiline and empty-file patches, and raw patch inputs. Document the patch format's newline, replace-all, and write limitations.

Dependency

Follows #60, now merged into main. This PR targets main and contains the follow-up fixes and regression coverage.

Validation

• Repository pre-push checks: formatting, both Clippy configurations, all workspace tests, and both WASM configurations. Rust 1.96.1 covers the declared 1.96 minimum.
• 392 tests passed; one existing doctest is ignored.
• Verified that all three pairing checks fail against pre-fix commit 1663f72 for wrong native result types or lost results, then pass with the fix.
• Local CLI smoke test exported six calls reusing one ID with matching results. Installed Codex CLI 0.156.1 applied all four generated patch fixtures, covering an edit, multiline file creation, an empty file, and a multi-file raw patch.

The smoke test checks serialization and native patch application. It does not run a model-backed resumed turn.