Resuming a session drops `reasoning_text` from all earlier assistant messages (kimi-k3)

[dzamoshchin]

Describe the bug

After --resume, earlier assistant messages are sent without their reasoning. With kimi-k3 (first-party, /chat/completions, not BYOK), a live session sends each assistant message's reasoning_text back on later requests. After resume, the same messages are sent without it. The same message across requests:

before resume: {"role": "assistant", "reasoning_text": "Let me run the test and look at stats.py.", "tool_calls": [...]}
after resume:  {"role": "assistant", "tool_calls": [...]}

The reasoning is still in the session's events.jsonl (reasoningText), so it is removed when the session is loaded, not lost when saved. Setting COPILOT_STRIP_REASONING_ON_RESUME=false keeps it.

Kimi K3 is trained with its reasoning preserved across turns, so after a resume it sees every earlier turn with empty reasoning. In our tests, K3 produced no reasoning on any request after a resume.

Separately, on the first request after resume, earlier tool_calls[].function.arguments are re-serialized compactly ({"command": "…"} becomes {"command":"…"}). That breaks prompt caching at the first assistant message.

Affected version

1.0.90-5

Steps to reproduce the behavior

1. Run copilot -p "..." --model kimi-k3 --reasoning-effort high --allow-all-tools --no-ask-user --log-level all --log-dir logs1 with a prompt that produces a few tool calls
2. Run copilot --resume=<session-id> -p "..." with the same flags and --log-dir logs2
3. Compare the Wire request: bodies: in logs1, assistant messages carry reasoning_text; in the first request in logs2, the same messages have none

Expected behavior

A resumed session sends the same assistant history, including reasoning_text, as the live session did.